Merge origin/master into llm-error-recovery-rfc

.agents/notes/AGENTS.md

@@ -0,0 +1,3 @@
+# AGENTS.md — Agent Notes
+
+Agent Notes are effectively RFCs written by agents: durable proposals and decision records that preserve rationale, alternatives, consequences, and verification contracts. Follow the [documentation standard](../../docs/AGENTS.md) and the [Agent Note contract](README.md).

.agents/notes/README.md

@@ -0,0 +1,111 @@
+# Agent Notes
+
+One kind of design doc lives here. An **Agent Note** records a decision or proposal that shapes this codebase — the *why* and *what we gave up*, the parts code and docs can't carry. This file is the front door and contract: where Agent Notes live, when to write one, and [the in-file format](#the-file-format).
+
+## Layout and naming
+
+Every Agent Note has two axes, both encoded in its **path** — `{lifecycle}/{class}/yyyy-mm-dd-topic-title.md`:
+
+- **Lifecycle** (the top-level folder) is the Agent Note's status, and an Agent Note moves between folders as that status changes:
+  - **`proposed/`** — proposals reviewed before implementation; not yet built (or only partly).
+  - **`implemented/`** — the decision shipped. The file records what was decided and what was rejected, and is **kept current with what actually shipped**: when the code later moves a file, renames a package, or changes a key/default, the Agent Note is updated in the same change to match (facts only — paths, names, structure — not the decision itself). See [implemented/AGENTS.md](implemented/AGENTS.md).
+  - **`rejected/`** — the proposal was considered and declined. Kept for the record so the rejection isn't re-litigated.
+- **Class** (the nested folder) is the *kind* of decision — see [Classification](#classification) below.
+
+The date in the filename is when the topic was **first proposed** (per git history). Cross-references between Agent Notes use relative markdown links (`[topic](../../implemented/architecture/2026-…-….md)`) — never bare prose or numbers — so they are mechanically checkable and survive moves between folders.
+
+The tree is the inventory: browse its lifecycle/class folders or search the repository. Do not add a centralized `INDEX.md`; the [no-index Agent Note](implemented/process/2026-07-19-remove-generated-agent-note-index.md) owns the rationale.
+
+## Classification
+
+Each Agent Note belongs to one path-encoded class from the closed set in `scripts/agent-note-tree.ts`; the classification gate rejects other folders. Adding a class requires updating the canonical set and this section. See the [classification Agent Note](implemented/process/2026-06-20-agent-note-classification.md).
+
+| Class | What it covers |
+|---|---|
+| `feature` | A new user- or model-facing capability. |
+| `bug-fix` | Corrects a defect or closes a gap a postmortem surfaced. |
+| `simplification` | Removes code, behavior, or surface area without adding a capability. |
+| `architecture` | A structural decision about the **shipped source** — how packages relate, what the runtime vocabulary is. |
+| `process` | Tooling, policy, or workflow **around** the code — gates, the package manager, vendoring — not runtime behavior. |
+| `testing` | Test infrastructure and strategy. |
+
+The `architecture` / `process` line: **architecture** is about the source we ship; **process** is the surrounding tooling and workflow. (`refactor` is deliberately absent — it overlaps `simplification`, whose discriminator, "does observable behavior change?", already covers it.)
+
+## When to write one
+
+Write an Agent Note when a decision is **durable** (it shapes the codebase beyond a single function or package), **contested** (there was a real alternative a reasonable engineer might have chosen), and **surprising** (a future reader would otherwise ask "why on earth is it done this way?"). A proposal for substantial future work starts in `proposed/`; a decision already made starts in `implemented/`. Pick the class folder that matches the decision (see [Classification](#classification)).
+
+Do NOT write one for a mechanical or local choice (a variable name, a one-file refactor), for anything already enforced and explained by a gate or a convention in AGENTS.md, or for a still-provisional decision tagged `TODO(...)` in the code — record those as TODOs and promote to an Agent Note only once they settle. An Agent Note is never edited into a *different decision*: supersede it with a new one and cross-link. (Editing an `implemented/` Agent Note to track where its already-made decision now *lives* — a moved file, a renamed package — is not a different decision and is required, not forbidden; see [implemented/AGENTS.md](implemented/AGENTS.md).)
+
+## The file format
+
+Every Agent Note follows one in-file format, enforced by `pnpm run verify-agent-note-format` ([scripts/verify-agent-note-format.ts](../../scripts/verify-agent-note-format.ts), part of `doc-sync`); the rationale for the format — and the alternatives it rejected — is [the uniform-format Agent Note](implemented/process/2026-07-05-uniform-agent-note-format.md).
+
+### The header block
+
+The first three lines of every Agent Note are exactly:
+
+```markdown
+# Agent Note: <title>
+
+Status: <status>
+```
+
+followed by a blank line. The `Status:` value is one of three forms, and must agree with the lifecycle folder the file sits in — the gate cross-checks them:
+
+- `Status: proposed`
+- `Status: implemented`
+- `Status: rejected — <why, in one line>`
+
+The status carries no dates and no parentheticals: the filename holds the first-proposed date, git holds everything else, and an "accepted in amended form" note is body content (state the amendment where the decision is stated). The rejection reason is the one status with content, because a rejected Agent Note's verdict is the fact readers come for.
+
+### The body skeleton
+
+Every Agent Note opens its body with `## Problem` — the motivation, written to stand without the solution. What follows depends on the lifecycle; recurring sections use these canonical names and nothing else, while genuinely bespoke technical sections (package topology, wire contracts, schemas) remain free-form between the required ones.
+
+#### `proposed/`
+
+```markdown
+## Problem
+## Proposal
+…bespoke sections…
+## Alternatives considered
+## Acceptance criteria
+## Risks
+```
+
+`## Proposal` is the intended change and may legitimately speak in the future tense — plans, migration steps, and open questions belong here while the work is unbuilt. `## Acceptance criteria` says what observable state means done. `## Risks` covers both what could go wrong and what the change knowingly gives up.
+
+#### `implemented/`
+
+```markdown
+## Problem
+## Decision
+…bespoke sections…
+## Alternatives considered
+## Consequences
+```
+
+`## Decision` describes shipped reality in the present tense, and the whole file is kept current with it per [implemented/AGENTS.md](implemented/AGENTS.md). `## Consequences` records what the trade-off cost **and** bought. Proposal-era headings are spec-speak here and the gate rejects them: `## Proposal`, `## Plan`, `## Migration plan`, and `## Acceptance criteria` may not appear in an implemented Agent Note (the [slop checklist](../../docs/AGENTS.md) names why). A `## Testing`, `## Deferred`, or `## Related` section is fine where it states present-tense fact.
+
+#### `rejected/`
+
+A rejected Agent Note is the proposal, frozen: it keeps whatever proposal-time sections it had (including `## Acceptance criteria` or `## Plan`), and the verdict lives on the `Status:` line. Only the header block, the `## Problem` opener, a `## Proposal` section, and the Alternatives-considered mandate below apply.
+
+### Alternatives considered — mandatory
+
+Every Agent Note carries an `## Alternatives considered` section: each genuine alternative and why it lost, one bold-led paragraph per alternative or a `### Why not <X>?` subsection per contested one. A decision recorded without what it beat invites re-litigation — the failure Agent Notes exist to prevent.
+
+Alternatives are recorded, never invented. An Agent Note dated before 2026-07-05 whose alternatives are not reconstructible from the record carries this exact comment in place of the section, which the gate accepts for pre-format files only:
+
+```markdown
+<!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->
+```
+
+### Moving between lifecycles
+
+Moving a file between lifecycle folders means updating the `Status:` line and re-satisfying that folder's skeleton in the same change — the gate fails the move otherwise. Concretely, `proposed/` → `implemented/` rewrites `## Proposal` into a present-tense `## Decision`, folds `## Acceptance criteria` and `## Risks` into `## Consequences` (or a present-tense `## Testing`/`## Verification` section for what now pins the behavior), and drops plans in favor of what shipped — the rewrite [implemented/AGENTS.md](implemented/AGENTS.md) requires, made mechanical. `proposed/` → `rejected/` only adds the reason to the `Status:` line and freezes the file.
+
+### Chinese counterparts
+
+A `.zh.md` counterpart mirrors its English sibling's structure section-for-section under the [i18n contract](../../docs/i18n/README.md); the machine-checked header tokens (`# Agent Note: ` and the `Status:` line) stay in English verbatim. The format gate skips `.zh.md` files — the pairing gate owns their consistency.

.agents/notes/implemented/AGENTS.md

@@ -0,0 +1,11 @@
+# AGENTS.md — Implemented Agent Notes
+
+These Agent Notes describe shipped decisions. Follow the [root instructions](../../../AGENTS.md), [documentation standard](../../../docs/AGENTS.md), and [Agent Note format](../README.md#the-file-format); `verify-agent-note-format` gates the lifecycle-specific structure.
+
+## Keep an implemented Agent Note current with what actually shipped
+
+Keep paths, symbols, defaults, and mechanisms current in the same change that alters them. Rewrite stale facts in place; do not append change history.
+
+### This is not a license to rewrite the *decision*
+
+Update factual realization in place. A reversal of the decision or its rationale requires a new Agent Note and cross-link; see the [Agent Note contract](../README.md).

.agents/notes/implemented/architecture/2026-06-11-content-block-vocabulary.md

@@ -0,0 +1,26 @@
+# Agent Note: Provider-neutral content-block vocabulary owned by dsh-llm
+
+Status: implemented
+
+## Problem
+
+The harness needs one internal language for messages that the loop, session log, and all plugins speak.
+
+## Decision
+
+Own the vocabulary: messages are arrays of typed content blocks (`text`, `reasoning`, `tool-call`, `tool-result`), with the union derived from the merge-extensible `ContentBlockMap` so plugins add block types via declaration merging. The same merge-extensible-map pattern types every "stringly" field (`MessageSource`, `FinishReason`, `TurnTrigger`, `TurnEndReason`). Streaming is a raw chunk protocol; `BlockAssembler` is the single shared assembly implementation. Adapters translate to provider wire formats — mapping cost lives in adapters, where it belongs.
+
+In-session context injection (`context/message`, `steering/message`) renders as tagged user-role envelopes (the system-reminder pattern) rather than a new role, so adapters carry zero burden. Live-adapter validation confirms this rendering for current DeepSeek behavior; a future provider-specific mismatch belongs in that adapter rather than a new canonical role.
+
+## Alternatives considered
+
+- **Mirror the DeepSeek/OpenAI chat-completions shape** — zero mapping cost for the first provider, but awkward for rich content (reasoning, tool results as structured blocks).
+- **Adopt Anthropic's Messages block structure verbatim** — battle-tested, but the canonical types would mirror a third-party API the harness does not target first.
+
+## Consequences
+
+- Reasoning has a core home without provider-specific shapes.
+- Multimodal blocks return only with coordinated adapter, UI, and compaction support; see [the drop-image Agent Note](../simplification/2026-07-04-drop-image-content-block.md).
+- Cache hints and assistant prefill remain absent until a shipping adapter can honor them; see the [producer-less variants](../simplification/2026-07-04-prune-producerless-vocabulary-variants.md) and [inert request knobs](../simplification/2026-07-04-drop-inert-request-knobs.md) Agent Notes.
+- Every adapter pays a translation cost; the first real adapters have since validated the streaming protocol, and new adapters should continue proving their provider-specific mapping in adapter-local tests.
+- IDs that cross package boundaries are branded (`CallId`, the shared agent/session `SessionId`) — nominal typing at zero runtime cost.

.agents/notes/implemented/architecture/2026-06-11-custom-schema-dsl.md

@@ -0,0 +1,21 @@
+# Agent Note: Custom typed tool-schema DSL instead of schemastery
+
+Status: implemented
+
+## Problem
+
+Tool parameters must reach the model as standard JSON Schema while giving tool authors typed `execute(args)` without casts. Schemastery already serves plugin config, but the tool-author API needs per-property `required: true` booleans rather than JSON Schema's separate `required` array.
+
+## Decision
+
+A small custom DSL in dsh-tools: `SchemaSpec` (per-property specs with `required: true` booleans), type-level `InferArgs<S>` mapping a spec to the argument type (required keys non-optional, others genuinely optional via `?`), a runtime `schemaSpecToJsonSchema()` converter, and `defineTool()` tying them together. Raw JSON-Schema `ToolDefinition`s remain accepted by `ToolRegistry.register()` — that's how MCP-sourced tools arrive.
+
+## Alternatives considered
+
+**Schemastery** (already vendored, used for plugin Config) was evaluated and rejected for this use: it targets validation / transformation against StandardSchema, not JSON Schema *generation*, so it would add indirection without producing the wire format cleanly.
+
+## Consequences
+
+- First-party tool authors get zero-cast typed args; the type gymnastics cost stays inside the core package (sanctioned by the AGENTS.md type-safety policy).
+- The DSL is deliberately small (string/number/boolean/object/array, enum, default, nested properties/items). Gaps vs full JSON Schema (unions, formats, constraints) are accepted until real tools demand them.
+- The `InferArgs` mapping is regression-tested at the type level after an early optionality bug.

.agents/notes/implemented/architecture/2026-06-11-dev-invariants-over-deep-readonly.md

@@ -0,0 +1,58 @@
+# Agent Note: Source-owned session immutability and dev-mode invariants
+
+Status: implemented
+
+## Problem
+
+The session log needs two different protections: immutable ownership of each stored fact, and checks for relationships among facts across time and service seams. Conflating them in an optional development plugin would leave production history vulnerable; trying to express both through TypeScript readonly types would not create a runtime boundary or describe relational rules.
+
+The session log is the durable source of truth for replay, request reconstruction, persistence, and user-visible history. Code outside the session package must be able to inspect that history without retaining a reference that can rewrite it later, and inputs accepted from callers must not remain connected to caller-owned mutable objects.
+
+Immutability of individual values is only half of the contract. A log can contain perfectly immutable records whose sequence, turn/step nesting, tool-call pairing, scoped delivery, or reconstructed model request is wrong. Those rules relate multiple records or services and cannot be established by freezing one object.
+
+TypeScript readonly types are not a sufficient runtime boundary. They disappear when the program runs, a cast can bypass them, and a recursive `DeepReadonly<T>` would spread through every log and message consumer even though some downstream request-processing APIs intentionally work with mutable values.
+
+## Decision
+
+Responsibility is split between an always-on storage boundary and optional development assertions.
+
+### Session owns immutable history
+
+`Session` accepts an event only after one recursive pass has materialized a lossless JSON snapshot. That pass rejects unsupported values and produces the exact detached record that enters the log, so validation and storage cannot observe different values from a stateful getter or retain caller-owned nested references.
+
+The accepted event and all of its descendants are deep-frozen before publication. `append()` returns that owned frozen event, `session/event` observers receive the same record, and `session.events` returns a frozen array snapshot. A previously returned array does not grow after a later append. Seed records pass through the same validation, snapshot, and freeze boundary before construction succeeds.
+
+This guarantee belongs in `Session`, not in an optional listener, because every composition relies on trustworthy history. A production deployment, a focused test, or a custom embedding receives the same storage semantics whether or not development support plugins are registered.
+
+### Derived requests remain detached
+
+`deriveMessages()` projects logged surface events into detached, deep-frozen `Message` objects and returns a fresh array snapshot. Request assembly can therefore combine derived history with other inputs without exposing a path back into the log. The cache reuses safe immutable projections rather than recloning the complete history for each model call.
+
+### The invariants plugin checks relationships
+
+`dsh-invariants` is a pure-listener development plugin. It does not freeze records and has no configuration; disposal removes only its assertions. It checks rules that require trace state or observation of another seam, including monotonic sequence numbers, turn and step nesting, tool-call/result pairing, legal agent-status transitions, subject-correct scoped dispatch, and equality between a loop-built request and the request reconstructed from its session-log prefix.
+
+When the plugin attaches to an existing or seeded session, it replays the immutable log to rebuild trace state. This makes hot reload safe in the middle of a turn without giving the plugin ownership of session storage.
+
+## Alternatives considered
+
+### Pervasive deep-readonly types
+
+[The rejected immutable-public-surfaces proposal](../../rejected/architecture/2026-06-11-immutable-public-surfaces.md) would apply a recursive readonly type across public log and message surfaces. That provides editor feedback but not a runtime guarantee: TypeScript types are erased and plugin code can cast through them. It also pushes readonly types into consumers where mutation is intentional. Runtime ownership at the `Session` boundary protects every caller without that type propagation.
+
+### Development-only freezing
+
+Freezing history only when an invariants plugin is installed would make the core guarantee composition-dependent. Code could pass development tests and still corrupt history in production or in a focused composition that omits the plugin. Storage immutability is therefore always on, while the more expensive relational checks remain opt-in development support.
+
+### Clone only when deriving messages
+
+Detaching `deriveMessages()` would protect the most common request path but leave other readers of `session.events`, append return values, and session-event observers able to mutate durable history. The log must protect its own boundary; derived projections are an additional isolation boundary, not a substitute.
+
+## Consequences
+
+- Every accepted live or seeded session event is detached from caller-owned inputs and deeply immutable before any observer can receive it.
+- `session.events` exposes stable immutable snapshots instead of the private growing array.
+- Request-side mutation cannot reach stored history through derived messages.
+- Development builds can enable relational assertions without changing storage behavior, and disposing or omitting the plugin does not weaken log immutability.
+- `dsh-invariants` has no `Config` surface because it has no behavior to tune.
+- The runtime boundary carries a recursive snapshot-and-freeze cost once per accepted event; later readers and cached projections reuse the owned immutable records.

.agents/notes/implemented/architecture/2026-06-11-event-sourced-sessions.md

@@ -1,12 +1,10 @@
-# RFC: Event-sourced sessions with derived message history
+# Agent Note: Event-sourced sessions with derived message history
 
-Status: implemented (accepted 2026-06-11)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
+## Problem
 
-## Context
-
-The MVP requires strict event-based tracing with fully replayable sessions (严格的基于事件的trace、logging系统，session完全可回放). Two models were considered: a mutable message array with events fired as notifications (simpler, but state and log can diverge), or event-sourcing where the log IS the state.
+The MVP requires strict event-based tracing with fully replayable sessions (严格的基于事件的trace、logging系统，session完全可回放).
 
 ## Decision
 
@@ -14,11 +12,15 @@ A `Session` is an append-only log of typed `SessionEvent`s — the single source
 
 Appends are synchronous (the hot path never blocks on I/O); `session/event` is a sync notification; persistence plugins buffer write-behind and drain at the awaited `session/flush` checkpoint fired at every turn end.
 
-Ordering contract: the loop appends to the session *before* emitting the corresponding Cordis event, and the `agent/step-result` waterfall runs before the `assistant/message` append so the log records what tool dispatch actually used (post-review fix; regression-tested).
+Ordering contract: the loop appends to the session *before* emitting the corresponding Cordis event, and the `agent/step-result` waterfall runs before the `assistant/message` append so the log records the message tool dispatch actually used. Regression tests pin that ordering.
+
+## Alternatives considered
+
+**A mutable message array with events fired as notifications** — simpler, but state and log can diverge; with event-sourcing the log IS the state, so divergence is structurally impossible.
 
 ## Consequences
 
 - Replay, trace, and telemetry are structurally guaranteed, not bolted on.
 - Persistence stays a plugin concern; the in-memory store ships in dsh-session.
-- The event vocabulary is merge-extensible (plugins add e.g. compaction events); it carries a TODO(review) marker until the first persistence plugin and real adapter exercise it.
+- The event vocabulary is merge-extensible (plugins add e.g. compaction events); [session persistence](2026-06-14-session-persistence.md) froze its shape once the log became durable.
 - Derivation cost grows with log length — compaction (future plugin) is the intended mitigation, not log mutation.

.agents/notes/implemented/architecture/2026-06-11-microkernel-event-taxonomy.md

@@ -0,0 +1,29 @@
+# Agent Note: Microkernel — extension via Cordis event taxonomy, one concrete loop
+
+Status: implemented
+
+## Problem
+
+The product principle is "everything is a plugin": hooks, /goal, /loop, dynamic workflows, compaction, sandboxing, permissions, UI, persistence, MCP, skills must all be writable as plugins without modifying the core.
+
+## Decision
+
+Pure Cordis event taxonomy. The loop's extension seams are typed events with deliberate dispatch modes:
+
+- **waterfall** (around-middleware) where plugins transform, veto, recover, or wrap: `agent/prompt-submit`, `agent/request`, `agent/request-error`, `agent/step-result`, `agent/turn-continuation`, `tools/pre-execute`, `tools/execute`, `tools/post-execute`, `llm/stream`, `system-prompt/assemble`.
+- **serial** (awaited in listener order; a bail value stops later listeners) for ordered checkpoints: every `agent/pre-step` and `agent/post-step` listener runs when all abstain, while the first stop returned from `agent/turn-stop` makes the terminal decision final.
+- **parallel** (awaited fan-out) where every listener must get an independent chance: the `session/flush` durability checkpoint.
+- **emit** (synchronous fire-and-forget) for notifications: turn/step boundaries, stream chunks, lifecycle, errors, and the contained immutable `tools/result` observation.
+
+The event vocabulary lives in interface packages (dsh-agent declares the agent/* events); `@deepseek-ai/dsh-agent-loop` is the only concrete loop plugin and is itself swappable — nothing outside it may depend on it.
+
+## Alternatives considered
+
+**A purpose-built middleware stack (koa-compose style)** and **an explicit phase state machine plugins insert into** — both would re-implement dispatch, disposal, and reload semantics that Cordis's native event system already provides; as Cordis effects, listeners get HMR and disposal for free.
+
+## Consequences
+
+- Every MVP feature maps to a listener (the [feature → mechanism map](../../../../docs/cookbook/extension-cookbook.md#the-feature--mechanism-map) is the proof obligation, kept current).
+- HMR and disposal come free: listeners and registrations are Cordis effects.
+- Waterfall semantics (call `next()` or short-circuit) are non-obvious and must be taught — documented in AGENTS.md and covered by composition tests.
+- The loop must be defensive: plugin exceptions are contained at turn level, steering from any seam is never stranded (regression-tested).

.agents/notes/implemented/architecture/2026-06-11-runtime-arg-validation.md

@@ -1,10 +1,8 @@
-# RFC: Runtime arg validation at the model boundary
+# Agent Note: Runtime arg validation at the model boundary
 
-Status: implemented (accepted 2026-06-13)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
-
-## Context
+## Problem
 
 `defineTool` ([the custom schema DSL](2026-06-11-custom-schema-dsl.md)) gives tool authors a typed `execute(args)` via the `InferArgs<S>` mapping. But that type is a compile-time claim about a value that arrives at runtime as model-generated JSON: nothing forced the model to honor the schema, so a malformed call — missing a required key, a string where a number was declared, an enum value outside the set — reached `execute` typed-in-name-only. The tool body then either crashed on the bad shape (a generic stack trace the model can't act on) or, worse, silently misbehaved. Meanwhile the converter already encodes the exact structure a validator would need to walk.
 
@@ -17,6 +15,8 @@ The validator mirrors `schemaSpecToJsonSchema` semantics exactly — same struct
 ## Consequences
 
 - The model gets actionable feedback on its own malformed calls instead of an opaque crash, closing the gap between `InferArgs`'s promise and runtime reality.
-- The validator and `InferArgs` must stay in agreement; that drift risk is to be closed by a property test ([property-based testing](../testing/2026-06-11-property-based-testing.md), not yet landed) generating args that satisfy `InferArgs` and asserting they pass `validateArgs`. Until then the agreement rests on the example tests and the shared converter structure.
+- The validator and `InferArgs` must stay in agreement; [a property test](../testing/2026-06-11-property-based-testing.md) generates args satisfying a spec and asserts they pass `validateArgs` (with targeted corruptions rejected), closing that drift risk mechanically.
 - `ToolArgsError` is a plain `Error` with a `code` field for now; if a harness-wide error taxonomy lands it becomes a subclass without changing callers that read `.message`.
 - Validation cost is negligible next to a model call.
+
+<!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->

.agents/notes/implemented/architecture/2026-06-11-structured-error-taxonomy.md

@@ -1,15 +1,11 @@
-# RFC: Structured error taxonomy
+# Agent Note: Structured error taxonomy
 
-Status: implemented (accepted 2026-06-14)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
-
-## Context
+## Problem
 
 Failures crossed seams as bare strings. A tool error flattened to a text block — name, code, and stack lost — so a future sandbox/retry plugin couldn't tell ENOENT from EACCES, and the model got less actionable feedback than it could. A non-Error throw degraded further: the loop wrapped it in `new Error(String(x))`, dropping any code. And `LlmError` was the only typed error in the system, with no shared base, so there was nothing for a consumer to `instanceof` against generically.
 
-This is the last of the runtime-validation / error-taxonomy pieces and the one the user was most skeptical of, so it was deliberately built **last and in isolation**: the earlier PRs (arg validation, dev invariants) threw plain `Error`s with a `code` field, decoupled from any shared base, so this change is a pure upgrade and is independently revertible without unpicking them.
-
 ## Decision
 
 A single `HarnessError extends Error` base in `dsh-llm` (the leaf package every other imports — no new dependency edge): a stable `code` distinct from `message`, `cause` chaining via `ErrorOptions`, and `name` defaulting to the subclass. `isHarnessError` narrows at seams.
@@ -23,4 +19,6 @@ A single `HarnessError extends Error` base in `dsh-llm` (the leaf package every
 - Errors are machine-routable end-to-end: a plugin can branch on `error.code` rather than substring-matching a message.
 - One base class is imported widely, but it lives in the package everyone already depends on, so the cost is a single import, not a new edge.
 - `deriveMessages` does not surface `error` into model history — the model still sees the text block; the structured field is for code and replay.
-- Reverting this PR returns the earlier errors to plain `Error`+`code` form; nothing else in the stack depends on the shared base.
+- Argument validation and dev invariants retain their existing codes and behavior; the shared base adds cross-seam routing metadata without changing model-facing text.
+
+<!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->

.agents/notes/implemented/architecture/2026-06-11-tool-schemas-in-prompt-assembly.md

@@ -1,17 +1,19 @@
-# RFC: Tool schemas are part of the system-prompt assembly
+# Agent Note: Tool schemas are part of the system-prompt assembly
 
-Status: implemented (accepted 2026-06-11)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
+## Problem
 
-## Context
-
-On the wire, tool schemas travel in a dedicated `tools` field of the model request, not in prompt text. Architecturally, though, "what the model is told it can do" is one coherent concern: prompt sections and the tool list are assembled from the same plugin contributions and consumed at the same moment. The alternative — the loop querying the tool registry separately from the prompt service — splits one concern across two seams.
+On the wire, tool schemas travel in a dedicated `tools` field of the model request, not in prompt text. Architecturally, though, "what the model is told it can do" is one coherent concern: prompt sections and the tool list are assembled from the same plugin contributions and consumed at the same moment.
 
 ## Decision
 
 `PromptAssembly { sections, tools }`: the system-prompt service collects ordered text sections AND tool schemas (the tool registry auto-contributes a provider). The loop consumes one assembly per step; adapters map `sections` to the provider's system slot and `tools` to the wire `tools` field. The `system-prompt/assemble` waterfall is therefore a single interception point for everything the model is told up front — tool filtering (ToolSearch / progressive disclosure) is an assembly rewrite, same as prompt edits.
 
+## Alternatives considered
+
+**The loop queries the tool registry separately from the prompt service** — splits one coherent concern across two seams, and every interception that wants to shape "what the model is told" (tool filtering, plan mode) would need two listeners on two surfaces instead of one assembly rewrite.
+
 ## Consequences
 
 - One waterfall governs the model's standing context; plugins like plan mode can swap prompt text and visible tools in one listener.

.agents/notes/implemented/architecture/2026-06-13-capability-seams.md

@@ -1,29 +1,30 @@
-# RFC: Capability seams — interface / implementation / consumer split
+# Agent Note: Capability seams — interface / implementation / consumer split
 
-Status: implemented (accepted 2026-06-13)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
-
-## Context
+## Problem
 
 The harness has swappable capabilities — bash execution today, sandboxed/remote executors and alternative model providers tomorrow. A capability has three concerns that change at different rates and for different reasons: the *contract* (what the capability is), the *implementation* (how it runs), and the *consumer surface* (what the model and other plugins program against). Bundling them in one package couples those rates of change — swapping a local executor for a sandboxed one would churn the tool schemas the model sees, even though the model-facing contract never changed.
 
-This is distinct from "who provides vs. needs a capability at runtime", which Cordis already answers with services + `inject` (a provider registers `ctx.bash`; a consumer declares `inject: ['bash']` and its fiber pends until the service exists). That mechanism is necessary but doesn't dictate package boundaries; this RFC does.
+This is distinct from "who provides vs. needs a capability at runtime", which Cordis already answers with services + `inject` (a provider registers `ctx.bash`; a consumer declares `inject: ['bash']` and its fiber pends until the service exists). That mechanism is necessary but doesn't dictate package boundaries; this Agent Note does.
 
 ## Decision
 
 A swappable capability is **three packages**:
 
-1. **Interface** — an abstract service + the vocabulary types, owning the `ctx.<key>` and depending only on cordis (e.g. `dsh-bash`: `BashExecutor`, `BashRunResult`, `BashTask`).
+1. **Interface** — an abstract service + the vocabulary types, owning the `ctx.<key>` and depending only on its vocabulary dependencies (e.g. `dsh-bash`: `BashExecutor`, `BashRunResult`, `BashProcess`).
 2. **Implementation** — a concrete subclass loaded as a plugin (e.g. `dsh-bash-local`: subprocesses, process-group kills, spill-file truncation). Sandboxed/remote backends are sibling packages implementing the same interface.
-3. **Consumer** — what the model and plugins see (e.g. `dsh-tool-bash`: the `bash`/`bash_output`/`bash_kill` tool schemas). Consumers `inject` the interface key and never import implementation types.
+3. **Consumer** — what the model and plugins see (e.g. `dsh-tool-bash`: the `bash` schema, with background handles registered into the generic task runtime). Consumers `inject` the interface key and never import implementation types.
 
 Implementation and consumer then evolve independently: a sandboxed executor replaces `dsh-bash-local` without touching a tool schema.
 
-Alternatives considered: **one combined package** — rejected because it recouples the three rates of change the split exists to separate (the whole point). **`@cordisjs/plugin-capability`** — a different axis entirely: it is a permission/capability-*security* service (named permissions with inheritance, tested against a session via `ctx.capability.test`), a candidate for the deferred permissions/sandbox work on the `tools/execute` veto seam, NOT a mechanism for swapping implementations. Confusing the two ("capability") is the trap this RFC names.
-
 The split is not mandatory when the parts are genuinely one concern: the LLM seam folds interface + consumer into `dsh-llm` (the consumer is the loop itself, not a swappable schema surface) with adapters as the implementation packages. Don't split preemptively — a capability with one conceivable implementation and one consumer stays one package until a second appears.
 
+## Alternatives considered
+
+- **One combined package** — rejected because it recouples the three rates of change the split exists to separate (the whole point).
+- **`@cordisjs/plugin-capability`** — a different axis entirely: it is a permission/capability-*security* service (named permissions with inheritance, tested against a session via `ctx.capability.test`), a candidate for the deferred permissions/sandbox work on the `tools/pre-execute` deny/ask seam, NOT a mechanism for swapping implementations. Confusing the two ("capability") is the trap this Agent Note names.
+
 ## Consequences
 
-More packages and more boilerplate per capability (a `package.json`/`tsconfig`/README trio, the inject wiring). Bought: implementations and consumers ship and version independently, and a new backend never risks the model-facing contract. The rule is documented in [AGENTS.md](../../../../AGENTS.md) § Conventions ("Capability seams are three packages") and [architecture.md](../../../architecture.md) § "Capability seams"; the bash trio is the reference template. When to fold vs. split is a judgment call the architecture doc spells out — this RFC records *why* the default is to split.
+More packages and more boilerplate per capability (a `package.json`/`tsconfig`/README trio, the inject wiring). Bought: implementations and consumers ship and version independently, and a new backend never risks the model-facing contract. The rule is documented in [AGENTS.md](../../../../AGENTS.md) § Conventions ("Capability seams are three packages") and [architecture.md](../../../../docs/architecture.md) § "Capability seams"; the bash trio is the reference template. When to fold vs. split is a judgment call the architecture doc spells out — this Agent Note records *why* the default is to split.

.agents/notes/implemented/architecture/2026-06-13-twin-llm-adapters.md

@@ -1,10 +1,8 @@
-# RFC: Two LLM adapters as a design-verification twin
+# Agent Note: Two LLM adapters as a design-verification twin
 
-Status: implemented (accepted 2026-06-13)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
-
-## Context
+## Problem
 
 `dsh-llm` owns a provider-neutral streaming vocabulary — the `StreamChunk` protocol (`block-start`, `text-delta`, `reasoning-delta`, `tool-call-delta`, `block-end`, `usage`, `finish`) and the content-block types ([the content-block vocabulary](2026-06-11-content-block-vocabulary.md)). A vocabulary defined against a single adapter risks baking that adapter's quirks into the "neutral" contract: anything the one implementation happens to do becomes the de-facto spec, and the abstraction is unverified until a second provider arrives — by which point the leak is expensive to fix.
 
@@ -17,8 +15,11 @@ Ship **two** adapters against the one contract from the start, deliberately buil
 
 The rule they enforce: **anything the StreamChunk vocabulary cannot express for BOTH implementations is a core-vocabulary bug**, caught immediately rather than at the next provider. The pair pinned down conventions now documented on `StreamChunk` in `dsh-llm/src/types.ts`: usage emitted before finish, nothing after finish, tool-call `arguments` as raw JSON strings end-to-end, and the two sanctioned error paths (throw from `stream()` *or* end with `finish {kind:'error'|'aborted'}`) that a consumer must handle on both sides — a divergence the library-backed adapter surfaced that a single hand-rolled adapter would have hidden.
 
-Alternatives considered: **a single adapter** — less code and half the e2e cost, but leaves the "provider-neutral" claim unverified; the vocabulary would encode DeepSeek-via-fetch assumptions silently. **A mock second adapter** — cheaper but doesn't exercise a real provider's wire quirks, so it proves little. The twin is real-on-real.
+## Alternatives considered
+
+- **A single adapter** — less code and half the e2e cost, but leaves the "provider-neutral" claim unverified; the vocabulary would encode DeepSeek-via-fetch assumptions silently.
+- **A mock second adapter** — cheaper but doesn't exercise a real provider's wire quirks, so it proves little. The twin is real-on-real.
 
 ## Consequences
 
-Double the adapter maintenance and double the key-gated e2e surface (both adapters cover V4 Flash and Pro across representative thinking/effort modes). Bought: a continuously-verified neutrality guarantee for the most leak-prone abstraction in the codebase, and a worked second example for adapter authors. The two share the core Config shape (`apiKey`/`baseURL`/`models`) so a deployment swaps mostly one line, but the reasoning knob differs — `dsh-llm-deepseek` takes `thinking`/`reasoningEffort`, `dsh-llm-pi-ai` takes a single `reasoning` level — so a swap translates that field. If the maintenance cost ever outweighs the verification value (e.g. once conformance tests from [architectural conformance](../../proposed/process/2026-06-11-architectural-conformance.md) cover the contract mechanically), retiring the twin to a single adapter + the conformance kit would be a new RFC superseding this one.
+The twin doubles adapter and key-gated e2e maintenance—both cover V4 Flash and Pro across representative reasoning modes—in exchange for continuous seam-neutrality validation and a second implementation example. Both use `apiKey`, `baseURL`, and `models`; the hand-rolled adapter exposes `thinking`/`reasoningEffort`, while pi-ai exposes one `reasoning` level. A future conformance suite could justify retiring one adapter through a superseding Agent Note.

.agents/notes/implemented/architecture/2026-06-14-session-persistence.md

@@ -0,0 +1,34 @@
+# Agent Note: Session persistence as an abstract service over the existing `SessionEvent`
+
+Status: implemented
+
+## Problem
+
+Sessions lived only in memory. The example `session-jsonl.ts` plugin (duplicated byte-for-byte in both examples) was write-only telemetry: it buffered `session/event` and appended JSON lines, with no read/replay path, no crash-safety (no fsync, no atomic write, a fire-and-forget dispose drain), no listing, and no format versioning. Nothing could rehydrate a past session from disk into a live agent, so durable resume ("continue yesterday's task"), durable forking, and the ACP `session/load` method ([ACP support](../feature/2026-06-14-acp-agent-client-protocol.md)) were all impossible.
+
+The [event-sourced model](2026-06-11-event-sourced-sessions.md) makes the append-only log the single source of truth and derives LLM history from it. Persistence had to stay faithful to that: persist the existing `SessionEvent` directly, with no parallel "persisted message" type that the log is converted to and from. The backend also had to be swappable — a file store now, a database store later — behind one interface.
+
+## Decision
+
+Persistence is an abstract **capability seam** ([capability seams](2026-06-13-capability-seams.md), the `dsh-bash` template), not loop or core logic:
+
+1. **Interface** (`dsh-session-persistence`, `ctx.sessionPersistence`) — an abstract `SessionPersistence` service: `create`/`append`/`load`/`list`. Its persisted unit IS the existing `SessionEvent` (`{ type, seq, time, data }`), reused verbatim — no conversion type.
+2. **Implementation** (`dsh-session-persistence-jsonl`) — an append-only JSONL log per session (a `SessionHeader` line then one `SessionEvent` per line, verbatim **including `assistant/chunk`**).
+
+Key choices recorded here because they are durable, contested, and surprising:
+
+- **The canonical durable log persists every `SessionEvent` verbatim, including `assistant/chunk`.** `deriveMessages()` skips chunks, and a chunk-filtered rollout (Codex's `policy.rs`) is tempting — but `seq = log.length` and the load-validation `events[i].seq === i` require a *contiguous* log; filtering chunks out would leave holes and break both the contract and resume. A chunk-filtered projection is possible later as a derived view with its own renumbering, but it is NOT the canonical log.
+- **Append-only; a crashed turn is closed, never truncated.** Events through a flushed `turn/end` are never rewritten, and the loop flushes only at turn end. Because one interrupted turn may contain substantial valid work, `load` preserves its contiguous, parseable events and appends error results for unanswered tool calls, a missing `step/end`, and `turn/end` with `{ kind: 'interrupted' }`. The synthetic results keep resumed provider transcripts valid. Only an incomplete final record is discarded; a parse error or sequence gap at or before the last real `turn/end` is corruption and makes the session unloadable.
+- **File backend canonical, DB backend a proven drop-in.** `SessionEvent` maps 1:1 onto a row `(session_id, seq, type, time, data)` — `append` is INSERT (in a transaction asserting the contiguous-seq contract), `load` is SELECT … ORDER BY seq. `dsh-session-persistence-sqlite` is exactly this: a `SessionPersistence` subclass with no interface change (opencode runs this exact shape on SQLite/WAL), and it passes the same `runPersistenceContract` suite as the JSONL backend — so the contract holds both backends to identical semantics (lazy materialization, interrupted-turn close on load, contiguous-seq), expressed once over file bytes and once over rows.
+- **Metadata is out-of-log.** Format version, cwd, and lineage are storage concerns, not replayable conversation state, so they live in a `SessionHeader` owned by `dsh-session` and attached to a `Session` via a new readonly `session.header` — never in `SessionEventMap`, never reaching `deriveMessages()`. The alternative (a merge-extensible `session/meta` event as log line 0) was rejected: an in-log event would ride along with a seeded/forked session for free, but metadata is not replayable state, so the explicit out-of-log header seam is the cleaner cost. (The header was originally split into an immutable `SessionHeader` plus a mutable `SessionSummary` whose union was `SessionMeta`; the mutable summary was later removed as dead state — see [Drop the mutable session summary](../simplification/2026-06-19-drop-mutable-session-summary.md).)
+- **`ctx.agents.create()` and `ctx.agents.resume()` are async factories; resume additionally crosses the persistence boundary.** `ctx.agents.resume({ resumeSessionId })` awaits `ctx.sessionPersistence.load`, recreates the live session with the loaded events (so `lastTurnNumber`/`deriveMessages` continue), and registers the fresh agent under the exact resumed id. The agent-loop does NOT hard-inject `sessionPersistence` (that would pend non-persistent demos forever); `resume` rejects with a clear error when it is absent.
+
+## Alternatives considered
+
+Each key choice above records its rejected alternative where the choice is stated: a **chunk-filtered canonical log** (Codex's `policy.rs` shape) — breaks the contiguous-seq contract; **truncating a crashed turn** — silently destroys a long autonomous run's real work; an **in-log `session/meta` event as line 0** — metadata is not replayable state; **hard-injecting `sessionPersistence` into the loop** — would pend non-persistent demos forever.
+
+Format versioning: the header carries a `version`; `load` rejects any non-current version (no migration — the pre-release session format is pinned at `SESSION_FORMAT_VERSION = 0` and absorbs shape churn, per the AGENTS.md pre-release stance). Stated honestly: append-only + flush is robust to partial trailing writes (tolerated on load) but not to fsync-less power loss mid-line; a DB/WAL backend is the stronger option later.
+
+## Consequences
+
+Two new packages and the metadata seam in `dsh-session` (`session.header`, the `create(id?, options?)` signature). Bought: durable resume/fork, a read/replay path, crash tolerance, and the foundation the ACP `session/load` ([ACP support](../feature/2026-06-14-acp-agent-client-protocol.md)) needs — all over the existing event-sourced log, with the backend swappable behind one interface. The reusable `runPersistenceContract` suite holds every backend to the same append-only, contiguous-seq, lazy-materialization, and serializability semantics. Persisting the full log also settles event fidelity: `assistant/chunk` remains verbatim.

.agents/notes/implemented/architecture/2026-06-15-turn-enclosure-invariant.md

@@ -1,10 +1,8 @@
-# RFC: Every session event is enclosed in a turn
+# Agent Note: Every session event is enclosed in a turn
 
-Status: implemented (accepted 2026-06-15)
+Status: implemented
 
-<!-- XXX: legacy ADR/RFC body format, not yet normalized to a unified RFC template. -->
-
-## Context
+## Problem
 
 A durable session-persistence backend (added in a companion change) uses the **turn** as its crash-recovery boundary: a crash can leave an unclosed final turn, which `load` closes with a synthetic `turn/end {kind:'interrupted'}` while preserving the turn's real events (see [session persistence](2026-06-14-session-persistence.md)). This recovery is only well-defined if nothing *legitimately* durable sits OUTSIDE a turn — between the last `turn/end` and the next `turn/start` — since such an event would be swept into the next turn's interrupted close.
 
@@ -15,20 +13,22 @@ That assumption did not hold. Two paths recorded events outside any turn:
 
 In case 2, if the injected `context/message` is the last event before a flush/dispose (no later turn appends a `turn/end`), `scanLog` treats it as crash debris and **drops it on resume** — the injected context is durably on disk but silently lost on reload. Case 1 was benign in isolation (a `user/message` is always followed by the turn it triggered) but made the "what may appear outside a turn" rule fuzzy.
 
-Two ways to fix it: relax the *reader* (let `scanLog` commit events that sit outside an open turn), or constrain the *producer* (make every event turn-enclosed so the reader's simple "last `turn/end`" rule is both correct and complete). We chose the producer-side invariant: a single, checkable rule beats a more permissive boundary scan that has to reason about partial turns *and* loose between-turn events.
-
 ## Decision
 
 **Every session event lives inside a turn** — between a `turn/start` and its matching `turn/end`. Concretely:
 
 - The loop appends queued `user/message` events **after** `turn/start` (inside the turn), not before it. `turn/end` is therefore owed the moment those messages are recorded, and the existing finalizer guarantees it.
-- An `agent.inject()` made while the agent is **running** appends its `context/message` into the already-open turn (unchanged).
+- An `agent.inject()` made while the agent is **running** joins the already-open turn. While the current step executes assistant tool calls, accepted context waits in arrival order until that batch settles, then appends after every recorded result and before the turn closes even when execution is interrupted.
 - An `agent.inject()` made while **idle** wraps its `context/message` in a one-shot turn: `turn/start{trigger:{kind:'injection'}}` → `context/message` → `turn/end{completed}`. A new `injection` variant joins the merge-extensible `TurnTriggerMap`.
 - The loop derives the next turn number from the log each iteration (`lastTurnNumber(session) + 1`) instead of keeping a private counter, so an idle injection's one-shot turn cannot collide with the next real turn's number.
 - The `dsh-invariants` plugin **enforces** the invariant in dev: a `user/message` / `context/message` / `steering/message` appended while no turn is open throws an `InvariantError`.
 
 The serializability invariant is enforced at the same source boundary (`Session.append` throws on non-JSON-serializable data), so "what may enter the log" is now governed in one place rather than discovered downstream by whichever backend happens to be watching.
 
+## Alternatives considered
+
+**Relax the reader instead of constraining the producer** — let `scanLog` commit events that sit outside an open turn. Rejected: a single, checkable producer-side rule beats a more permissive boundary scan that has to reason about partial turns *and* loose between-turn events.
+
 ## Consequences
 
 The turn is now the *single* durability/replay boundary, so [session persistence](2026-06-14-session-persistence.md)'s crash-recovery rule is complete, not merely sufficient: an interrupted final turn is closed (with a synthetic `turn/end {interrupted}`) and its real events preserved, with zero risk of conflating between-turn context into it, because there is no between-turn context. `scanLog` stays simple (one possibly-open final turn, never a loose between-turn event), and an idle background-task notice survives persist + resume.
@@ -37,4 +37,4 @@ Costs: `agent.inject()` while idle now writes three log lines instead of one, an
 
 The rule is intentionally producer-enforced and dev-checked rather than reader-tolerated: a future backend (SQLite/WAL) inherits the same clean boundary for free, and a plugin that records an event outside a turn fails loudly in dev instead of silently losing data on the next reload.
 
-The invariant also constrains where the loop may record an `error` event. A failure detected while a turn is open is appended INSIDE the turn (before `turn/end`); but a failure that surfaces once the turn is already closed — a rejecting `session/flush` (which runs as the post-`turn/end` durability checkpoint) or a throwing `agent/turn-end` listener (after `closeTurn` already appended `turn/end`) — has no in-turn position left. Appending an `error` there would land it past the last `turn/end`, exactly the crash-tail position a backend discards. So those post-turn failures are reported via the `agent/error` event and the logger only, never as a `SessionEvent`; the turn stays balanced and persistence keeps its buffered events for the next checkpoint. If durable operational diagnostics are ever needed, they belong on a separate telemetry channel, not the replayable session log.
+Failures detected during a turn are logged before `turn/end`. A later flush failure has no valid in-turn position, so it is reported through `agent/error` and logging rather than appended as a session event. This preserves a balanced replay log; durable operational diagnostics require a separate telemetry channel.

.agents/notes/implemented/architecture/2026-06-17-filesystem-capability-seam.md

@@ -0,0 +1,158 @@
+# Agent Note: Filesystem capability seam — ctx.fs, local backend, and model-facing filesystem tools
+
+Status: implemented
+
+## Problem
+
+The harness has a concrete `bash` capability seam (`dsh-bash` / `dsh-bash-local` / `dsh-tool-bash`), but filesystem operations are about to be added as model-facing tools without an equivalent seam. If `read`, `write`, and `edit` directly use `node:fs`, the model-facing tool package will own filesystem execution policy, local path resolution, atomic write behavior, text decoding, symlink behavior, and edit semantics all at once.
+
+That couples three concerns that change independently:
+
+1. The filesystem contract: what operations plugins can ask for.
+2. The backend: local disk now, sandboxed/remote/project-scoped filesystem later.
+3. The consumer surface: model-facing `read` / `write` / `edit` schemas and result formatting.
+
+Without a `ctx.fs` interface, swapping local filesystem access for a sandboxed or remote backend would churn the tool schemas, demos, and prompt guidance even when the model-facing contract should stay stable. It also makes permission/sandbox boundaries harder to reason about: a `cwd` option can look like a sandbox even though it is only a base path unless an explicit backend or `tools/execute` policy enforces containment.
+
+We need the filesystem tools to land in the same capability-seam shape as bash before they become a public package surface.
+
+## Decision
+
+Filesystem access is a first-class capability seam following [the capability-seam Agent Note](2026-06-13-capability-seams.md):
+
+1. `@deepseek-ai/dsh-fs` (`packages/fs/fs`) owns the abstract `ctx.fs` service, the filesystem vocabulary types, and the `fs/*` policy event vocabulary.
+2. `@deepseek-ai/dsh-fs-local` (`packages/fs/fs-local`) provides the first implementation, backed by the local filesystem.
+3. `@deepseek-ai/dsh-tool-fs` (`packages/fs/tool-fs`) provides the model-facing `read`, `write`, and `edit` tools over `ctx.fs`, and is the executor that dispatches the `fs/*` events.
+
+The consumer package depends only on the interface package, never on `dsh-fs-local`. A deployment that wants a different backend loads a different provider for `ctx.fs` without changing the tool schemas or model-facing prompt guidance.
+
+The read-before-write/edit and observed-state policy is a fourth package, `@deepseek-ai/dsh-fs-policy` (`packages/fs/fs-policy`), contributed through the `fs/*` event gate rather than living on `ctx.fs`; a deployment loading `dsh-tool-fs` also loads `dsh-fs-policy` to get read-before-write/edit. This Agent Note established the three-package seam; the split of policy off the provider base class is decided by [the split-fs-seam Agent Note](../simplification/2026-06-26-fsspec-style-fs-seam.md), and its realization as an event-gate plugin (not a method service) by [the event-gate Agent Note](2026-06-26-file-context-as-event-gate.md). This document is updated to describe that landed four-package shape.
+
+The first backend is deliberately local-only: `dsh-fs-local` implements `ctx.fs` against the host filesystem. Future sibling backends can provide sandboxed, remote, virtual, or project-scoped filesystems behind the same interface.
+
+The first consumer is deliberately text-file-only: `dsh-tool-fs` exposes model-facing `read`, `write`, and `edit` tools for UTF-8 text files. Future consumers can add directory listing, search/glob, binary-safe operations, file watching, or higher-level project operations without changing the local backend package, as long as the needed capability exists on `ctx.fs`. Direct directory listing was later added by [Add direct directory listing to the filesystem seam](2026-07-03-filesystem-directory-listing-seam.md).
+
+Filesystem permissions and sandboxing are not implied by this split. The local backend resolves relative paths from its configured base directory, but containment policy is a separate decision: either a stricter `ctx.fs` implementation enforces it, or a permission/sandbox plugin wraps `tools/execute` and vetoes calls before they reach the consumer.
+
+Read-before-write/edit and observed state belong to `dsh-fs-policy`, not `ctx.fs`. Through the `fs/*` event gate, the policy records versions per opaque actor and supplies optional mutation expectations; the provider enforces freshness atomically. `dsh-tool-fs` emits the events without depending on the policy. See the [split-seam](../simplification/2026-06-26-fsspec-style-fs-seam.md) and [event-gate](2026-06-26-file-context-as-event-gate.md) Agent Notes.
+
+## Package topology
+
+The filesystem seam uses the same dependency direction as the bash trio:
+
+```text
+@deepseek-ai/dsh-tool-fs  --depends on-->  @deepseek-ai/dsh-fs  <--depends on--  @deepseek-ai/dsh-fs-local
+        consumer                                interface                         implementation
+```
+
+`@deepseek-ai/dsh-fs` depends only on `cordis` plus the repo-wide `HarnessError` base from `@deepseek-ai/dsh-llm`. It declares the `ctx.fs` key, the abstract `FileSystem` service, the vocabulary types shared by backends and consumers, the filesystem error vocabulary, and the `fs/*` policy event vocabulary. It carries no observed-state store and no owner-derivation shape; the events pass an opaque `object` actor that the provider never reads, and the `dsh-fs-policy` plugin owns the owner-derivation shape and the observed-state store on top of those events.
+
+`@deepseek-ai/dsh-fs-local` depends on `@deepseek-ai/dsh-fs` and `cordis`. It subclasses `FileSystem`, registers itself as `ctx.fs`, owns local-backend configuration such as the base directory, and contains all direct `node:fs` / `node:path` access. It holds no observed-state store — freshness is a version token the backend mints and the policy plugin records.
+
+`@deepseek-ai/dsh-tool-fs` depends on `@deepseek-ai/dsh-fs`, `@deepseek-ai/dsh-tools`, `@deepseek-ai/dsh-system-prompt`, and `cordis`. It registers model-facing tools and prompt sections. It must not import `node:fs`, `node:path`, or `@deepseek-ai/dsh-fs-local`; filesystem execution always goes through `ctx.fs`. If the implementation needs concrete agent or session helper types, those dependencies belong in `tool-fs`; they must not leak back into `dsh-fs`.
+
+The root `tool-fs` plugin registers the full filesystem tool suite (`read`, `write`, and `edit`) by composing the per-tool registration helpers. It injects `fs` and never imports an implementation package.
+
+## `ctx.fs` contract
+
+`@deepseek-ai/dsh-fs` owns a semantic filesystem service. It is higher-level than `readFile` / `writeFile` so `tool-fs` does not reimplement path resolution, versioning, text decoding, binary rejection, pagination, atomic replacement, symlink behavior, or literal edit semantics.
+
+The interface covers these semantic operations:
+
+- Resolve a model/plugin-supplied path into a backend-defined target.
+- Stat target metadata without reading file contents.
+- Read a bounded UTF-8 text page from a target.
+- Create or replace a UTF-8 text file.
+- Edit an existing UTF-8 text file by literal replacement.
+
+The provider seam also carries the freshness hooks that policy builds on — but the observed-state store and owner derivation live in the `dsh-fs-policy` plugin, not on `ctx.fs`:
+
+- The backend mints an opaque `version` token per target (in `stat` and in every read/mutation outcome).
+- `writeText`/`editText` take an OPTIONAL version expectation: omit it for an unconditional bare-provider mutation, or supply it to guard the mutation inside the backend's atomic critical section.
+- The `dsh-fs-policy` plugin decides that expectation on `fs/write-intent`/`fs/edit-intent` and records observed versions on `fs/observed`, keyed by an owner it derives from the opaque event actor (normally `exec.agent.session`).
+
+Authorization is version freshness, not a full/partial view distinction: any read records the target's version, and a later write/edit is authorized as long as the file is still at that version — so a windowed read of lines 100-150 authorizes an edit of line 120. The observed-state store is a `WeakMap<owner, Map<targetKey, version>>` inside `dsh-fs-policy`; `dsh-fs` holds none of it and treats the actor as opaque. (This Agent Note first modeled a `FileState` cache with `full`/`partial` views on `ctx.fs`; the split-fs-seam and event-gate Agent Notes replaced that with the freshness-based policy plugin described here.)
+
+Path resolution is explicit and allowed to be async. Local resolution may only normalize a path, but sandboxed/remote/project-scoped backends may need I/O to resolve a user-supplied path into a stable target identity.
+
+Resolved targets must expose at least three concepts:
+
+- The original input path, for diagnostics.
+- An opaque `targetKey`, used for stale guards and file-state lookup. The local backend might use a realpath-like key; a remote backend might use a workspace URI or file id. Consumers must not parse or assume this is a local absolute path.
+- A `displayPath`, used for model/UI-facing output. It may be a local absolute path, workspace-relative path, or remote URI depending on the backend.
+
+Read and mutation results must include an opaque file `version`. The local backend derives its token from bigint stat metadata (`dev`, `ino`, `size`, `mtimeNs`, and `ctimeNs`) so same-size rewrites and inode replacement invalidate consumers reliably; a remote backend can use a revision id or hash-like token. The `dsh-fs-policy` plugin records versions for stale checks; consumers may display related metadata but must not interpret the version token.
+
+The provider hands back decoded text: `readText` returns a whole regular text file, `streamText` streams the same text semantics for large files. Both own regular-file checks, bounded line/output handling is NOT theirs — line windowing, numbered-line rendering, and total-line accounting live in the executor (`dsh-tool-fs`), which reads through `ctx.fs` and renders the model-facing window. The provider owns UTF-8 decoding and binary/NUL rejection; it does not know about line windows or views.
+
+Observed-state recording is not on `ctx.fs`: after a successful read the executor emits `fs/observed`, and the `dsh-fs-policy` plugin records `{ version }` for the deriving owner. There is no `full`/`partial` view — a read at any window records the version, and freshness (not view completeness) authorizes a later write/edit.
+
+Full-file writes create or replace UTF-8 text files. Backends may create parent directories when that behavior is supported and documented. Existing non-regular targets are rejected. `writeText` takes an optional expectation: `createIfAbsent` creates a missing target and rejects an existing one with `FS_NOT_OBSERVED` (the path the policy uses for an unobserved owner); `replaceIfVersion` replaces only when the target exists at the observed version, else `FS_STALE_VERSION`; omitting the expectation is the unconditional bare-provider create-or-overwrite. The policy plugin chooses which expectation to supply from the owner's observed state.
+
+Literal edit is a provider primitive (`editText`), not composed in `tool-fs` from a read plus write. Literal matching, duplicate-match rejection, CRLF preservation, binary rejection, optional stale-version checking, and atomic read-modify-write must stay together inside the backend's mutation critical section. `editText` takes the same optional version expectation; the stale check runs before literal matching so an edit against an old read reports `FS_STALE_VERSION`. A remote backend may implement edit as a native compare-and-edit operation; the consumer does not force local-style composition.
+
+The policy plugin, not `ctx.fs`, gates on prior observation: an `edit` requires a prior observation by the owner (else `FS_NOT_OBSERVED`), and the recorded version is passed to `editText` as the CAS basis. With the policy plugin absent, `ctx.fs` alone is a complete unconstrained seam (unconditional write/edit); the tool is never method-coupled to the policy.
+
+Filesystem contract failures are thrown as `FsError extends HarnessError`, and the tool registry converts them into `isError` tool results with structured `{ name, code }` metadata. `dsh-fs` owns this vocabulary rather than each tool inventing messages. The codes are `FS_NOT_FOUND`, `FS_NOT_TEXT`, `FS_STALE_VERSION`, `FS_NOT_OBSERVED`, `FS_NOT_REGULAR_FILE`, `FS_AMBIGUOUS_EDIT`, `FS_EDIT_NOT_FOUND`, and `FS_ABORTED`. (An earlier draft included `FS_PARTIAL_OBSERVATION`; freshness-based authorization has no partial/full distinction, so it was dropped. Directory-listing-specific codes were added later by [Add direct directory listing to the filesystem seam](2026-07-03-filesystem-directory-listing-seam.md).)
+
+## Tool consumer behavior
+
+`@deepseek-ai/dsh-tool-fs` is the model-facing consumer. It owns tool names, JSON schemas, argument validation at the model boundary, prompt sections, and result formatting. It does not own filesystem execution.
+
+The first tool suite contains:
+
+- `read`: inspect a UTF-8 text file and return line-numbered content with pagination guidance.
+- `write`: create or fully replace a UTF-8 text file.
+- `edit`: update an existing UTF-8 text file by replacing literal text, requiring a unique match by default and allowing an explicit replace-all mode.
+
+Each tool follows the same execution shape:
+
+1. Validate and normalize model arguments.
+2. Call the appropriate `ctx.fs` operation.
+3. Format the result as `ContentBlock[]` for the model.
+4. Let thrown backend/tool errors flow through `ToolRegistry.execute()`, which converts them into `isError` tool results.
+
+The package registers prompt guidance through `ctx.systemPrompt.section(...)` and registers schemas through `ctx.tools.register(...)`. Tool schemas still flow into the normal prompt assembly path via `SystemPrompt.assemble()` and `ToolRegistry.schemas()`; no agent-loop changes are required.
+
+The tool package keeps model-facing contracts stable when backends change: a local backend and a remote backend may resolve paths differently internally, but the `read` / `write` / `edit` schemas do not change solely because the backend changes.
+
+The default deployment requires a prior `read` before updating an existing file with `write` or `edit`. `tool-fs` does not implement this by checking whether a tool named `read` ran: it dispatches the `fs/write-intent`/`fs/edit-intent` events (passing the execution context as the opaque actor), and the `dsh-fs-policy` plugin derives the owner, gates on prior observation, and supplies the version expectation. Any windowed read authorizes a later write/edit as long as the file is unchanged. Creating a new file with `write` does not require prior observation.
+
+The root plugin registers the full suite by composing the per-tool registration helpers. It injects `fs`, `tools`, and `systemPrompt`.
+
+## Testing
+
+Tests follow the package boundary, not only the user-visible tools: the service seam in `dsh-fs`; real filesystem behavior through the `ctx.fs` interface in `dsh-fs-local` (resolution, symlinks, streaming, binary/UTF-8 rejection, unconditional and version-guarded writes, literal-edit semantics, line-ending preservation, structured `FsError` codes); the consumer surface in `dsh-tool-fs` against the real local provider (mock only the model/clock, never the collaborator); and integration through `ctx.tools.execute()` with and without `dsh-fs-policy`, world-verified by reading files back from disk rather than trusting the returned `ContentBlock[]`. The observed-state/owner-derivation policy is tested in `dsh-fs-policy`, not here.
+
+The defensive-pattern classes this repo has been bitten by are pinned directly:
+
+- **Atomic-write temp-file safety.** Write/edit stage through a private random `0700` directory next to the target with an exclusive owner-only (`'wx'`, `0o600`) temp file, cleanup on failure, and a final atomic rename — mirroring the bash spill-file rules, because predictable world-readable temp paths invite symlink races and disclosure. Tests assert the permissions and that a pre-existing temp path is not clobbered; this primitive is a standing requirement of the seam.
+- **`targetKey` identity through symlinks.** Two input paths resolving to the same realpath share one observed-state entry: a `read` via path A satisfies the read-before-edit guard for an `edit` via symlink path B, and a stale write through one path is detected through the other.
+- **Concurrency / stale races.** Two concurrent write/edit operations against the same target settle deterministically — one succeeds, the other is rejected with `FS_STALE_VERSION` — and a successful edit refreshes recorded state so the same owner's next edit proceeds.
+- **HMR safety and disposal.** Disposing the backend's fiber withdraws the `ctx.fs` provider; a later provider starts with no inherited state.
+
+## Alternatives considered
+
+- **Model-facing tools directly over `node:fs`** — the tool package would own execution policy, path resolution, atomic writes, text decoding, and edit semantics at once, coupling the three independently-changing concerns the Problem names and churning schemas on any backend swap.
+- **One combined `dsh-fs-tools` package** — the pre-seam shape; rejected for the same interface/implementation/consumer split as bash, and the combined name never became public surface.
+- **Observed-state on `ctx.fs`** — the shape this Agent Note first landed; superseded by [the split-fs-seam Agent Note](../simplification/2026-06-26-fsspec-style-fs-seam.md) and [the event-gate Agent Note](2026-06-26-file-context-as-event-gate.md): a sandboxed/remote backend must not inherit model-facing observation policy, so the provider keeps only the version token and the optional version-guarded mutation.
+
+## Consequences
+
+**`cwd` can be mistaken for a sandbox.** The local backend's base directory is a resolution default, not automatically a containment boundary. If containment is required, it must be enforced by the backend contract or by a permission/sandbox plugin on `tools/execute`.
+
+**The interface can become too local.** Returning fields such as `absolutePath` from `ctx.fs` would make remote, sandboxed, or virtual backends awkward. The contract should expose display metadata without requiring consumers to understand host paths.
+
+**The interface can become too thin.** If `ctx.fs` only mirrors `node:fs` primitives, `tool-fs` will reimplement binary detection, pagination, atomic writes, and edit semantics. That recreates the coupling this Agent Note is trying to avoid.
+
+**Edit semantics are race-prone by nature.** Literal edit is a read-modify-write operation; the guard is the backend's atomic mutation critical section plus the optional version expectation, so concurrent edits settle deterministically — one wins, the other gets `FS_STALE_VERSION`.
+
+**Observed state does not belong on `ctx.fs`.** Recording what an execution context has seen is workflow policy, not raw filesystem I/O. This Agent Note first placed it inside the filesystem seam; the split-fs-seam Agent Note then established that a sandboxed/remote backend should not inherit model-facing observation policy, and moved it into the `dsh-fs-policy` plugin. The provider seam keeps only what write/edit safety genuinely needs at the storage layer — a backend-minted version token and an optional version-guarded mutation — while the policy plugin owns owner derivation, observed-state, and read-before-edit gating over the `fs/*` events.
+
+**The `resolve`-then-operate shape costs an extra round-trip per call.** Each tool may resolve a path to an `FsTarget` and then issue the read/write/edit as a separate `ctx.fs` call. For the local backend this is negligible (resolution is in-memory path normalization), but a remote/sandboxed backend may turn each step into its own request, so a single `read` can become two network round-trips. Backends where the round-trip matters can cache or fold resolution internally while preserving the observable contract.
+
+**Observed-state persistence is deferred.** Observed state lives in memory (the `WeakMap` inside `dsh-fs-policy`), so a resumed session conservatively requires files to be read again before write/edit until a future session-event or persistence mechanism makes observation replayable.
+
+**Error codes become part of the seam.** `FsError` codes make stale-version and observation failures machine-routable through the existing structured error taxonomy. The cost is that `dsh-fs` imports the shared `HarnessError` base from `dsh-llm`; that dependency is intentional and stays limited to the error vocabulary.
+
+**Package churn is front-loaded.** The three-package split adds boilerplate before there is more than one backend. This is intentional: filesystem access is a likely sandbox/remote boundary, and changing the package surface after shipping model-facing tools would be more expensive.

.agents/notes/implemented/architecture/2026-06-18-agent-lifecycle-and-ownership-seams.md

@@ -0,0 +1,48 @@
+# Agent Note: Agent lifecycle and ownership seams
+
+Status: implemented
+
+## Problem
+
+Several ACP and tool-bash limitations were symptoms of the same missing seam: plugins could create or resume agents through `ctx.agents`, but they could not own and dispose one agent independently, and long-running bash tasks carried no stable owner in the executor itself. ACP aborted and awaited agents on disconnect but could not unregister just that session's agent; `session/cancel` could not cancel queued-but-not-yet-started work; and `tool-bash` kept task ownership in a plugin-local `Map`, so an HMR reload could make an old task look unowned.
+
+## Decision
+
+Three seams: the queue-aware cancel, the `AgentHandle` disposer, and the bash owner token.
+
+### 1. Queue-aware `Agent.cancel(reason?)`
+
+A new `cancel()` verb on the `Agent` interface — the single public stop primitive. (It originally shipped alongside a narrower step-only `abort()`; that verb was later removed as unused, leaving `cancel()` the only public way to stop work.) It clears the inbox's queued + steering FIFOs, aborts the in-flight step if any, and drives a **turn-scoped cancellation marker** the driver loop checks at every turn-decision point — so a prompt that is queued-but-not-yet-started never runs, a cancel landing in the pre-step / continuation window drops the about-to-run turn (ending it `aborted`), and a later prompt cannot be batched into the cancelled turn. `whenIdle()` reaches post-cancel quiescence. ACP `session/cancel` maps to `cancel()`. The marker is armed ONLY when there is something to cancel, so an idle no-op cancel cannot strand the next prompt.
+
+### 2. `AgentHandle` async disposer
+
+`ctx.agents.create`/`resume` (and the `AgentFactory` interface) return `AgentHandle = { agent: Agent; dispose(): Promise<void> }`. The disposer is a **consumer capability** — a registry observer holding only the bare `Agent` cannot tear it down. The caller fiber and registered factory provider are structural co-owners: caller unload enforces structured ownership, while provider unload must stop old instances whose scoped dependency surface resolves through that provider. All three paths reach the same memoized teardown: stop the loop, await its exit and idle flushes (true quiescence, not just the `disposed` status flip), detach the agent, detach its session, and unwind its scope. Each public ID becomes reusable when its exact registry entry detaches; there is no separate reservation-release phase. Config-created agents are already owned by the `AgentLoop` fiber (the handle is discarded). ACP holds each session's disposer in its `SessionRecord` and runs it on disconnect/teardown, so a bare client disconnect leaves no registered agent and no session-store entry — even when `session/load` races teardown (the just-resumed handle is disposed before the closed-guard throw).
+
+**Teardown ORDER is load-bearing for durability**, and the implementation folds the session lifecycle into the agent's SINGLE composite cordis effect (`SessionStore.prepare`/`enter`/`announce`, replacing a sibling-effect split). A fiber unload disposes sibling effects concurrently (`Promise.all`), which would race removing the session store's append publication hooks against the loop's closing `session/flush` and drop the closing `turn/end`; inside one effect the disposers run as an ordered LIFO chain (loop stopped + `await agent.done` BEFORE the session detaches), so the loop's final flush is captured on BOTH the handle's `dispose()` and a fiber unload. The contained `agent/disposed` and `session/disposed` notifications cannot reject the chain or skip later teardown.
+
+### 3. Bash owner token in the seam
+
+Background-task ownership moved from a `tool-bash` plugin-local `Map<string, Agent>` into the executor. `BashExecRequest` gains an optional `owner?: string`; the resolved `BashExecSpec` carries it as required-but-nullable `owner: string | undefined` (a forgotten owner is a visible `undefined`, never a silently-absent property). The executor stores the token on its task and exposes it via a new `BashExecutor.ownerOf(id): string | undefined` seam (NOT on the public `BashTask` — one read path, no redundant API). `tool-bash` deletes its `Map` entirely: it stamps `exec.agent?.id` (the shared registry/session id) as the owner at `start`, and `bash_output`/`bash_kill` compare `ctx.bash.ownerOf(id)` to the caller's token with `!== undefined` semantics (an empty-string token is still a real owner). The completion notice finds the live agent by scanning `ctx.get('agents')?.list()` for `agent.id === ownerToken` (read via `ctx.get` — `onTaskDone` runs on the bash fiber, a foreign fiber, where the `ctx.agents` proxy would throw). Because ownership now lives on the task in the executor (disposed with the `dsh-bash` fiber), it SURVIVES a `tool-bash` HMR reload — closing the old `XXX(tool-bash-owner-hmr)` gap. (The `onTaskDone` listener is still effect-scoped to `tool-bash`'s `apply`, so a completion landing during the reload gap still drops its one notice — the pre-existing reload-gap drop — but the ownership fence itself is HMR-proof.)
+
+## Verification
+
+These invariants hold and are pinned by tests:
+
+- ACP disconnect/session close leaves no registered agent AND no session-store entry for that session, even when `session/load` races teardown.
+- `session/cancel` before a queued prompt starts prevents that prompt from running and cannot batch the next prompt into the cancelled turn.
+- A `tool-bash` HMR reload does NOT make an existing background task readable or killable by a different session (ownership survives on the executor).
+- Existing non-ACP demos still work without managing handles explicitly; config-created agents remain owned by the `AgentLoop` plugin fiber.
+
+## Session owner tokens are unique among live agents
+
+The bash owner-token comparison relies on the shared `Agent.id`/`SessionId` being unique among live agents. Concurrent same-ID operations may both prepare privately, but publication enters the session and agent in order; `SessionStore.enter()` rejects a duplicate live session id, and every losing transaction rolls its private state back. A programmatic caller therefore cannot publish two live agents with one session token. The access *policy* (token comparison) stays in `tool-bash` (the consumer); the bash seam stores only an opaque `owner` string and never interprets it — the correct interface/implementation/consumer split.
+
+## Alternatives considered
+
+- **A public `BashTask.owner` field** instead of the `BashExecutor.ownerOf(id)` seam — rejected: one read path, no redundant API.
+- **Sibling cordis effects for the agent's session lifecycle** — rejected: a fiber unload disposes sibling effects concurrently (`Promise.all`), racing removal of the store-owned append publication hooks against the loop's closing `session/flush`; the single composite effect's ordered LIFO chain is what captures the closing `turn/end` on both disposal paths.
+- **A separate step-only `abort()` beside `cancel()`** — shipped originally, then removed as unused; `cancel()` is the single public stop primitive ([the public-stop-surface Agent Note](../simplification/2026-06-20-public-agent-stop-surface.md)).
+
+## Consequences
+
+This touched public interfaces (`Agent`, `AgentFactory`, the bash seam) deliberately, not as a local ACP patch. The simple synchronous `Agent.send()` ergonomics were preserved; the async lifecycle path is additive, for owners that need it.

.agents/notes/implemented/architecture/2026-06-18-session-surface.md

@@ -0,0 +1,70 @@
+# Agent Note: Session surface — an ordered projection over the event log
+
+Status: implemented
+
+## Problem
+
+The event log is authoritative, but history manipulation had no durable shared mechanism. Plugins such as compaction would otherwise rewrite derived requests through order-sensitive listeners, leave no provenance, and require repeated changes to `deriveMessages()`.
+
+## Decision
+
+Add a **surface** — a derived, cached order of event sequences (the subset of events that produce LLM messages) — maintained by `surfaceOp` markers in the event log.
+
+### Two new top-level fields on `SessionEvent`
+
+Every `SessionEvent` gains two optional fields (structural metadata, like `seq`/`time`):
+
+- **`sourceEventSeqs?: number[]`** — seq numbers of events that are provenance sources (e.g., the `assistant/chunk` seqs that built an `assistant/message`, or the surface nodes shadowed by a compaction marker). A present `[]` is valid only on `assistant/message` and records a known empty provider stream; omission there means legacy or otherwise unrecorded provenance. Other surface events require a non-empty list when the field is present. Provenance is a core design principle; without it, the replace-range operation cannot be validated on replay.
+- **`surfaceOp?: SurfaceOp`** — how this event entered the surface. Absent for non-surface events.
+
+### SurfaceOp: two operations
+
+```ts
+export type SurfaceOp =
+  | 'append'                                    // normal tail append
+  | { op: 'replace'; start: number; end: number }  // shadow [start, end] inclusive
+```
+
+1. **Append** — add the new event seq to the tail. Used by `user/message`, `assistant/message`, `tool/result`, `context/message`, `steering/message`. The loop passes `surfaceOp: 'append'` on all such appends and records `sourceEventSeqs` where applicable: every successful `assistant/message` records its complete `assistant/chunk` source set, including `[]`, while `tool/result` records its `tool/call` source.
+
+2. **Replace** — remove entries from `start` through `end` (both inclusive) and insert the new event seq in their place. Both `start` and `end` must be present in the current surface; `start === end` replaces one entry. The event's `sourceEventSeqs` must contain every shadowed surface seq. The shadowed events remain in the log but are no longer on the surface.
+
+### SurfaceManager: delta-based, not full rebuild
+
+A `Session` owns one `SurfaceManager` that maintains an ordered `number[]` of event seqs. The manager validates each seed or append candidate without applying it before commit, then processes only committed events since its previous synchronization rather than rescanning the entire log. `Session.surface` exposes the same manager through the readonly `SessionSurface` contract, so acceptance, derived history, compaction, and workspace context share one incremental state. Replace locates its inclusive endpoints by array position and splices the replacement seq into that range; no second manager, link objects, or seq-to-node map duplicates the order.
+
+Delta processing is O(1) when no new events and O(new events) when new events arrive.
+
+`deriveMessages()` uses the surface when surface markers exist, falling back to the existing linear scan for sessions without markers (backward compatibility).
+
+### Persistence
+
+The new fields are serialized as top-level JSON properties. The JSONL backend requires zero changes — `JSON.stringify`/`JSON.parse` preserve everything transparently. The SQLite backend's `events` table carries two nullable TEXT columns (`source_event_seqs`, `surface_op`). The on-disk `SCHEMA_VERSION` is bumped to reflect the column set, and — per the pre-release bump-and-reject policy — a database written by any other build is REJECTED on open rather than migrated (there is no persisted user data to upgrade). The session format `version` is pinned at `SESSION_FORMAT_VERSION = 0` (the "unstable / pre-release" stance): the optional surface fields are absorbed without bumping it.
+
+### Crash recovery
+
+The `repair.ts` module synthesizes `tool/result` closers for orphaned tool calls after a crash. These closers carry `surfaceOp: 'append'` and `sourceEventSeqs` pointing to the orphaned `tool/call` event, so the rehydrated surface is valid.
+
+### Invariants
+
+The dev-mode invariants plugin validates: `sourceEventSeqs` references (only `assistant/message` may use an empty list; otherwise no duplicates, references earlier events, and references known seqs) and `surfaceOp` (replace `start ≤ end`, both endpoints are on the tracked surface, the range is non-reversed in surface position, and `sourceEventSeqs` includes every node the range shadows).
+
+Every surface-eligible event must carry `surfaceOp` or it would disappear from derived history. Typed `append` overloads enforce this for literal event types; runtime checks in `append` and the seed constructor cover widened unions and loaded logs. Invalid seeds are rejected rather than upgraded under the pre-release format policy.
+
+## Alternatives considered
+
+- **Per-plugin `agent/request` wrapping** (the pre-surface pattern for history manipulation) — listener-ordering fragility, no durable record of what was changed, and every new manipulation forces another change to core `deriveMessages()`.
+- **Half-open `[start, endExclusive)` replace ranges** — rejected: endpoints are named by surface event seqs, and single-entry replacement (`start === end`) reads naturally with inclusive semantics.
+- **Linked node objects plus a seq map** — rejected: production did not read predecessor links, the only successor use was the next array position, and replacement already required linear `indexOf` lookup. A single seq array preserves the same asymptotic behavior with one representation to validate.
+- **Full rebuild behind a dirty flag** instead of delta processing — O(N²) over a session's lifetime: every single-event append would rescan all prior events.
+
+## Consequences
+
+- **`packages/core/session`**: `surface.ts` (`SurfaceManager`) maintains one ordered seq array for candidate acceptance and live projection; `SessionSurface` is its readonly public view. `SurfaceOp`/`SurfaceIntent` and the top-level session-event fields record how entries join it. `append()` requires a `SurfaceIntent` for surface events, `deriveMessages()` walks the surface as the sole derivation path, and `repair.ts` emits surface-aware closers. The seed constructor rejects a surface-eligible seed event missing its `surfaceOp` marker (see § Invariants).
+- **`packages/core/agent-loop`**: All surface-capable appends pass surface opts. Chunk seqs are collected for `assistant/message` provenance; `tool/call` seqs are captured for `tool/result` provenance.
+- **`packages/session-persistence/session-persistence-sqlite`**: Two new nullable TEXT columns (`source_event_seqs`, `surface_op`) on the `events` table; `SCHEMA_VERSION` bumped (bump-and-reject, no migration).
+- **`packages/support/invariants`**: Surface-related validation rules.
+- **`packages/session-persistence/session-persistence-jsonl`**: No changes required.
+- **`packages/session-persistence/session-persistence`**: Abstract interface unchanged.
+
+The surface is the foundation for future history manipulation. A compaction or tool-result-prune plugin appends one of the existing message-producing event types (a `user/message` carrying the summary, say) with `surfaceOp: { op: 'replace', start, end }` and `sourceEventSeqs` covering the shadowed entries — the new event takes the range's place on the surface while the plugin's own trace events (e.g. `compaction/start`, `compaction/end`) stay off it. Replay preserves the decision deterministically.

.agents/notes/implemented/architecture/2026-06-18-shared-persistence-write-coordinator.md

@@ -0,0 +1,44 @@
+# Agent Note: Shared persistence write coordinator
+
+Status: implemented
+
+## Problem
+
+`dsh-session-persistence-jsonl` and `dsh-session-persistence-sqlite` intentionally prove the same `SessionPersistence` contract over different storage media, but their write-path orchestration was duplicated: per-session state, `session/created` adoption, backend-specific prefix reads, write-behind buffers, serialized flush chains, HMR seeding, and dispose drains. The pure seed-prefix collision and serializability guards had already moved into the seam package; the remaining orchestration was still correctness-heavy and received the same fixes twice. A code-level diff showed the two backends were byte-identical — or same-algorithm — for ALL of it: the four maps (`states`/`buffers`/`chains`/`inits`), `installWritePath`, `initFor`, `onCreated`'s four cases, `flush`, `drain`, `serialize`, `adopt`, `adoptLivePrefix`, `assertVersion`, and the `create`/`append`/`load` skeletons. Only the storage primitives (write bytes vs. INSERT rows) differed.
+
+## Decision
+
+Extract a backend-agnostic `PersistenceCoordinator` into `dsh-session-persistence`. The coordinator owns the orchestration once; each first-party backend composes one (`new PersistenceCoordinator(ctx, this)`), implements a small `PersistenceBackend` hook interface, and delegates its four public service methods (`create`/`append`/`load`/`list`) to it.
+
+Composition, not inheritance. The coordinator is a concrete class the backend holds, not a base class the backend extends. The Agent Note's risk — "a coordinator must not make unusual backends fight an inheritance hierarchy" — is avoided: a backend exposes only the hooks; it cannot reach the coordinator's private orchestration state, and the public `SessionPersistence` service shape is unchanged, so a third-party backend MAY still implement the abstract service directly without the coordinator at all.
+
+The coordinator retires each live session from its `session/disposed` notification: it waits for that exact Session object's initialization, serializes a final drain, and then removes the owned state, buffer, and init entries. Failed drains retain their buffers for backend teardown to retry. Settled per-id chain tails remove themselves only when they are still the current tail, so a completion cannot erase a newer operation for the same id. Backend teardown unregisters the write-path listeners before awaiting all admitted retirements, remaining buffers, and chains, then closes the backend.
+
+### The hook interface (`PersistenceBackend<TornMarker>`)
+
+Six methods (five required + an optional lifecycle hook) — the only seam between the coordinator and storage:
+
+- `name` — backend label for the dispose-failure `AggregateError`.
+- `loadStored(id)` — read a stored prefix by id, scanning ANY storage scope (every JSONL cwd bucket; SQLite's id is globally unique). Used by resume/load and, via `!== undefined`, the create-collision probe.
+- `loadLive(id, cwd)` — read a stored prefix SCOPED to `cwd`. **Deliberately distinct from `loadStored`**: HMR live-adoption must only adopt a persisted log at the SAME cwd as the live session; a same-id log at a different cwd is a collision, not a resume. Collapsing the two reintroduces a cross-cwd adoption bug. SQLite ignores `cwd`.
+- `appendBatch(meta, events, isMaterialized)` — durably append a contiguous batch, lazily materializing the session ATOMICALLY when not yet materialized (the materialize-write and the first event batch must commit together — a crash between them must not leave a materialized-but-empty session; this is why there is no separate `materialize` hook).
+- `commitRepair(meta, tornMarker, closers)` — make a crash repair durable: truncate the torn tail (iff `tornMarker !== undefined`) and append `closers`. **NOT required to be atomic** — JSONL legitimately truncates-then-appends in two fsync'd steps, SQLite does DELETE+INSERT in one transaction. Used by `load` (truncate + synthetic closers) and live-adoption (truncate only, `closers = []`).
+- `list()` — list all stored metadata.
+- `close?()` — optional lifecycle teardown (SQLite closes its db handle; JSONL omits it), awaited in the dispose effect AFTER the quiescence drain so a close failure never masks a drain error.
+
+### The opaque torn marker
+
+The single design choice that keeps the seam clean: the crash-repair "where is the torn tail" token is OPAQUE to the coordinator. The coordinator computes the synthetic closers (it owns `interruptedTurnClosers` from `dsh-session`), but it only ever tests `tornMarker !== undefined` and passes the value straight back to `commitRepair` — it never inspects it. Each backend picks its own marker type: JSONL uses the byte offset to truncate to, SQLite the seq to delete from (both happen to be `number`). The JSONL backend folds its `committedBytes < buffer.byteLength` comparison INSIDE the hook so the returned marker is already `number | undefined`; without that fold the coordinator would have to know about byte lengths.
+
+## Testing
+
+The shared `runPersistenceContract` (public-API contract) keeps running for every backend. `runCoordinatorContract` (`tests/coordinator-contract.ts`) holds the write-path orchestration — adoption, HMR, collision, session and backend disposal drains, and crash-tail repair — and runs once per backend through a `CoordinatorFixture` (an in-memory reference + jsonl + sqlite). Coordinator-specific tests pin retirement map cleanup, same-id chain-tail races, failed-drain retry, and close ordering. The per-backend specs retain storage mechanics only (JSONL: path safety, fsync rollback, bucket listing; SQLite: schema version, `scanRows`, transaction rollback). A through-coordinator torn-tail→load→`commitRepair` test per real backend (via a `corruptTail` fixture hook) keeps the coordinator's torn-marker repair branch covered under the 100% per-file gate — the contract crash test only produces synthetic closers, never a torn marker, so it could not reach that branch.
+
+## Alternatives considered
+
+- **A base class the backends extend** — rejected for composition: a backend exposes only the hooks, cannot reach the coordinator's private orchestration state, and a third-party backend may still implement the abstract service directly without the coordinator at all.
+- **A wider hook surface** — each candidate hook folded away: there is no separate `materialize` hook (the materialize-write must commit atomically with the first event batch inside `appendBatch`), no separate create-collision probe (it is `loadStored(id) !== undefined`), and no coordinator pass-through for `list()` (listing needs none of the orchestration).
+
+## Consequences
+
+The coordinator adds one indirection, an opaque torn marker, and detached session-retirement tasks, but centralizes correctness-heavy orchestration previously duplicated by every backend. Session disposal remains an observe-only event, so the session owner does not await persistence retirement; the coordinator contains failures, preserves uncommitted buffers, and makes backend teardown the quiescence boundary. Its hook surface stays narrow: collision checks reuse `loadStored`, materialization stays atomic inside `appendBatch`, and listing bypasses the coordinator. New backends implement storage primitives rather than copy the event-buffer-flush lifecycle.

.agents/notes/implemented/architecture/2026-06-20-branded-ids.md

@@ -1,31 +1,31 @@
-# RFC: Branded IDs everywhere they belong
+# Agent Note: Branded IDs everywhere they belong
 
-Status: proposed
+Status: implemented
 
 ## Problem
 
-The harness already brands three identifiers — `CallId` (`packages/llm/llm/src/brand.ts`), `SessionId` (`packages/core/session/src/types.ts`), and `AgentId` (`packages/core/agent/src/types.ts`) — using the `Branded<B> = string & { readonly [BRAND]: B }` machinery and a zero-cost cast factory per type. `brand.ts` also states the governing policy: *"Branding is for IDs that cross package boundaries and could plausibly be confused; not every string needs a brand."* That policy is right; the problem is that it is only half-applied. Two gaps let a structurally-identical-but-semantically-wrong string slip through the type checker today.
+The harness brands `CallId` (`packages/llm/llm/src/brand.ts`) and the shared agent/session `SessionId` (`packages/core/session/src/types.ts`) using the `Branded<B> = string & { readonly [BRAND]: B }` machinery (owned by the type-only `@deepseek-ai/dsh-brand` package at `packages/util/brand/` — see its [README](../../../../packages/util/brand/README.md)) and a zero-cost cast factory per type. `dsh-brand` also states the governing policy: *"Branding is for ids that cross package boundaries and could plausibly be confused; not every string needs a brand."* That policy is right; the problem is that it is only half-applied. Two gaps let a structurally-identical-but-semantically-wrong string slip through the type checker today.
 
 **Gap 1 — unbranded cross-boundary IDs in the bash seam.** The background-task id is a plain `string`: `BashTask.id: string` (`packages/bash/bash/src/types.ts`), carried as `string` through the whole executor seam (`BashExecutor.get`/`ownerOf`/`readOutput`/`kill(id: string)` in `packages/bash/bash/src/index.ts`) and validated/passed as `string` by the model-facing tools (`validateTaskId`, `assertTaskAccess`, the `task_id` schema arg in `packages/bash/tool-bash/src/index.ts`). It is generated by a per-executor counter — `` `bash-${this.nextTaskId++}` `` in `packages/bash/bash-local/src/index.ts` — which gives it **exactly the same `name-N` shape as `SessionId`'s default** (`` `session-${++counter}` `` in `packages/core/session/src/index.ts`). A bash task id and a session id are trivially swappable at a call site and the compiler says nothing. This is the headline case the user asked about, and it is a model-facing id (the model passes `task_id` back to `bash_output`/`bash_kill`), so a confusion here is reachable from untrusted input.
 
-The bash **owner token** is the related sub-case: `BashExecRequest.owner?: string` and `BashExecSpec.owner: string | undefined` (`packages/bash/bash/src/types.ts`) are documented as a deliberately *opaque* isolation key, but in every live caller the value IS the owning agent's `session.header.id` (`callerToken = (exec) => exec.agent?.session.header.id` in `packages/bash/tool-bash/src/index.ts`) — i.e. a `SessionId` wearing a `string` disguise. It is compared for access control (`owner !== callerToken(exec)`), so a mismatched-but-well-typed string here is a cross-session isolation bug the type system currently cannot catch. This is the same `session.header.id`-as-owner alias that the [unify-the-agent-id-and-the-session-id](../simplification/2026-06-20-unify-agent-and-session-id.md) proposal calls the "bash owner-token alias hole".
+The bash **owner token** is the related sub-case: `BashExecRequest.owner?: string` and `BashExecSpec.owner: string | undefined` (`packages/bash/bash/src/types.ts`) are documented as a deliberately *opaque* isolation key, but in every live caller the value IS the owning agent's shared `Agent.id`/`SessionId` (`callerToken = (exec) => exec.agent?.id` in `packages/bash/tool-bash/src/index.ts`) wearing a different seam-local name. It is compared for access control (`owner !== callerToken(exec)`), so a mismatched-but-well-typed string here is a cross-session isolation bug the type system currently cannot catch. This is the shared id alias covered by the [unified agent/session identity decision](../simplification/2026-06-20-unify-agent-and-session-id.md).
 
-**Gap 2 — brand erosion at the seams of the *already-branded* IDs.** Even `CallId`/`SessionId`/`AgentId` decay back to bare `string` at exactly the places confusion is most likely: the registry/store `Map` key types and most public method params. Representative sites: `SessionStore.store = new Map<string, Session>()` and `create`/`prepare(id?: string)`/`get(id: string)` (`packages/core/session/src/index.ts`); `AgentRegistry.store = new Map<string, Agent>()` and `register`/`get(id: string)` (`packages/core/agent/src/index.ts`); `ToolPresenter.pending = new Map<string, …>()` keyed by call id and `call(callId: string)`/`result(callId: string)` (`packages/ui/acp/src/index.ts`); the ACP session-id surface beyond the store map — `SessionRecord.sessionId: string`, `bySession = new WeakMap<Agent, string>()`, `loadingIds = new Set<string>()`, `requireSession(sessionId: string)`, and the exported `streamSessionEventUpdate(sessionId: string, …)` (`packages/ui/acp/src/index.ts`); and the persistence coordinator's `Map<string, …>` keyed by session id (`packages/session-persistence/session-persistence/src/coordinator.ts`). A brand that is dropped at the `Map` key buys nothing on lookups — the value of the existing brands is partly unrealized.
+**Gap 2 — brand erosion at the seams of the *already-branded* IDs.** Even `CallId` and `SessionId` decay back to bare `string` at exactly the places confusion is most likely: registry/store key types and public method params. Representative sites include the session store, the agent registry (both keyed by the shared `SessionId`), `ToolPresenter`'s call-id map, ACP's session-id records and loading set, and the persistence coordinator. A brand that is dropped at a collection key buys nothing on lookups — the value of the existing brands is partly unrealized.
 
-## Proposal
+## Decision
 
 A type-only change. Brands are zero-cost casts; nothing about runtime behavior, serialization, comparison, or the wire format changes. The work is in three parts, all honoring the existing "not every string" policy.
 
-- **Brand the bash task id.** Add `BashTaskId = Branded<'BashTaskId'>` plus its same-named factory in `packages/bash/bash/src/types.ts` (the package that *owns* the id), importing `Branded` from `@deepseek-ai/dsh-llm` exactly as `SessionId`/`AgentId` already do. Thread it through `BashTask.id`, the `BashExecutor` seam methods (`get`/`ownerOf`/`readOutput`/`kill`), the generation site in `dsh-bash-local` (brand the counter output once, at creation), and the `dsh-tool-bash` validate/access surface (`validateTaskId` returns a `BashTaskId`; `task_id` is branded at the tool boundary where the model's string arrives).
+- **Brand the bash task id.** Add `BashTaskId = Branded<'BashTaskId'>` plus its same-named factory in `packages/bash/bash/src/types.ts` (the package that *owns* the id), importing `Branded` from `@deepseek-ai/dsh-brand` exactly as `SessionId` does. The brand primitive lives in the dependency-free `dsh-brand` utility package precisely so `dsh-bash` can brand its ids by depending on it alone — it never pulls in `dsh-llm` (or `dsh-session`) just to reach `Branded`. Thread it through `BashTask.id`, the `BashExecutor` seam methods (`get`/`ownerOf`/`readOutput`/`kill`), the generation site in `dsh-bash-local` (brand the counter output once, at creation), and the `dsh-tool-bash` validate/access surface (`validateTaskId` returns a `BashTaskId`; `task_id` is branded at the tool boundary where the model's string arrives).
 
-- **Mint a distinct `OwnerToken` brand.** Add `OwnerToken = Branded<'OwnerToken'>` in `packages/bash/bash/src/types.ts`; type `BashExecRequest.owner` / `BashExecSpec.owner` / `BashExecutor.ownerOf` as `OwnerToken | undefined`. The `dsh-tool-bash` consumer casts the agent's `session.header.id` (a `SessionId`) into an `OwnerToken` at the boundary — the one place the two vocabularies meet. The bash seam never imports `dsh-session`. (Rationale in the next section.)
+- **Mint a distinct `OwnerToken` brand.** Add `OwnerToken = Branded<'OwnerToken'>` in `packages/bash/bash/src/types.ts`; type `BashExecRequest.owner` / `BashExecSpec.owner` / `BashExecutor.ownerOf` as `OwnerToken | undefined`. The `dsh-tool-bash` consumer casts the agent's shared `id` (`SessionId`) into an `OwnerToken` at the boundary — the one place the two vocabularies meet. The bash seam never imports `dsh-session`. (Rationale in the next section.)
 
-- **Stop the brand erosion.** Propagate the existing brands to the `Map` key types and public method params listed under Gap 2 — `Map<SessionId, Session>`, `get(id: SessionId)`, `Map<AgentId, Agent>`, `Map<CallId, …>`, the ACP `SessionRecord.sessionId: SessionId` surface, the coordinator's `Map<SessionId, …>`. This is the larger mechanical share of the diff and the part that makes the *existing* brands actually load-bearing on lookups, not just on the struct fields.
+- **Stop the brand erosion.** Propagate the existing brands to the `Map` key types and public method params listed under Gap 2 — `Map<SessionId, Session>`, `Map<SessionId, Agent>`, `get(id: SessionId)`, `Map<CallId, …>`, ACP's `SessionId` surface, and the coordinator's `Map<SessionId, …>`. This is the larger mechanical share of the diff and the part that makes the *existing* brands actually load-bearing on lookups, not just on struct fields.
 
 Illustrative shape (the factory pattern is identical to the three existing brands):
 
 ```ts ignore-check
-import type { Branded } from '@deepseek-ai/dsh-llm'
+import type { Branded } from '@deepseek-ai/dsh-brand'
 
 /** A background bash task handle (generated `bash-N` by the local executor). */
 export type BashTaskId = Branded<'BashTaskId'>
@@ -40,7 +40,9 @@ export function OwnerToken(id: string): OwnerToken {
 }
 ```
 
-## Why a distinct OwnerToken brand (not SessionId)
+## Alternatives considered
+
+### Why not typing `owner` as `SessionId`?
 
 The obvious shortcut is to type `owner` as `SessionId` directly — it always *is* one. We reject that. The bash executor seam is a capability seam (interface `dsh-bash`, implementation `dsh-bash-local`, consumer `dsh-tool-bash`) and its owner token is *documented as deliberately opaque*: the executor "never interprets it (no access policy lives in the seam — that is the consumer's job)" (`packages/bash/bash/src/types.ts`). Typing the seam's field as `SessionId` would import `dsh-session`'s vocabulary into a package that must not know what an owner token *means* — it would couple a generic execution backend to the session model and contradict the opaque-token design. A sandboxed or remote executor that replaces `dsh-bash-local` should not inherit a session dependency. The distinct `OwnerToken` brand keeps the seam decoupled: `dsh-bash` knows only "an owner is some opaque branded token," and the `dsh-tool-bash` consumer — which already decides the access policy — is the single boundary that casts its `SessionId` into an `OwnerToken`. The brand still delivers the safety win (you cannot pass a `BashTaskId` or a raw string where an owner is expected) without the coupling.
 
@@ -48,21 +50,18 @@ The obvious shortcut is to type `owner` as `SessionId` directly — it always *i
 
 Kept deliberately narrow per the "not every string needs a brand" policy. Each of these is a plausible future brand, deferred with a reason, not a commitment:
 
-- **`ModelId`** (`GenerateOptions.model`, the `LlmService` adapter-registry key) — a real cross-package lookup key (config → agent → llm → adapter); a reasonable next brand, left out only to keep this RFC's blast radius focused.
+- **`ModelId`** (`GenerateOptions.model`, the `LlmService` adapter-registry key) — a real cross-package lookup key (config → agent → llm → adapter); a reasonable next brand, left out only to keep this Agent Note's blast radius focused.
 - **`ToolName`** (the `ToolRegistry` key) — author-defined, human-readable, and rarely confused with another id; the weakest candidate, likely not worth a brand.
 - **`ErrorCode`** (`HarnessError.code`) — a closed vocabulary (`ABORTED`, `NO_ADAPTER`, …), not a per-instance id; better served by a string-literal union than a brand, if anything.
 - **Numeric ordinals** — turn number, step number, and the event `seq` are `number`, not `string`, so `Branded<string>` does not apply; a parallel `number & { readonly [BRAND]: B }` variant could brand them, but they are positional ordinals rarely passed across boundaries, so the payoff is low.
-- **Validated construction** — the brand factories are pure casts with no runtime check, and every boundary (ACP `sessionId`, provider-issued `call.id`, the empty-string fallback in `dsh-llm-deepseek`) trusts the raw string today. A `SessionId.parse()` / `isValid()` companion that throws on malformed input at boundaries is a genuine gap, but it is a *runtime-behavior* change with its own design (what is "malformed"? what do we do on failure?) and belongs in its own RFC, not bundled into this type-only pass.
+- **Validated construction** — the brand factories are pure casts with no runtime check, and every boundary (ACP `sessionId`, provider-issued `call.id`, the empty-string fallback in `dsh-llm-deepseek`) trusts the raw string today. A `SessionId.parse()` / `isValid()` companion that throws on malformed input at boundaries is a genuine gap, but it is a *runtime-behavior* change with its own design (what is "malformed"? what do we do on failure?) and belongs in its own Agent Note, not bundled into this type-only pass.
 
-## Acceptance criteria
+## Verification
 
-- `BashTaskId` and `OwnerToken` are defined in `dsh-bash` and threaded end-to-end: the executor seam, the `dsh-bash-local` generation site, and the `dsh-tool-bash` model-facing surface all speak the brands; `dsh-bash` gains no dependency on `dsh-session`.
-- No collection keyed by an in-scope branded id (`CallId`/`SessionId`/`AgentId`/`BashTaskId`) is keyed by bare `string` — this covers `Map`, `WeakMap` value slots, and `Set` membership (e.g. the ACP `bySession`/`loadingIds`), not just `Map<string, …>`; the corresponding public method params and exported function signatures (e.g. `streamSessionEventUpdate`) take the brand, not `string`.
-- Brands are constructed via the cast factory at each boundary where a raw string enters (provider call id, ACP session id, model-supplied `task_id`); no `as` casts scattered at call sites.
-- `pnpm run typecheck` and `pnpm run doc-sync` are green; the change is observably type-only (no snapshot, no e2e behavioral diff).
+The landed invariants: `BashTaskId` and `OwnerToken` are defined in `dsh-bash` and threaded end-to-end (executor seam, the `dsh-bash-local` generation site, the `dsh-tool-bash` model-facing surface) with no `dsh-bash` dependency on `dsh-session`; no collection keyed by an in-scope branded id (`CallId`/`SessionId`/`BashTaskId`) is keyed by bare `string`; public method params and exported signatures keep the brand; and brands are constructed via the cast factory at each boundary where a raw string enters (provider call id, ACP session id, model-supplied `task_id`), never as scattered `as` casts.
 
-## Risks / what we give up
+## Consequences
 
-- **Mechanical churn across two surfaces.** Propagating brands touches the bash seam (interface + impl + consumer) and the ACP session-id surface plus the persistence coordinator. The risk is broad but low-severity: a missed site is a compile error, not a silent bug. It ships as its own PR, converged with Codex, and stacks naturally near the [unify-the-agent-id-and-the-session-id](../simplification/2026-06-20-unify-agent-and-session-id.md) work (both touch the session-id / owner-token boundary; if that proposal lands first, `OwnerToken` still stays distinct from the unified id for the decoupling reason above).
-- **Brands do not validate.** A brand is a confusability guard, not a correctness proof: a *wrong* session id that is still a well-formed string passes the type checker exactly as before. This RFC does not close that gap (see Out of scope) — it only stops the *category* error of passing the wrong *kind* of id.
-- **The "where to stop" line stays a judgment call.** Branding `BashTaskId` but not `ToolName`, `OwnerToken` but not `ModelId`, is a taste call about which strings "could plausibly be confused." Reasonable reviewers may want more or fewer; the policy in `brand.ts` is the tie-breaker, and this RFC errs toward the ids that are model-facing or used for access control.
+- **Mechanical churn across two surfaces.** Propagating brands touches the bash seam (interface + impl + consumer) and the ACP session-id surface plus the persistence coordinator. The churn is broad but low-severity: a missed site is a compile error, not a silent bug. The change is observably type-only — no snapshot or e2e behavioral diff. It sits next to the [unified agent/session identity decision](../simplification/2026-06-20-unify-agent-and-session-id.md) because both touch the session-id / owner-token boundary; `OwnerToken` stays distinct from the unified id for the decoupling reason above.
+- **Brands do not validate.** A brand is a confusability guard, not a correctness proof: a *wrong* session id that is still a well-formed string passes the type checker exactly as before. This Agent Note does not close that gap (see Out of scope) — it only stops the *category* error of passing the wrong *kind* of id.
+- **The "where to stop" line stays a judgment call.** Branding `BashTaskId` but not `ToolName`, `OwnerToken` but not `ModelId`, is a taste call about which strings "could plausibly be confused." Reasonable reviewers may want more or fewer; the policy in `brand.ts` is the tie-breaker, and this Agent Note errs toward the ids that are model-facing or used for access control.

.agents/notes/implemented/architecture/2026-06-20-extract-example-app-packages.md

@@ -0,0 +1,55 @@
+# Agent Note: Extract example apps into packages
+
+Status: implemented
+
+## Problem
+
+An example folder is supposed to be *thin* — the variable wiring of a demo, not the demo's machinery. Before this change it was thick. Each example carried a hand-rolled `start.ts` boot bootstrap, an infra preamble (`timer`, and — for the stdio demos — `logger` + `hmr`), nested includes of three shared YAML fragments (`base.yml` / `base-core.yml` / `acp-agent/acp-tail.yml`), and per-example `agent-loop`/persistence/system-prompt config. The actual app — the spine of services every agent needs — was spread across the leaf and those includes.
+
+The leaf configs also owned a coupled front door. ACP requires stdout purity and creates agents through `session/new`; stdio requires a console logger and a pre-created `main`. Prose warnings were the only guard against combining these incorrectly, while three `start.ts` files duplicated the Loader bootstrap and lifecycle code.
+
+## Decision
+
+Each example is now **mostly an invocation of an app package**, splitting the wiring along the existing [interface / implementation / consumer seam](2026-06-13-capability-seams.md): the **app package owns the composition**, the leaf `cordis.yml` owns only the **swappable choices** (which LLM adapter, which bash executor, model, prompt, persistence root).
+
+- **`@deepseek-ai/dsh-agent-spine-demo`** ([packages/examples/agent-spine-demo](../../../../packages/examples/agent-spine-demo)) composes the providerless, executor-less, UI-less spine and forwards the loop's agent-list config. Its dependency on the concrete loop is intentional because this package composes the spine rather than extending it; swapping the loop means supplying another bundle.
+- **`@deepseek-ai/dsh-stdio-demo`** ([packages/examples/stdio-demo](../../../../packages/examples/stdio-demo)) and **`@deepseek-ai/dsh-acp-demo`** ([packages/examples/acp-demo](../../../../packages/examples/acp-demo)) bake in their front doors. Stdio includes `ui-stdio`, a console logger, and `main`; ACP includes the bridge and JSONL persistence but no stdout logger or pre-created agent. Leaves may add plugins, but the safe composition is now the default artifact.
+- **`start.ts` is gone.** Each app package exposes a `bin` (`dsh-stdio-demo` / `dsh-acp-demo`); the `demo:*` scripts invoke it (e.g. `dsh-stdio-demo ./cordis.yml`). The Loader-boot tail, `.env` loading, and fail-loud guards live in the shared [`@deepseek-ai/dsh-app-boot`](../../../../packages/ui/app-boot) package (unit-tested under the per-file coverage gate — see [share the app bins' boot glue](../simplification/2026-07-04-share-app-bin-boot-glue.md)); each bin is a thin self-executing composition over those helpers plus its app-specific lifecycle (the ACP bin: snapshot-mode selection and stdin-dispose). The `bin.ts` files themselves stay coverage-excluded (self-executing CLI entries, like the old `start.ts`) and are driven by the keyless Loader-path tests.
+- **Each leaf `cordis.yml` collapses** to backends + config: the LLM adapter (`llm-deepseek` with apiKey/models, or `llm-replay`), the bash executor (`bash-local`), `hmr` for the stdio demos (see the amendment below), and one app entry carrying the app's config (model, system prompt, persistence root — surfaced as the app package's own `Config`, which routes each value to wherever the app wires it: stdio onto its pre-created agent, acp onto the bridge plugin).
+- **echo-agent folds onto `dsh-stdio-demo`**, swapping the LLM backend to the local `mock-llm` and adding the local `echo-tool` (plus `bash-local`, which the spine's `tool-bash` injects) at the leaf — the clean demonstration of "swap the backend, keep the app". `mock-llm.ts` / `echo-tool.ts` stay as example-local teaching plugins.
+- **`base.yml`, `base-core.yml`, and `acp-agent/acp-tail.yml` are retired** — the spine they shared now lives in `dsh-agent-spine-demo`.
+
+`bash-local` and the LLM adapter stay **leaf choices**: the bundle ships `tool-bash` (the consumer schema), the leaf picks the executor implementation, so a sandboxed executor or replay adapter swaps in without touching the app.
+
+### Amendment on implementation: `hmr` stays a leaf entry
+
+The proposal listed `hmr` among the stdio app's baked-in front-door cluster. Validating against the code, baking `hmr` into the `dsh-stdio-demo` package fights cordis in two ways, so it ships as a **leaf `cordis.yml` entry** instead:
+
+1. `@cordisjs/plugin-hmr` is a Loader-only, subprocess-only dev plugin — its constructor throws without `node --expose-internals` + a live `loader` service, so it can only run in the real `demo:*`/bin subprocess, never in the in-process unit/coverage tier.
+2. The in-process test tier (vitest) cannot even *import* the vendored `hmr` module (its class-decorator `@Inject` form fails under Vite's transform), so a package whose `apply` statically imported it could never satisfy the per-file 100% coverage gate on its headline function.
+
+Crucially, `hmr` is **not** a stdout-purity footgun the way the console logger is — a stray `hmr` in the ACP config would not corrupt the JSON-RPC frames — so leaving it at the leaf costs none of the safety the coupling argument is about. The **logger** (the real coupling) stays baked in: the stdio app includes it, the ACP app omits it.
+
+## Alternatives considered
+
+### Why not keep the wiring in shared YAML includes?
+
+The old `base*.yml`/`acp-tail.yml` includes already deduped the *config*, but a YAML include cannot **encapsulate** the front-door coupling — it can only describe it in a comment and trust every leaf to obey. It also cannot own a `bin`, so the boot glue stayed copied across three `start.ts` files. A package turns "the ACP app never logs to stdout" from a prose warning into a property of the artifact: there is no logger entry in the leaf to get wrong.
+
+## Verification
+
+- Example directories contain only their config, README, and tests: `start.ts`, the infrastructure preamble, and the shared YAML includes are gone.
+- `demo:echo`, `demo:repl`, and `demo:acp` invoke the app-package bins.
+- Each new package has a README and per-file 100% coverage; each app package also has a keyless real-Loader-path bin smoke that catches export-shape failures described in [postmortem 0001](../../../../docs/postmortem/0001-acp-default-export-drops-inject.md).
+- The ACP replay transcript remains unchanged because the plugin set and load order did not change.
+
+## Consequences
+
+- **The bare-plugin-tree pedagogy.** echo-agent's inlined `cordis.yml` showed every plugin at once; the spine now lives behind a bundle, so seeing the whole tree means opening `dsh-agent-spine-demo`. The app package's README carries that teaching weight.
+- **A layer of indirection.** "What does this demo load?" becomes a package read, not a single YAML scan.
+
+## Related
+
+- Supersedes [Make the shared example base providerless](../../rejected/architecture/2026-06-20-providerless-example-base.md): renaming `base.yml` to the providerless core is moot once the spine moves into `dsh-agent-spine-demo` and the `base*.yml` files are deleted.
+- Builds on the [capability-seams](2026-06-13-capability-seams.md) interface/implementation/consumer split — backends and presentation stay leaf choices; the spine is the shared bundle.
+- Complements [Reorganize packages into a modular hierarchy](2026-06-20-package-hierarchy.md): the new app/core packages slot into existing groups under that hierarchy (`core` for the reusable spine bundle, `ui` for the app-specific front doors).

.agents/notes/implemented/architecture/2026-06-20-generic-long-running-tool-runtime.md

@@ -0,0 +1,128 @@
+# Agent Note: The background task runtime (`ctx.tasks`) and generic task control tools
+
+Status: implemented
+
+## Problem
+
+Background bash originally combined two responsibilities: the bash executor ran processes and also managed task ids, ownership, incremental reads, cancellation, completion listeners, and model-facing control tools. Adding background subagents required the same lifecycle and interaction contract. Implementing that contract independently for every long-running capability would duplicate isolation, cleanup, notification, and prompt behavior while teaching the model a different collect-and-stop protocol for each producer.
+
+The task registry, control tools, and completion notices form one harness capability. Bash and subagents should supply execution-specific hooks without owning generic task behavior.
+
+## Decision
+
+The `tasks/` package group owns background-task semantics:
+
+- `@deepseek-ai/dsh-tasks` registers running work as `ctx.tasks` and owns task ids, authorization, snapshots, reads, cancellation, waiting, completion listeners, and cleanup.
+- `@deepseek-ai/dsh-tool-tasks` exposes `task_output`, `task_list`, and `task_kill`, injects completion notices, and supplies the background-task system-prompt guidance.
+
+Long-running tools are producers. `dsh-tool-bash` adapts a `BashProcess` into incremental output and process cancellation; `dsh-tool-subagent` adapts a child run into final output and child disposal. The execution seams remain independent of sessions and the task registry.
+
+`TaskService` is a concrete, process-local service. TODO(task-service-backend): separate its public contract from the implementation when a second backend defines the required lifecycle; a systemd-backed runtime is one plausible driver, but this PR does not speculate about its durability, reconnect, ownership, or observation semantics.
+
+## Runtime contract
+
+The literal types live in the [task data-structure catalog](../../../../docs/core-data-structures/tasks.md). A producer calls `ctx.tasks.start()` with a kind, label, optional owning `Agent`, and a `run()` function. The runtime completes all failable preflight work before calling `run()` and invokes it once. After `run()` returns hooks, registration commits without another failable step; a producer cannot start work that lacks a collectable task id.
+
+The producer hooks define three responsibilities:
+
+- `cancel(reason?)` synchronously requests termination, is idempotent, and must cause `done` to settle.
+- `done` never rejects and settles only after the producer has released the task's resources.
+- Optional `readOutput()` returns the next consuming output delta. Omitting it declares a final-output task whose terminal result comes from `TaskOutcome.output`.
+
+Statuses are `running`, `stopping`, `completed`, `killed`, and `failed`. Producer-specific information such as an exit code or stop reason belongs in `detail`; the registry does not interpret it. Task kinds form a merge-extensible string union, and task ids are branded and generated as `<kind>-N`, with a counter per kind.
+
+The runtime attaches one continuation to `done`, records the first terminal outcome, resolves waiters, and invokes completion listeners with per-listener error containment. First-wins settlement matters during teardown: if `cancel` throws, the runtime force-fails the record and warns that work may be orphaned rather than waiting forever for a promise that may never settle. A later producer outcome cannot overwrite that diagnosis or notify twice. A `cancel` that returns without eventually settling `done` still blocks teardown because the runtime cannot distinguish it from a slow, valid stop.
+
+Task registrations are not effects of the producer tool fiber. Reloading a tool or control-surface plugin therefore does not kill work owned by an agent and backend. The task service's own disposal cancels all live tasks and awaits contract-compliant producers.
+
+## Authorization and owner lifecycle
+
+Task ids are runtime-global and predictable, so every access is authorized by the registry. `get`, `read`, `wait`, and `kill` accept the calling `Agent`; `list` returns only tasks visible to that caller. An owned task is accessible only to the exact owning session. Unowned tasks are open to non-agent callers and die with the task service.
+
+The snapshot stores the owner's branded `SessionId` for authorization, while lifecycle operations retain the exact live `Agent` instance. These identities serve different purposes: session equality grants access, but exact object identity selects cleanup and completion delivery. Reusing an agent or session id cannot redirect an old scope's cleanup or notices to a replacement.
+
+The first task for an owner attaches one asynchronous effect to `owner.ctx`. Agent-scope disposal cancels that owner's live tasks, awaits their terminal records, and removes their snapshots. This effect survives producer reloads and joins the agent's existing quiescence boundary. The task service retains the effect disposer so service reload can detach callbacks from still-live agent scopes after global teardown.
+
+For contract-compliant producers, `AgentHandle.dispose()` resolves only after owned background work has stopped. Work intended to outlive an agent must be started unowned; survival across runtime restarts requires a separate durable-job design.
+
+## Service surface
+
+`TaskService` provides:
+
+- `start(spec)` for preflighted, atomic registration.
+- `get(id, caller?)` and `list(caller?)` for non-consuming snapshots.
+- `read(id, caller?)` for a consuming stream delta or an idempotent final result.
+- `kill(id, caller?, reason?)` for cancellation.
+- `wait(id, timeoutMs, caller?, signal?)` for bounded terminal waiting.
+- `onTaskDone(listener)` for effect-scoped observation with exact-owner delivery and listener containment.
+- `attachSurface(name)` for the control-surface availability fence.
+
+`wait` returns the terminal snapshot when the task settles or the live snapshot when its timeout expires. Aborting a wait cancels only that wait. If settlement has already assigned terminal delivery to the waiter, the terminal snapshot still wins. Waiters unregister synchronously on abort so a same-tick settlement cannot suppress a completion notice on behalf of a reader that receives nothing.
+
+A producer loaded without any control surface would let callers start work they cannot collect or stop. `dsh-tool-tasks` therefore calls `attachSurface()` for its lifetime, and `start()` fails before producer execution when no surface is attached. This check occurs at start rather than plugin load because sibling plugins may activate concurrently. Custom non-model surfaces can attach themselves without teaching the registry tool names.
+
+## Model-facing control surface
+
+`dsh-tool-tasks` registers three kind-independent tools with generic ACP cards:
+
+- `task_output(task_id, wait?, timeout_ms?)` reads output and always appends `[status: ...]`. Stream tasks return only output since the previous read; final-output tasks return their result after settlement. Reads are non-blocking unless `wait: true`, whose timeout is defaulted and capped by plugin config. A wait timeout reports the still-running status and does not stop the task.
+- `task_list()` returns caller-visible tasks as `<id> [<kind>] <status> — <label>`, or `(no background tasks)`.
+- `task_kill(task_id, reason?)` requests cancellation immediately. The optional logged reason is forwarded to the producer. Terminal tasks report their existing status; a throwing producer cancel fails the call and leaves the task running.
+
+Stream reads share one task-scoped consuming cursor because the owning model is the intended reader. A UI or multiple independent readers need a separate non-consuming observation API; sharing this cursor would let readers consume one another's output.
+
+The system prompt tells the model to retain task ids, continue independent work instead of busy-polling or duplicating a running task, collect relevant tasks before its final answer, and kill work that no longer matters. Completion injects a logged `context/message` into the exact owner's session; it becomes durable context for the next request but does not wake an idle agent.
+
+The runtime marks a terminal task `reported` when a read or wait delivers it, when a live waiter has claimed delivery at settlement, or when the model explicitly kills it. Reported tasks do not inject redundant completion notices. Listener failures are logged independently, do not stop later listeners, and are not awaited by waiters or teardown.
+
+## Producer opt-in
+
+Each producer owns whether its schema exposes `run_in_background` through defaulted config. `dsh-tool-bash` and each `dsh-tool-subagent` instance use `enableRunInBackground`, defaulting to true. A disabled instance omits the parameter and also rejects a forced background argument at execution because the generic argument validator permits undeclared keys. Schema omission advertises the capability; the execution check enforces it.
+
+`ctx.tasks` does not rewrite producer schemas. A bundle forwards configuration only for producers it owns. If a background call reaches `start()` without an attached surface, the runtime fence fails before execution.
+
+## Producer integrations
+
+The bash seam exposes `resolve`, `run`, and `start`. `start(spec)` returns a `BashProcess` with incremental reads, cancellation, exit facts, and a non-rejecting quiescence promise. The local executor retains live handles only so its own disposal can kill and join processes. Foreground callers continue to use `resolve` and `run` directly.
+
+For background bash, `dsh-tool-bash` registers the calling agent as owner. Its hooks map `kill()` to cancellation, `done` to a completed or killed `TaskOutcome`, and `readOutput()` to the process's bounded incremental output plus spill and sandbox notices. Generic task tools own ids, status lines, listing, waiting, and completion notices.
+
+For background subagents, `dsh-tool-subagent` creates a task-owned `AbortController` and begins provider startup inside the task starter. Cancellation aborts the same signal before or after provider readiness. `done` awaits both the child result and child disposal, maps completed output to a final result, maps abort to `killed`, and maps other stop reasons or infrastructure failures to `failed`. Intermediate child history remains in the child session and is not exposed through `readOutput()`.
+
+## Alternatives considered
+
+### Per-capability control tools
+
+Separate bash and subagent output/stop tools duplicate ids, isolation, cleanup, notification, and guidance while increasing the model's schema and protocol burden. One runtime keeps execution-specific behavior in producers without cloning the task lifecycle.
+
+### An immediate abstract task-runtime backend
+
+The current `TaskStart.run()` contract passes in-process callbacks and exact `Agent` objects. A durable backend changes identity, restart, ownership, and observation semantics, so extracting an interface before a second implementation exists would freeze the wrong boundary.
+
+### Consumer-owned authorization or cleanup events
+
+Consumer-owned checks invite inconsistent or missing isolation on each new surface. A broadcast cleanup event makes every listener filter every agent and provides no registration disposer. Central authorization plus one owner-scoped effect gives every consumer the same fence and an awaited, removable lifecycle hook.
+
+### Blocking output or a separate wait tool
+
+Blocking by default would serialize the parent while background work runs. Waiting without reading would add another model call and schema without returning useful information. `task_output(wait: true)` makes blocking explicit and combines it with result delivery.
+
+The wait uses the shared deadline primitives but not the generic tool-timeout policy. A wait timeout is a successful observation that returns `[status: running]`; the generic policy would replace it with a timeout error. No tool-call timeout controls task lifetime after a task id has been returned.
+
+### Runtime-owned output sinks
+
+A push sink would centralize buffering, but bash already owns bounded buffers, truncation, and spill files behind its executor seam. Pulling formatted deltas preserves that ownership. A durable backend that owns storage may justify revisiting the producer interface.
+
+### Random ids, promotion, or lifecycle session events
+
+Authorization, not unguessability, is the access boundary, and ids do not derive filesystem paths; sequential branded ids keep transcripts readable. Foreground-to-background promotion requires a user interaction contract the SDK does not prescribe. Starts, reads, and notices are already logged as tool and context events, so dedicated task session events would duplicate model-visible facts.
+
+## Testing
+
+Unit coverage pins preflight atomicity, per-kind ids, stream and final reads, wait timeout and abort races, cancellation, first-wins settlement, listener containment, notice suppression, owner isolation, stale owner instances, owner cleanup, service teardown, and the no-surface fence. Producer tests cover bash process mapping, subagent startup cancellation, terminal mapping, and disposal. Snapshot coverage pins the control-tool schemas and prompt guidance.
+
+## Consequences
+
+Bash commands and subagents share one id vocabulary, listing, notice format, prompt habit, and set of control tools. New long-running producers implement execution hooks instead of another registry and tool family. The [tool cookbook](../../../../docs/cookbook/adding-a-tool.md) points producers to this contract.
+
+Owned background bash now stops with its agent instead of surviving it. Background processes have no executor timeout; callers must kill irrelevant work or rely on owner/service disposal. Stream reads support one consuming reader, completion notices do not wake idle agents, and a producer that returns from `cancel` without settling `done` can still stall teardown. Durable jobs, independent observation cursors, and foreground promotion remain separate designs.

.agents/notes/implemented/architecture/2026-06-20-package-hierarchy.md

@@ -1,4 +1,4 @@
-# RFC: Reorganize packages into a modular hierarchy
+# Agent Note: Reorganize packages into a modular hierarchy
 
 Status: implemented
 
@@ -8,7 +8,7 @@ Status: implemented
 
 This was not just cosmetic. Because every top-level package looked like part of the same public surface, future removal was harder, and publish/lint/doc scripts had to encode intent through comments or hand-maintained static lists rather than reading it off the layout.
 
-## What landed
+## Decision
 
 Packages are grouped by modular role at a uniform `packages/<group>/<pkg>/` depth. Group directories are pure containers (no `package.json`); every package keeps its `@deepseek-ai/dsh-<pkg>` name — this is repo structure and maintenance policy, not package renaming.
 
@@ -51,7 +51,7 @@ packages/
 
 The package list had been enumerated in five places. The uniform depth-2 layout lets most of them be derived instead:
 
-- `tsconfig.base.json` and `tsconfig.typecheck.json` each map every package through a single `@deepseek-ai/dsh-*` `paths` wildcard listing one candidate per group, in place of 18 per-package entries. (One subtlety this introduced: a path candidate contains `/*/`, which a naive regex comment-stripper mistakes for a block comment — `scripts/doc-typecheck.ts` reads the `paths` map via the TypeScript JSONC API rather than stripping comments by hand for exactly this reason.)
+- `tsconfig.base.json` maps every package through a single `@deepseek-ai/dsh-*` `paths` wildcard listing one candidate per group, in place of per-package entries. Root `tsconfig.json` reuses that source map and carries the explicit project references that keep package/vendor typecheck boundaries intact. (One subtlety this introduced: a path candidate contains `/*/`, which a naive regex comment-stripper mistakes for a block comment — `scripts/doc-typecheck.ts` reads the JSONC config through TypeScript's parser rather than stripping comments by hand for exactly this reason.)
 - `scripts/publint-all.ts` derives its list by reading the hierarchy (`packages/<group>/<pkg>`), resolving the `TODO(package-inventory)`.
 - `tsconfig.build.json`'s project `references` stay an explicit list — TypeScript project references have no wildcard form. Generating these from a manifest is left to a follow-up (see [discover package inventories](../../proposed/process/2026-06-20-discover-package-inventory.md)).
 
@@ -62,6 +62,12 @@ Two doc-sync/hygiene gates keep the structure and its references honest, so the
 - `scripts/verify-package-paths.ts` flags a `packages/<path>` reference (in Markdown or a `.ts` comment/string) that does not resolve **and** names a real package in a segment — i.e. a stale path to a moved package. A path naming a package that exists nowhere (a forward-looking proposal) is left alone, so the gate applies uniformly across proposed/implemented/rejected.
 - `scripts/check-workspace-constraints.ts` asserts the `packages/<group>/<pkg>` shape: group dirs carry no `package.json`, and no package sits flat at the root or nests deeper. Group names stay open — a new group may be added without editing the gate; only the depth-2 shape is fixed.
 
-## What we gave up
+## Alternatives considered
+
+- **A third tier (`adapters/` / `impls/` under each family)** — rejected: uniform depth 2 keeps the workspace glob a clean `packages/*/*` and lets one `@deepseek-ai/dsh-*` tsconfig wildcard resolve every package.
+- **Nesting persistence under `core/session/`** — rejected: the storage backends form a parallel capability family mirroring `llm/` and `bash/`, while the session log itself stays core product API.
+- **`ui-stdio` under `ui/`** — rejected: it is example-coupled dev support, not a product surface; `acp` is the only `ui/` member because an editor actually drives it.
+
+## Consequences
 
 The restructure churned imports, workspace globs, doc links, build references, and package paths in one coordinated move. That churn is acceptable pre-release (per the AGENTS.md foundation-over-blast-radius stance) because it stops the flat layout from fossilizing support packages as product contracts, and it is a one-time cost: the wildcard `paths`, the glob-derived publint list, and the shape gate mean a new package needs no further structural edits.

.agents/notes/implemented/architecture/2026-06-21-mandatory-app-attribution-headers.md

@@ -0,0 +1,82 @@
+# Agent Note: Mandatory `User-Agent` attribution for provider requests
+
+Status: implemented
+
+## Problem
+
+LLM provider requests should identify the product making them. That is useful for provider-side support, abuse investigation, compatibility debugging, and traffic analytics. Before this Agent Note the harness only partially did this: the hand-rolled DeepSeek adapter sent a hand-copied `User-Agent` constant (`packages/llm/llm-deepseek/src/adapter.ts`), while the pi-ai-backed twin sent no harness-owned headers at all (`packages/llm/llm-pi-ai/src/adapter.ts`). New adapters could therefore omit attribution silently, and a library-backed adapter could drift from the hand-rolled adapter even though [the twin-adapter Agent Note](2026-06-13-twin-llm-adapters.md) exists to keep the provider seam honest across both implementations.
+
+The immediate prompt came from OpenRouter's [App Attribution](https://openrouter.ai/docs/app-attribution) docs. OpenRouter creates app pages and rankings from `HTTP-Referer` plus display/category headers. That is valuable, but it is not the HTTP standard for application identity. The risk is adopting OpenRouter's exact header set as if it were universal, then leaking provider-specific headers to direct DeepSeek requests, future OpenAI/Anthropic/Vertex adapters, test servers, or proxies that log unknown fields indefinitely.
+
+## Investigation
+
+- **OpenRouter's mechanism is provider-specific.** Their current docs say app attribution is tracked through `HTTP-Referer` (required), `X-OpenRouter-Title`, and `X-OpenRouter-Categories`; `X-Title` is only accepted for backward compatibility. Their API reference calls the headers optional and says they make the app discoverable on OpenRouter. This is a concrete OpenRouter contract, not an IETF or OpenAI-compatible API standard.
+- **In agent tooling, `HTTP-Referer` is an OpenRouter-aware convention, not a general agent convention.** It is common enough that OpenRouter SDKs and OpenRouter examples expose it directly, and frameworks that target OpenRouter usually need a way to pass it through. But agent protocols such as ACP negotiate names, versions, and capabilities in their own initialize messages, while model-provider requests still need HTTP-level identity. "Accepted in the agent world" therefore means "recognized by OpenRouter integrations," not "portable across agent runtimes or providers."
+- **Coding agents identify the product and version in `User-Agent`.** Public implementations vary in environment detail and provider-specific side headers, but product identity is the common contract; there is no universal exact format.
+- **The standards-track general client identity header is `User-Agent`.** RFC 9110 section 10.1.5 defines `User-Agent` as the user-agent software identity, says it is used for interoperability reports and analytics, and says a user agent SHOULD send it on each request unless configured not to. This is the only standard header that directly matches "what product is making this HTTP request."
+- **`Referer` is standard, but OpenRouter's `HTTP-Referer` is not the standard field.** RFC 9110 section 10.1.3 defines `Referer` as the URI from which the target URI was obtained and spends significant text on privacy restrictions. OpenRouter instead asks for `HTTP-Referer`, using it as an app URL identifier. That name and meaning are OpenRouter-specific even though it resembles the CGI environment variable form of the standard `Referer` header.
+- **`From` is standard but not suitable as a mandatory default.** RFC 9110 section 10.1.2 defines `From` as an email address for the human responsible for a user agent. Robotic agents SHOULD send it so servers can contact an operator, but non-robotic agents should not send it without explicit user configuration because of privacy and security policy concerns. The harness can support an operator contact later, but must not invent one or require it globally.
+- **Request-body `user` or `metadata` fields are not app attribution.** Some model APIs expose a stable end-user identifier, request metadata, labels, or project/account headers. Those are useful for abuse monitoring, internal billing, dashboards, or trace correlation, but they either identify the end user rather than the product, are provider-specific body schema, or are not guaranteed to be forwarded through OpenAI-compatible gateways. They are not a substitute for a static application identity header.
+- **SDK telemetry headers identify the SDK, not the app.** Official and third-party SDKs often send library/version headers. Those help the SDK maintainer debug their client, but they do not identify the harness as the application unless the application explicitly supplies a product attribution layer.
+- **pi-ai has a first-class header hook.** `@earendil-works/pi-ai`'s `StreamOptions.headers` merges caller headers last over provider defaults, so a library-backed adapter can satisfy the same wire contract as the hand-rolled one without wrapping or upstream work. The mock-server suites assert arrival on the wire for both adapters.
+
+## Decision
+
+Provider request attribution is mandatory at the LLM adapter boundary, using the standard `User-Agent` header only. The rule: every product LLM adapter sends a static, non-secret application identity on every provider HTTP request, and every adapter has tests proving that `User-Agent` reaches the wire (a mock server asserting received headers; for a library-backed adapter, the library's header hook feeding the same mock-server assertion).
+
+Do **not** implement OpenRouter app attribution in this Agent Note. `HTTP-Referer`, `X-OpenRouter-Title`, `X-Title`, and `X-OpenRouter-Categories` are OpenRouter-specific product-surface headers, not provider-neutral model-request attribution. They can be proposed later by an OpenRouter adapter or explicit OpenRouter mode, with its own privacy/product decision, tests, and docs. Until then, even requests pointed at OpenRouter send only the shared `User-Agent` attribution from this Agent Note.
+
+The provider-neutral identity is owned by `dsh-llm` (`packages/llm/llm/src/attribution.ts`), not by individual adapters. `AppIdentity` contains only public product facts needed to build `User-Agent`, and the default `APP_IDENTITY` settles the values the proposal left open:
+
+- product token for `User-Agent`: `deepseek-harness` (continuity with the pre-Agent Note wire value and the repo/org identity)
+- version: read from the owning package's manifest via `createRequire`, never a hand-copied constant
+- app URL: `https://github.com/deepseek-ai/deepseek-harness-sdk` - the planned public home; a `FIXME` in `attribution.ts` blocks release until that repository actually exists
+
+The default is mandatory and non-empty. White-label deployments pass their own `AppIdentity` to `attributionHeaders(identity)` - the override seam is the function parameter, with no deployment config plumbing until a consumer needs it - and omission falls back to the harness default rather than suppressing attribution. There is no per-request API for the model, user prompt, session id, cwd, user email, API key owner, or local machine identity to influence these fields.
+
+Wire mapping (`attributionHeaders`; header names lowercase in code - HTTP field names are case-insensitive on the wire):
+
+| Target | Mapping |
+|---|---|
+| All HTTP-based adapters | `User-Agent: {product}/{version} (+{url})` - the parenthesized `+url` comment stays within RFC 9110's conservative product/comment syntax. |
+| Direct DeepSeek endpoint | `User-Agent`; do not send OpenRouter-only headers unless DeepSeek documents an equivalent contract. |
+| OpenRouter endpoints | `User-Agent` only for now. Do not send `HTTP-Referer`, `X-OpenRouter-Title`, `X-Title`, or `X-OpenRouter-Categories` under this Agent Note. |
+| Future providers | `User-Agent` only unless a later provider-specific Agent Note accepts additional headers. Do not reuse `HTTP-Referer` by analogy. |
+
+Endpoint detection is not part of this Agent Note because no endpoint-specific mapping is accepted here. If OpenRouter support lands later, detection must be explicit: either a dedicated OpenRouter provider package or an explicit `provider: 'openrouter'` / `attributionTarget: 'openrouter'` config, not arbitrary path fragments or model names.
+
+## Verification
+
+The landed contract:
+
+- `dsh-llm` documents the mandatory `User-Agent` attribution contract for `LlmAdapter` authors (`LlmAdapter` JSDoc, package README, and the adapter-contract section of `docs/core-data-structures/llm-streaming.md`).
+- A shared helper (`attributionHeaders` / `userAgent`) constructs the app identity and the standard `User-Agent` value from package metadata, so adapters do not hand-copy version constants.
+- `dsh-llm-deepseek` sends the shared `User-Agent` on every request and its mock-server suite asserts the exact value.
+- `dsh-llm-pi-ai` sends the same `User-Agent` through pi-ai's `StreamOptions.headers` hook and its mock-server suite asserts the exact value.
+- No adapter sends OpenRouter-specific attribution headers (`HTTP-Referer`, `X-OpenRouter-Title`, `X-Title`, `X-OpenRouter-Categories`) as part of this Agent Note.
+- No app-attribution field carries secrets, local paths, session ids, prompt text, model output, user email, or per-user stable identifiers.
+- The adapter READMEs state the `User-Agent` attribution policy and explicitly avoid documenting OpenRouter app attribution as implemented behavior.
+
+## Alternatives considered
+
+**OpenRouter app attribution now.** Rejected for this Agent Note. Sending `HTTP-Referer` plus `X-OpenRouter-Title` would satisfy OpenRouter rankings, but those headers are a provider-specific product feature, not the provider-neutral model-request attribution this Agent Note is trying to standardize. Supporting them should be an explicit OpenRouter adapter/mode decision later, not hidden inside the first shared attribution helper.
+
+**OpenRouter headers everywhere.** Rejected. It would treat a custom OpenRouter contract as a universal standard and send fields with misleading semantics to providers that did not ask for them. It also risks using `HTTP-Referer` as a generic app URL field even though standard HTTP already has `User-Agent` for product identity and `Referer` for a different browsing-context concept.
+
+**Only provider account/project identity.** Rejected. Organization/project headers, API keys, cloud accounts, and billing projects identify who pays or owns the request, not which application is sending traffic. They also expose no public app title/category and do not help gateways like OpenRouter build app rankings.
+
+**End-user `user`/`metadata` fields.** Rejected for this Agent Note. Those are valuable for abuse monitoring and customer support but describe the human or tenant behind a request. App attribution must be static product identity and safe to send on every request.
+
+**Config-only opt-in attribution.** Rejected. A default-off setting is exactly how adapters keep drifting. The policy is mandatory default attribution with overrideable public values, not optional attribution.
+
+**Product-named token (`deepseek-harness-sdk`).** Considered for the `User-Agent` token, since the product name is DeepSeek Harness SDK. `deepseek-harness` won on continuity: it is the identity providers already see from this codebase, it matches the org/repo identity and package scope, and it keeps wire attribution stable while display copy carries the product name.
+
+## Consequences
+
+**Providers see that traffic comes from the harness.** That is the point, but it means deployments that previously blended into generic SDK traffic become identifiable. Mitigation: send only static public product data and let forks/white-label deployments pass their own `AppIdentity`.
+
+**The app URL points at a repository that does not exist yet.** `deepseek-ai/deepseek-harness-sdk` is the planned public home; until it is created the URL is a dangling promise. The `FIXME` marker on the constant blocks a release from shipping with it unresolved (see `docs/development.md` marker semantics).
+
+**Header support differs by client library.** The hand-rolled adapter sets headers directly; the pi-ai-backed adapter depends on pi-ai continuing to honor `StreamOptions.headers` (merged last over provider defaults). The wire-level mock-server tests are the guard: if a pi-ai upgrade stops delivering the header, the suite goes red. This is useful pressure on the abstraction: a provider adapter that cannot set mandatory headers cannot fully implement the harness LLM contract.
+
+**OpenRouter rankings do not benefit yet.** `User-Agent` is the correct baseline for provider-neutral HTTP identity, but it will not create OpenRouter app pages or rankings because OpenRouter requires `HTTP-Referer` for that product feature. That is deliberate: public app marketplace participation is a separate product decision, not a prerequisite for mandatory request attribution.

.agents/notes/implemented/architecture/2026-06-24-web-capability-seam.md

@@ -0,0 +1,332 @@
+# Agent Note: Web capability seam - stable tools over multiple providers
+
+Status: implemented
+
+## Problem
+
+The harness needs model-facing web tools without binding the model contract to one vendor's API shape. Search is the immediate pressure point: supporting both Exa search and Perplexity search from the start — two deliberately different provider shapes (Exa returns a flat `results[]` of `{title, url, highlights, publishedDate}`; Perplexity returns a generated answer plus citations) — is what proves the normalized seam does not just mirror one vendor. Fetch is a separate capability: an anonymous public HTTP(S) fetch backend has transport, security, redirect, decoding, and size-limit concerns that are not the same as provider-backed search.
+
+The model-facing surface must stay stable while backends change. A search provider swap should not change how the model asks for a query, and a fetch implementation swap should not change how the model asks for a URL. Conversely, a provider package should not expose its own model-facing tool schema just because it has extra provider-specific knobs.
+
+Putting search and fetch directly in `dsh-tool-web` would make the model-facing tool own provider selection, backend request mapping, transport policy, result normalization, prompt guidance, presentation, and schema registration at once. Letting each provider register its own tool has the opposite problem: tool availability, names, descriptions, and parameters would depend on whichever provider packages happen to load, and provider-specific fields would leak into the model contract.
+
+There is also a provider-selection question. Existing `tool-bash` and `tool-fs` can rely on Cordis `inject` because there is one backend service key. Web has two independent capabilities (`search` and `fetch`) and potentially multiple providers per capability. `inject: ['web']` proves the seam exists; it does not prove a usable search or fetch provider exists, and it does not define which provider should win when several are registered.
+
+## Decision
+
+Web access is a first-class capability seam following [the capability-seam Agent Note](2026-06-13-capability-seams.md):
+
+1. `@deepseek-ai/dsh-web` (`packages/web/web`) owns `ctx.web`, provider registration, provider selection, shared request/result vocabulary, and web-specific errors.
+2. Provider packages implement concrete backends and register capabilities with `ctx.web`, for example `@deepseek-ai/dsh-web-search-exa`, `@deepseek-ai/dsh-web-search-perplexity`, `@deepseek-ai/dsh-web-search-deepseek`, and `@deepseek-ai/dsh-web-fetch-local`.
+3. `@deepseek-ai/dsh-tool-web` (`packages/web/tool-web`) owns the model-facing `web_search` and `web_fetch` tool schemas, prompt sections, argument validation, result formatting, and tool-owned presentation over `ctx.web`.
+
+Providers do not register tools. Providers register capabilities. `dsh-tool-web` is the only owner of model-facing names, descriptions, prompt guidance, JSON schemas, and presentation.
+
+Search and fetch are separate tools but one web-access seam. `ctx.web` owns provider selection, abort/error vocabulary, and deployment configuration for both parallel registries. Their request schemas and provider logic remain separate; the shared service is the product boundary for reaching the web.
+
+`dsh-tool-web` registers model-facing web tools when the product has enabled those tools and the `ctx.web` seam is present. Backend availability is an execution-time concern, not a schema-registration concern:
+
+- `web_search` is registered when web search is enabled for the product/app, `web_fetch` when web fetch is.
+- A tool is never unregistered merely because its selected provider is missing, misconfigured, missing credentials, ambiguous, or temporarily unavailable.
+- The provider is resolved at execution time, and a structured `WebError` is returned when the selected capability cannot run.
+
+This keeps the model schema stable without making plugin load order, credential state, or HMR timing part of the model-facing contract. If web search is enabled but no usable search provider exists, `web_search` remains visible and execution fails with a structured `WebError` such as `WEB_PROVIDER_UNAVAILABLE` or `WEB_PROVIDER_CONFIGURED_UNAVAILABLE`. If a provider appears after `dsh-tool-web`, the next execution can use it without changing the schema. If a provider disappears mid-call, execution fails with a structured `WebError` instead of silently choosing another provider or falling through to `UNKNOWN_TOOL`.
+
+The seam deliberately exposes no observation surface — no registry-change event and no aggregated capability-status query. Unavailability is a fact a caller observes by executing: `search()`/`fetch()` resolve the provider at call time and throw the structured `WebError` that names what failed. [The observation-surface Agent Note](../simplification/2026-07-04-drop-unconsumed-web-observation-surface.md) records that judgment: derived-on-call selection and enablement-based registration leave no consumer that needs a change signal or an availability probe distinct from executing and routing the error, and a future provider-status panel reintroduces the smallest signal or query it actually consumes.
+
+## Package topology
+
+The three-package interface/implementation/consumer split follows bash and filesystem, but the *interface* package is closer to the LLM seam. `LlmService` (`packages/llm/llm/src/index.ts`) is a name-keyed provider registry: `registerAdapter(models, adapter)` stores adapters in a `Map`, returns a disposer, throws `DUPLICATE_ADAPTER` on duplicate keys, and throws `NO_ADAPTER` at resolution time. `ctx.web` follows that registry shape, but has two capability kinds and a richer selection policy (a configured provider id, or auto-select when exactly one usable provider is registered), so the `WebError` an execution throws can explain why a search or fetch capability cannot run.
+
+The dependency direction mirrors bash and filesystem:
+
+```text
+@deepseek-ai/dsh-tool-web  --depends on-->  @deepseek-ai/dsh-web  <--depends on--  @deepseek-ai/dsh-web-search-exa
+        consumer                                 interface                       implementation
+                                                                 <--depends on--  @deepseek-ai/dsh-web-search-perplexity
+                                                                                  implementation
+                                                                 <--depends on--  @deepseek-ai/dsh-web-search-deepseek
+                                                                                  implementation
+                                                                 <--depends on--  @deepseek-ai/dsh-web-fetch-local
+                                                                                  implementation
+```
+
+At runtime, provider packages register capabilities with `ctx.web`; `tool-web` registers stable tools with `ctx.tools` and executes through the seam:
+
+```mermaid
+flowchart LR
+  exa["@deepseek-ai/dsh-web-search-exa"] -->|registerSearchProvider| web["@deepseek-ai/dsh-web / ctx.web"]
+  perplexity["@deepseek-ai/dsh-web-search-perplexity"] -->|registerSearchProvider| web
+  deepseek["@deepseek-ai/dsh-web-search-deepseek"] -->|registerSearchProvider| web
+  fetchLocal["@deepseek-ai/dsh-web-fetch-local"] -->|registerFetchProvider| web
+  toolWeb["@deepseek-ai/dsh-tool-web"] -->|search/fetch| web
+  toolWeb -->|ctx.tools.register| webSearch["tool: web_search"]
+  toolWeb -->|ctx.tools.register| webFetch["tool: web_fetch"]
+```
+
+`@deepseek-ai/dsh-web` depends only on Cordis and low-level harness support. It declares `ctx.web`, provider interfaces, request/result types, the provider availability contract, and error codes. It does not import tool, agent, session, LLM, or provider packages.
+
+Provider packages depend only on `dsh-web` and Cordis. They own credentials, endpoints, wire mapping, parsing, and `WebError` translation, using platform `fetch`. Each provider injects the shared service and registers a backend; only `dsh-web` owns the `ctx.web` key. Provider-private protocol shapes do not create dependencies on `ctx.llm` or a Cordis HTTP service.
+
+`@deepseek-ai/dsh-tool-web` depends on `@deepseek-ai/dsh-web`, `@deepseek-ai/dsh-tools`, `@deepseek-ai/dsh-system-prompt`, and Cordis. It never imports concrete provider packages.
+
+## `ctx.web` contract
+
+`ctx.web` is a provider registry plus a provider-selecting execution surface. The registry half stays close to `LlmService`: a `Map<id, provider>` per capability kind, `registerSearchProvider` / `registerFetchProvider` methods that return disposers, duplicate ids that throw `WebError`, and execution-time resolution that throws when the selected provider is absent or unusable. The authoritative signatures live in `packages/web/web/src/types.ts`; the seam's shape:
+
+```ts
+interface WebSearchProvider {
+  readonly id: string
+  available(): boolean
+  search(request: WebSearchRequest, signal?: AbortSignal): Promise<WebSearchResult>
+}
+
+interface WebFetchProvider {
+  readonly id: string
+  available(): boolean
+  fetch(request: WebFetchRequest, signal?: AbortSignal): Promise<WebFetchResult>
+}
+
+interface WebService {
+  registerSearchProvider(provider: WebSearchProvider): () => void
+  registerFetchProvider(provider: WebFetchProvider): () => void
+
+  search(request: WebSearchRequest, signal?: AbortSignal): Promise<WebSearchResult>
+  fetch(request: WebFetchRequest, signal?: AbortSignal): Promise<WebFetchResult>
+}
+```
+
+The optional signal is execution control, not business input: `tool-web` passes `exec.signal` directly so turn cancellation, tool timeout, and agent disposal reach provider network requests, stream readers, and expensive decoding. The seam does not pass `ToolExecution` through — that would make `dsh-web` depend on `dsh-tools`.
+
+Provider ids are stable strings and unique within their capability kind. Registering a duplicate search provider id or duplicate fetch provider id fails rather than silently replacing the old provider. Provider registration returns a disposer and follows the existing `ctx.tools.register()` / `ctx.systemPrompt.section()` pattern: the mutation is wrapped in `ctx.effect()` so the registration is torn down with the contributing fiber.
+
+## Provider availability and selection
+
+Provider availability and capability selection are separate concepts, but both stay minimal. A provider reports only whether that concrete implementation is usable by cheap local checks such as credential presence or parseable endpoint config. A provider `available()` must not make network calls.
+
+`LlmService` has no status type at all: availability is expressed as registry membership plus a resolution-time throw. `ctx.web` follows the same discipline. The seam exposes no aggregated capability-status query — `search()` / `fetch()` derive the selection on each call from the configured provider id, the registered providers, and each provider's cheap local `available()` boolean, and a selection failure is the structured `WebError` thrown at execution time. A caller that needs to know whether a capability can run executes and routes that error; nothing is stored as mutable service state.
+
+The boolean is an input to selection, not a health system. `tool-web` never calls a provider's `available()` directly — its only path into the seam is `search()` / `fetch()` — so selection policy has one owner.
+
+Selection must not depend on registration order. Cordis load order, config ordering, and HMR timing are not product semantics.
+
+| Situation | Execution behavior |
+|---|---|
+| A configured provider id is registered and `available() === true` | runs that provider |
+| A configured provider id is not registered | fails with `WEB_PROVIDER_CONFIGURED_MISSING` |
+| A configured provider id is registered but unavailable | fails with `WEB_PROVIDER_CONFIGURED_UNAVAILABLE` |
+| No provider id is configured and exactly one provider for that kind is registered and available | runs that single provider |
+| No provider id is configured and no provider for that kind is registered | fails with `WEB_PROVIDER_UNAVAILABLE` |
+| No provider id is configured and multiple usable providers for that kind are registered | fails with `WEB_PROVIDER_AMBIGUOUS` rather than choosing by registration order |
+| No provider id is configured and providers exist but none are usable | fails with `WEB_PROVIDER_UNAVAILABLE` |
+
+The "single provider auto-selects" rule is for tests, demos, and simple deployments. Product configs set explicit provider ids:
+
+```yaml
+- id: web
+  name: '@deepseek-ai/dsh-web'
+  config:
+    searchProvider: exa
+    fetchProvider: local-http
+
+- id: web-search-exa
+  name: '@deepseek-ai/dsh-web-search-exa'
+
+- id: web-search-perplexity
+  name: '@deepseek-ai/dsh-web-search-perplexity'
+
+- id: web-search-deepseek
+  name: '@deepseek-ai/dsh-web-search-deepseek'
+
+- id: web-fetch-local
+  name: '@deepseek-ai/dsh-web-fetch-local'
+
+- id: tool-web
+  name: '@deepseek-ai/dsh-tool-web'
+```
+
+Operational overrides feed the same explicit selection path: `DSH_WEB_SEARCH_PROVIDER=perplexity` is equivalent to config `searchProvider: perplexity`, not a hidden priority chain inside `dsh-tool-web`.
+
+`ctx.web.search()` and `ctx.web.fetch()` resolve the provider at execution time using the selection rules above. If the selected capability is unavailable, they throw `WebError` with a structured code such as `WEB_PROVIDER_UNAVAILABLE`, `WEB_PROVIDER_CONFIGURED_MISSING`, `WEB_PROVIDER_CONFIGURED_UNAVAILABLE`, or `WEB_PROVIDER_AMBIGUOUS`. If no provider is explicitly configured and no usable provider exists, the execution error is the generic `WEB_PROVIDER_UNAVAILABLE` case; there is deliberately no diagnostic summary of every unavailable provider.
+
+## Search request and result schema
+
+The `web_search` model-facing tool is small. The only model-facing argument is:
+
+- `query`: required string.
+
+`max_results` is NOT exposed to the model. It is a `dsh-tool-web`-layer decision: the tool sets the result bound — the `searchMaxResults` plugin config, default `8` (aligning with OpenCode's Exa default), mirroring `dsh-tool-fs`'s `readLimit` — and passes it to the seam as `maxResults` on the `WebSearchRequest`. Keeping it off the model schema means the model just asks a question and the product controls how much context comes back; the field can be promoted to a model-facing argument later without breaking the seam.
+
+`maxResults` flows tool → seam → provider, and the bound is enforced on the way back:
+
+- `dsh-tool-web` owns the value and puts it on `WebSearchRequest.maxResults`.
+- `ctx.web` passes the request through to the selected provider unchanged.
+- A provider applies `maxResults` at the request layer when its API supports it (Exa's `numResults`), as a cost/latency optimization.
+- `ctx.web` enforces the bound on the result: if a provider returns more than `maxResults` sources — because its API has no result-count control (Perplexity) or ignored the hint — the seam truncates `sources[]` to `maxResults` and sets `WebSearchResult.truncated` to `true` before returning. This makes the bound a single cross-provider guarantee the model-facing layer can rely on, rather than something each provider must remember to honor.
+
+The seam request carries no provider-specific controls — no Perplexity model selection, search recency, domain filters, Exa `livecrawl`, Exa `type`, regional hints, generated-answer budgets, or search depth. Such a field is added only when it has provider-neutral semantics that both the tool schema and selected providers can honor honestly.
+
+```ts
+interface WebSearchRequest {
+  readonly query: string
+  /** Upper bound on returned sources; the seam truncates to it. Omitted = no bound. `dsh-tool-web` always sets it. */
+  readonly maxResults?: number
+}
+
+interface WebSearchResult {
+  readonly content?: string
+  readonly sources: readonly WebSearchSource[]
+  readonly truncated: boolean
+}
+
+interface WebSearchSource {
+  readonly url: string
+  readonly title?: string
+  readonly snippet?: string
+  readonly publishedAt?: string
+}
+```
+
+`content` is optional provider-generated answer text, search context, or summary. `sources[]` is the portable citation surface. A source always has a URL; title, snippet, and `publishedAt` are optional because not every provider returns them. `title` is not required: Perplexity-style citations may provide only URLs, and forcing adapters to invent titles would make the seam lie. `dsh-tool-web` renders a `title ?? hostname(url)`-style fallback label for display. `publishedAt` is an optional publication/crawl timestamp as an ISO-8601 string — Exa returns it as `publishedDate` on each result and Perplexity returns a `date` on search results, so it is real provider data, not derived; the seam carries it as a string and leaves date parsing to the consumer.
+
+Exa search maps each entry of the provider's flat `results[]` into a `WebSearchSource`: `url` ← `url`, `title` ← `title`, `snippet` ← the first `highlights[]` entry (an entry with no highlight has no portable snippet and is dropped), `publishedAt` ← `publishedDate`. Exa returns no provider-generated answer, so `content` is omitted. Perplexity search maps `choices[0].message.content` to `content` and prefers the structured top-level `search_results[]` for `sources[]` — `url` ← `url`, `title` ← `title`, `snippet` ← `snippet` (often empty), `publishedAt` ← `date` — falling back to the URL-only `citations[]` array only when `search_results` is absent (those sources carry just a `url`). If a provider returns fewer structured fields than the seam supports, the adapter omits those optional fields.
+
+Full page retrieval remains the job of `web_fetch(url)`. Search snippets are discovery context, not fetched page bodies.
+
+## Fetch request and result schema
+
+The `web_fetch` implementation is an anonymous public HTTP(S) fetch provider, `local-http`. It fetches bytes from a concrete URL, applies the basic transport hygiene below (http/https-only, credential rejection, byte/time caps, cross-origin redirect blocking), decodes textual content, and returns only the minimal model-useful result: final URL, status code, body, and truncation. It carries no browser cookies, editor credentials, git credentials, internal auth tokens, or implicit access to private services. (Full SSRF / private-network blocking is deferred — see [Deferred work](#deferred-work).)
+
+The seam request stays smaller than OpenCode's model-facing tool:
+
+- `url`: required HTTP(S) URL.
+
+The seam request deliberately does not include a per-call timeout, `format`, `prompt`, or provider-specific extraction controls. Cancellation is the direct optional execution signal, while the fetch provider owns one deployment-configured timeout backstop. `format` is a presentation decision over a fetched resource; `prompt` is a higher-level LLM summarization instruction; extraction APIs such as Firecrawl, Exa, Tavily, or Parallel may not expose a concrete HTTP response. If the product later needs provider-backed page extraction, that is a separate `web_extract` capability or a deliberate widening of this seam — extract semantics are never smuggled into `web_fetch` by making every HTTP field optional.
+
+HTTP status is part of the fetched resource state, not automatically a tool failure. A successful network fetch of a `404` or `500` response returns `WebFetchResult` with the status code and a bounded decoded body when the content type is supported. `WebError` is for failures to safely retrieve or represent the resource: invalid or blocked URL, redirect policy violation, timeout, abort, response too large, unsupported content type, provider failure, or network failure.
+
+```ts
+interface WebFetchRequest {
+  readonly url: string
+}
+
+interface WebFetchResult {
+  readonly url: string
+  readonly statusCode: number
+  readonly body: WebFetchBody
+  readonly truncated: boolean
+}
+
+type WebFetchBody =
+  | { readonly kind: 'html'; readonly content: string }
+  | { readonly kind: 'text'; readonly content: string }
+```
+
+`WebFetchResult.url` is the final URL after allowed redirects. The request URL is already present in `WebFetchRequest`, so there is no separate `requestedUrl`/`finalUrl` pair.
+
+`WebFetchBody` is a closed discriminated union because body kinds require coordinated changes to the seam, provider, and tool rather than independent plugin extension. Exhaustive switches make a new kind fail compilation at every renderer until handled. Separate object arms leave room for kind-specific fields.
+
+The provider owns safe resource retrieval: URL validation, HTTP transport, redirect policy, timeout, abort propagation, byte caps, charset decoding, content-type classification, and binary rejection. `dsh-tool-web` owns presentation: HTML-to-markdown, HTML-to-text, truncation formatting for the model, and future summaries.
+
+The fetch provider's resource controls:
+
+- Only `http:` and `https:` URLs are accepted; credentials in URLs are rejected.
+- Maximum URL length, response byte cap, decoded body character cap, timeout, and redirect hop cap are enforced.
+- Abort signals propagate through network fetches and expensive decoding.
+- Only same-origin redirects are followed automatically; a cross-origin redirect fails with `WEB_REDIRECT_BLOCKED`, requiring a fresh tool call and therefore a fresh provider/permission decision. (Claude Code's WebFetch uses this same model — it does not auto-follow a cross-host redirect; it returns the redirect target to the model for a fresh call.)
+- Requests carry an explicit product user agent rather than silently impersonating a browser.
+
+SSRF / private-network protection (blocking private, loopback, link-local, multicast, and otherwise non-public destinations, with DNS-resolve-then-validate to defeat rebinding and per-hop re-validation on redirects) is **deferred** — see [Deferred work](#deferred-work). Until it lands, `web_fetch` is an SSRF primitive and must not be enabled in a deployment that can reach sensitive internal network targets.
+
+## Tool consumer behavior
+
+`dsh-tool-web` owns two `ToolDefinition`s: `web_search` and `web_fetch`. It owns model-facing JSON schemas, snake_case argument names, prompt sections, result rendering to `ContentBlock[]`, `presentCall`, and `presentResult`.
+
+`dsh-tool-web` must not enumerate providers or call provider `available()` directly. Its only path into the seam is `ctx.web.search()` / `ctx.web.fetch()`. That keeps provider selection in one layer; otherwise the tool package could decide one provider is usable while execution resolves a different state.
+
+Tool registration is a minimal stable sync: on plugin startup the `dsh-tool-web` `Config` (`search?: boolean`, `fetch?: boolean`, both default `true`) enables or disables each web tool; an enabled tool is registered with a fiber-scoped disposer via the effect-based registry; neither tool is disposed merely because its selected provider is missing, unusable, or ambiguous; disposing the `tool-web` fiber tears down its registrations automatically.
+
+Provider availability changes affect execution results and diagnostics, not whether the model-facing schema exists. If a product wants no web tools at all, it disables `dsh-tool-web` or the individual web tool in config; if it wants web tools but the backend is misconfigured, the model sees a structured tool error at execution time.
+
+The prompt guidance explains the semantic split — `web_search` for discovery and current information, `web_fetch` when the model needs the content of a specific URL — and the prompt and tool result tell the model to cite relevant URLs with markdown links.
+
+The model-facing output is text-first because tool results are `ContentBlock[]`, but the seam outcome stays structured so UI presentation and future adapters do not have to scrape rendered text.
+
+## Errors
+
+`dsh-web` defines `WebError extends HarnessError` with stable codes, covering only states that callers may reasonably branch on:
+
+- `WEB_PROVIDER_UNAVAILABLE`
+- `WEB_PROVIDER_CONFIGURED_MISSING`
+- `WEB_PROVIDER_CONFIGURED_UNAVAILABLE`
+- `WEB_PROVIDER_AMBIGUOUS`
+- `WEB_DUPLICATE_PROVIDER`
+- `WEB_INVALID_URL`
+- `WEB_BLOCKED_URL`
+- `WEB_REDIRECT_BLOCKED`
+- `WEB_FETCH_TOO_LARGE`
+- `WEB_FETCH_TIMEOUT`
+- `WEB_ABORTED`
+- `WEB_UNSUPPORTED_CONTENT_TYPE`
+- `WEB_PROVIDER_ERROR`
+
+`WEB_DUPLICATE_PROVIDER` is thrown synchronously from `registerSearchProvider` / `registerFetchProvider` when an id is already registered for that capability kind (the analogue of `LlmService`'s `DUPLICATE_ADAPTER`); it is a registration-time programming error, not an execution outcome, but shares the `WebError` code space so callers see one taxonomy. `WEB_PROVIDER_ERROR` is the catch-all for a provider's own failure surfaced through the seam, including network/transport failure in `web-fetch-local` (DNS, connection refused, TLS); there is deliberately no separate `WEB_NETWORK` code — the provider sets a descriptive message so the model and logs can tell a network failure from a provider API failure.
+
+Tool execution lets these errors flow through `ToolRegistry.execute()`, which already converts `HarnessError` into an error tool result with structured metadata. The model gets a readable error message; hooks, tests, and UI code can route on the stable code.
+
+## Testing
+
+Each layer is pinned at its own seam: the registry/selection/truncation/abort contract and the `WebError` codes in `dsh-web`; per-provider request/response mapping over recorded fixtures (Perplexity fixtures include URL-only citations so the optional source fields stay honest) plus a self-skipping with-key smoke per real provider; real local-HTTP behavior in `web-fetch-local`; and enablement-driven registration, structured execution errors, and result formatting through the real tool registry in `dsh-tool-web`. A real-Loader smoke guards the two export shapes ([postmortem 0001](../../../../docs/postmortem/0001-acp-default-export-drops-inject.md)): `dsh-web` is a default-exported service, while the providers and `tool-web` are namespace plugins where a stray `export default` would drop `inject`.
+
+## Alternatives considered
+
+### Let each provider register its own model-facing tool
+
+This matches the most flexible provider-plugin systems: every provider can expose its full native schema. It is rejected for the harness because it gives provider packages ownership of model-facing names, descriptions, prompt guidance, and result formatting. Multiple search providers would produce duplicate tool names or provider-specific tool names, and the model would learn backend details instead of a stable product capability.
+
+### Put provider dispatch directly in `dsh-tool-web`
+
+This resembles OpenCode's local web search: one stable `websearch` tool dispatches to Exa or Parallel internally. It is acceptable for a small product path but wrong as a harness foundation. The tool package would own provider selection, credentials, request mapping, transport, response parsing, and presentation, making it hard to add Exa and Perplexity without baking their differences into the tool schema.
+
+### Split search and fetch into two seams (`dsh-search`, `dsh-fetch`)
+
+Tempting because the two halves share no request schema and no business logic, so each would map cleanly onto the bash/fs three-package template, and the `Search`/`Fetch` method-pair duplication on `WebService` would disappear. Rejected because the shared machinery — provider-id registry, registration-order-independent selection policy, abort propagation, the `WebError` taxonomy, and the product-facing "how this harness reaches the web" config surface — is real and would otherwise be duplicated across two near-identical seams. One `ctx.web` middle layer gives the product a single thing to inject and configure and gives provider selection one owner. The price is the parallel `searchX`/`fetchX` method pairs, which is accepted deliberately.
+
+### Choose the first registered provider
+
+Rejected. Registration order is not a product policy. It can change with config order, plugin loading, HMR, or refactors. Provider selection must be explicit, or automatic only when exactly one usable provider exists.
+
+### Treat Firecrawl/Exa/Tavily/Parallel extraction as fetch
+
+Rejected for the first version. Those providers often return extracted or summarized content rather than a concrete HTTP response. If the product needs extraction, design `web_extract` or deliberately widen the fetch seam later.
+
+### Mirror Claude Code's `url + prompt` WebFetch shape
+
+Rejected for the seam. `prompt` turns fetch into LLM summarization and couples public-web retrieval to a model provider. The harness seam should fetch and decode deterministically; `dsh-tool-web` can later offer summaries as a presentation mode without making `ctx.web` depend on `ctx.llm`.
+
+## Consequences
+
+**The search schema is deliberately thin.** Exa and Perplexity both expose useful provider-specific controls; a control is added only once it can be defined provider-neutrally and enforced honestly by both tool registration and provider execution.
+
+**Perplexity citations can be sparse.** A citation may be only a URL. Making `title` and `snippet` optional keeps the seam truthful but means `tool-web` renders fallback labels.
+
+**Stable tool registration defers misconfiguration to execution.** Keeping the tool visible is correct when the product enabled web access, but product apps that expect web search should surface the structured `WEB_PROVIDER_CONFIGURED_MISSING` / `WEB_PROVIDER_CONFIGURED_UNAVAILABLE` / `WEB_PROVIDER_AMBIGUOUS` failures loudly so users do not discover setup problems only after the model calls the tool.
+
+**Provider state can change after startup.** A tool can be visible in the request assembled at step start and lose its provider before execution. The execution path resolves again and fails with a structured error.
+
+**Fetch is a network boundary, not just a read-only tool.** `web_fetch` can reach sensitive network targets or exfiltrate data through URLs. Only the basic transport hygiene ships (http/https-only, credential rejection, byte/time caps, cross-origin redirect blocking); SSRF / private-network blocking is deferred (see [Deferred work](#deferred-work)), so until it lands `web_fetch` must not be enabled where it can reach internal targets.
+
+**Large web content can damage context quality.** Providers enforce byte/character caps and report `truncated`; `tool-web` formats bounded model output with clear continuation or follow-up guidance.
+
+## Deferred work
+
+- SSRF / private-network protection for `web_fetch`: block private, loopback, link-local, multicast, and otherwise non-public destinations so `web_fetch` is not an SSRF primitive. Doing it correctly is more than a URL-string check — it needs DNS-resolve-then-connect-to-the-validated-IP (to defeat DNS rebinding / TOCTOU), per-hop re-validation across redirects, and IPv6 edge handling (private ranges, IPv4-mapped addresses). Neither reference implementation surveyed does IP-level blocking (OpenCode does a prefix check then fetches; Claude Code relies on a centralized hostname blocklist plus a "private URLs will fail" prompt), so there is no implementation to copy and this is the harness's only SSRF defense — it warrants its own focused design/spike. Until it lands, `web_fetch` must only be enabled in deployments that cannot reach sensitive internal targets.
+- A `pdf` `WebFetchBody` kind: the `local-http` provider decodes text-extractable PDFs (best-effort, capped, `truncated`) into a `{ kind: 'pdf'; content; pageCount? }` arm, and `tool-web` renders it. This is fetch, not `web_extract` — PDF retrieval is a concrete HTTP 200 plus deterministic local decoding, not provider-side extraction of a non-HTTP resource. Adding it is a coordinated change across `dsh-web` (declare the arm), the provider (decode + narrow "binary rejection" to "reject binary except text-extractable PDF"; scanned/image PDFs needing OCR stay out of scope), and `tool-web` (render). The closed `WebFetchBody` union makes the consumer side fail to compile until the new arm is handled.
+- Provider-backed extraction as a separate `web_extract` capability, rather than widening `web_fetch` silently.
+- Permission policy integration once the deferred permission system lands.
+- Provider-neutral search controls beyond `query` and `maxResults`, once Exa and Perplexity can both honor them honestly.
+
+## Open questions
+
+- Should product app packages probe web configuration at startup (treating `WEB_PROVIDER_CONFIGURED_MISSING`, `WEB_PROVIDER_CONFIGURED_UNAVAILABLE`, and `WEB_PROVIDER_AMBIGUOUS` as fatal when web is explicitly configured), or leave misconfiguration to surface at the first execution?
+- Where should permission policy for public web access live once the deferred permission system lands: a dedicated web permission plugin on `tools/execute`, provider config, or both?

.agents/notes/implemented/architecture/2026-06-26-file-context-as-event-gate.md

@@ -0,0 +1,169 @@
+# Agent Note: Make `dsh-fs-policy` an event-gate plugin, not a method interface
+
+Status: implemented
+
+## Problem
+
+[The split-fs-seam Agent Note](../simplification/2026-06-26-fsspec-style-fs-seam.md) put `ctx.fileContext` between the model-facing tools and the `ctx.fs` provider: `dsh-tool-fs` injects `fileContext` and routes every `read`/`write`/`edit` through its methods. That makes `fileContext` **in-path and mandatory**. The tool cannot reach `ctx.fs` without it, the policy layer owns the fs I/O and the read windowing, and a deployment that does not want observed-state policy cannot simply drop the package — `dsh-tool-fs` would fail to resolve `ctx.fileContext`.
+
+This couples three things that should be separable:
+
+1. **What the tool does** — resolve a path, read a window, write/edit a file. This is the tool's job and needs only `ctx.fs`.
+2. **The freshness/observation policy** — "edit requires a prior read", "write/edit must be based on the version you read". This is the `dsh-fs-policy` plugin's job.
+3. **The recording of observed state** — a side effect that should never block the tool from functioning.
+
+Because the tool calls `fileContext` methods, removing the policy layer is a breaking change rather than a graceful loss of an *add-on*. The policy is load-bearing for the tool to even run, not an opt-in tightening.
+
+## Decision
+
+Invert the control flow. **`dsh-tool-fs` becomes the executor and calls `ctx.fs` directly**; **`dsh-fs-policy` becomes a gate + recorder plugin** that participates through events, never through a method the tool calls and never by registering a `ctx.fileContext` service.
+
+```text
+tool          dsh-tool-fs       executor: resolves, reads windows, writes/edits via ctx.fs;
+                                emits fs policy events; renders results
+policy        dsh-fs-policy  plugin: listens to fs/write-intent +
+                                fs/edit-intent (single-slot waterfall) and fs/observed
+                                (emit) events; adds observed-state + freshness.
+provider seam dsh-fs            ctx.fs: text IO + ATOMIC mutation primitives whose version
+                                guard is OPTIONAL; owns the fs policy event vocabulary
+provider      dsh-fs-local      local implementation of ctx.fs
+```
+
+The model is additive: bare `ctx.fs` performs atomic, unconstrained text I/O, while `dsh-fs-policy` adds observed state, read-before-edit, and version guards. Removing the policy therefore leaves the tools usable but unconstrained. Shipped agent configs load the policy; the bare mode exists to keep policy optional at the service boundary, not as the normal deployment stance.
+
+`dsh-tool-fs` no longer injects `fileContext`. It injects `fs` and `tools`/`systemPrompt`.
+
+## The policy is enforced by provider CAS, not by `dsh-fs-policy` stat
+
+`dsh-fs-policy` enforces "you must write/edit based on the version you read" **without ever calling `stat` or comparing versions itself**. It supplies the observed version as the CAS basis and lets the provider's mutation critical section detect staleness:
+
+- "Have you read this file?" is the one thing `dsh-fs-policy` decides locally — a `WeakMap` lookup, no I/O. No record ⇒ `FS_NOT_OBSERVED`.
+- "Is the version you read still current?" is decided **inside `ctx.fs.editText`/`writeText`**, in the same atomic lock that performs the read-match-rename. `dsh-fs-policy` passes `vObserved` as the expectation; the provider raises `FS_STALE_VERSION` if the file has moved on.
+
+This is deliberate. If `dsh-fs-policy` stat-ed and compared versions in its waterfall handler, there would be a TOCTOU gap between that check and the tool's actual write — the file could change in between, so the check would be a false guarantee that the provider's lock has to back up anyway. Putting the version check in the provider's critical section is both race-free and zero extra `stat`. So `dsh-fs-policy` does **no** filesystem I/O; the "must be based on the latest read" guarantee is *realized* by CAS, and `dsh-fs-policy` only chooses the basis (`vObserved`) and gates on prior observation.
+
+## Provider contract change: the version guard is optional
+
+For the bare provider to be unconstrained, the version guard on its two mutations becomes **optional** — present ⇒ guarded, absent ⇒ unconditional:
+
+```ts ignore-check
+// writeText: expected is now optional. The FsWriteIntent union is UNCHANGED.
+writeText(target: FsTarget, content: string, expected?: FsWriteIntent, signal?: AbortSignal): Promise<FsWriteOutcome>
+//   undefined          → unconditionally create-or-overwrite (bare default)
+//   createIfAbsent     → create only, reject an existing file (dsh-fs-policy, unobserved)   [unchanged]
+//   replaceIfVersion   → overwrite only at the observed version, else FS_STALE_VERSION    [unchanged]
+
+// editText: expected becomes optional (was the required { version: FsVersion }).
+editText(target: FsTarget, edit: FsEditRequest, expected?: { version: FsVersion }, signal?: AbortSignal): Promise<FsEditOutcome>
+//   undefined    → unconditionally replace literal text in the current content (bare default);
+//                  a missing target still reports FS_STALE_VERSION
+//   { version }  → edit only at that version, else FS_STALE_VERSION (the current behavior)
+```
+
+The `FsWriteIntent` union itself does not change — the third "unconditional" state is expressed by *omitting* `expected`, so both mutations share one symmetric shape (`expected?`: omit = no guard, present = guarded). This keeps full backward compatibility for the guarded paths `dsh-fs-policy` uses; only the previously-impossible "no guard" case is new, and it is the bare-provider default. The mutation still runs inside the backend's per-target lock either way, so an unconditional write/edit is still atomic (no torn files); "unconditional" drops the *version* precondition, not the atomicity. `editText` reports a missing target as `FS_STALE_VERSION` on both guarded and unguarded paths, preserving one edit failure code for "the target cannot be edited at this moment".
+
+## Event vocabulary (owned by `dsh-fs`)
+
+The events live in `@deepseek-ai/dsh-fs`, not in `dsh-fs-policy`. This is forced by the decoupling contract: `dsh-tool-fs` is the emitter, so it must reference the event types, and it must keep compiling even though `dsh-fs-policy` no longer provides a method service. `dsh-fs` is the package both `dsh-tool-fs` and `dsh-fs-policy` already depend on, so it is the only home that lets the emitter and the policy listener share a vocabulary without the emitter depending on the policy plugin.
+
+These events carry existing `dsh-fs` vocabulary (`FsTarget`, `FsVersion`, `FsWriteIntent`) plus an opaque actor — not model-facing concepts (no line windows, numbered lines, or rendered footers leak down).
+
+**The two `fs/*` decision events are single-slot, first-wins waterfalls.** `dsh-fs-policy` returns without calling `next()`, so it owns the slot in the default deployment; a listener registered earlier or with `prepend` would replace that policy. Permission, audit, and sandbox concerns remain on the composable `tools/execute` waterfall.
+
+The actor is typed `object` in `dsh-fs` — a pure opaque carrier the provider seam never reads or narrows. The owner-derivation (`actor.agent?.session`) and the `{ agent?: { session? } }` structural shape stay entirely inside `dsh-fs-policy`, which narrows the `object` actor to that shape in its listeners. `dsh-fs` owns the event names and the fs vocabulary; it does NOT own the policy layer's runtime owner structure.
+
+```ts
+import type { FsTarget, FsVersion, FsWriteIntent } from '@deepseek-ai/dsh-fs'
+
+interface Events {
+  /**
+   * Single-slot decision: produce the write expectation for the next
+   * ctx.fs.writeText. The default returns undefined (unconditional create-or-
+   * overwrite — the bare provider). The policy listener returns createIfAbsent
+   * (unobserved) or { kind: 'replaceIfVersion', version: vObserved } (observed).
+   * The listener does NOT call next(): one decision, not a composable chain. @mode waterfall
+   */
+  'fs/write-intent'(target: FsTarget, actor: object | undefined, next: () => FsWriteIntent | undefined | Promise<FsWriteIntent | undefined>): Promise<FsWriteIntent | undefined>
+  /**
+   * Single-slot decision: produce the optional version guard for the next
+   * ctx.fs.editText. The default returns undefined (unconditional edit of the
+   * current content — the bare provider; no stat). The policy listener returns
+   * { version: vObserved }, or throws FS_NOT_OBSERVED if the actor is unset or
+   * has not observed the target. Does NOT call next(): one decision. @mode waterfall
+   */
+  'fs/edit-intent'(target: FsTarget, actor: object | undefined, next: () => { version: FsVersion } | undefined | Promise<{ version: FsVersion } | undefined>): Promise<{ version: FsVersion } | undefined>
+  /**
+   * Record that an actor observed a target at a version, after a successful
+   * read/write/edit. Fire-and-forget (plain emit). Listeners MUST be
+   * synchronous, side-effect-only recorders (`dsh-fs-policy`'s is a WeakMap
+   * write); the tool does not guard the emit, so a throwing listener surfaces as
+   * the tool's isError result. No listener ⇒ nothing recorded.
+   * @mode emit
+   */
+  'fs/observed'(target: FsTarget, version: FsVersion, actor: object | undefined): void
+}
+```
+
+The `fs/*` decision events are **unbound waterfalls dispatched by the tool** (like `agent/request`, which the loop dispatches with no `this`), not service-bound waterfalls (like `llm/stream`). The dispatcher is the `dsh-tool-fs` plugin, which is not a service.
+
+## Tool contract (`dsh-tool-fs`)
+
+The tool keeps its model-facing schemas (`read`/`write`/`edit`, byte-for-byte unchanged) and prompt sections. The prompt guidance stays policy-first because a deployment loading the fs tools is expected to also load `dsh-fs-policy`: the model is still told to read before overwriting or editing, and any wording that says the "backend" requires that should be corrected to say the fs-policy plugin requires it. The bare-provider fallback does not change the prompt stance.
+
+`dsh-tool-fs` gains the executor responsibilities relocated from the old `fileContext` method service, including **read rendering** (`read-render.ts`: `buildWindow` + `formatReadOutput`, `READ_MAX_BYTES`, `READ_MAX_LINE_LENGTH`, `FileReadOutcome`/`FileTextLine`, plus `STREAM_MIN_SIZE` in `read.ts`), which is the tool's rendering detail now that the tool owns the read. Those read-rendering types and helpers move into `dsh-tool-fs`; the policy plugin must not remain a type dependency for the tool.
+
+`dsh-tool-fs` is a single root plugin that registers all three tools (`read`/`write`/`edit`), mirroring `dsh-tool-bash`. It injects `fs` (plus `tools`/`systemPrompt`), never `fileContext`. (The original proposal also exposed each tool as a `/read`/`/write`/`/edit` subpath plugin for focused deployments; that was dropped on implementation — no consumer needed a single-tool deployment, and the subpath publishing forced bespoke `tsdown`/`tsconfig`/`files`/workspace-constraint handling no sibling tool package carries. The per-tool registration helpers (`applyReadTool`/`applyWriteTool`/`applyEditTool`) remain internal modules the root plugin composes.)
+
+`stat` budget is minimized by letting the waterfall produce the expectation lazily — the bare default returns `undefined` (no guard) and never stats:
+
+- **read** — one `stat` (type + size routing + version), then `readText`/`streamText`, then `buildWindow`, then an `emit('fs/observed', target, info.version, exec)`. The post-read confirming `stat` from the old `fileContext.read` is dropped; a writer racing between the routing stat and the read can at worst make a *later* guarded edit spuriously `FS_STALE_VERSION` (fail-closed: the model re-reads, never writes against the wrong version, since `editText` re-checks in its lock).
+- **write** — `expectation = await ctx.waterfall('fs/write-intent', target, exec, () => undefined)`, then `ctx.fs.writeText(target, content, expectation)`, then an `emit('fs/observed', target, outcome.version, exec)`. **Zero stat in the tool** with or without `dsh-fs-policy`.
+- **edit** — `expectation = await ctx.waterfall('fs/edit-intent', target, exec, () => undefined)`, then `ctx.fs.editText(target, edit, expectation)`, then an `emit('fs/observed', target, outcome.version, exec)`. **Zero stat in the tool** in both cases: the bare default is `undefined` (unconditional edit), so the tool never stats to manufacture a basis. If the target is absent, the provider reports `FS_STALE_VERSION` even on the unguarded path.
+
+The tool passes `exec` (the tool-execution context) as the `actor` argument on every dispatch, so `dsh-fs-policy` can derive its observed-state owner. The tool does not know whether the policy plugin is present: it always provides the bare default behavior in the `next` thunk, and `dsh-fs-policy` short-circuits the thunk before it runs in the default deployment.
+
+**`fs/observed` fires after a successful operation.** Its listeners must be synchronous, non-throwing recorders; the tool does not guard the plain emit, so a throwing listener would report failure after a mutation already succeeded. Async or fallible observation needs a separate event contract.
+
+## Policy plugin contract (`dsh-fs-policy`)
+
+`dsh-fs-policy` is a plugin, not a service. It does not register `ctx.fileContext`, has no public method surface, and exposes no `read`/`write`/`edit`/`resolve` methods. It attaches three listeners via `ctx.on()` registrations (each returning a disposer for HMR). It keeps the observed-state `WeakMap<owner, Map<targetKey, { version }>>` and the structural owner derivation (narrowing the event's opaque `object` actor to its own `{ agent?: { session? } }` shape), but does not inject `fs` — every handler operates only on its own `WeakMap`, never on `ctx.fs`.
+
+- `fs/write-intent` listener: `prior = getObserved(owner, key)`; return `prior ? { kind: 'replaceIfVersion', version: prior.version } : { kind: 'createIfAbsent' }`. It does NOT call `next()`: it fully owns the single decision slot.
+- `fs/edit-intent` listener: `prior = getObserved(owner, key)`; if no `owner` or no `prior`, throw `FS_NOT_OBSERVED`; else return `{ version: prior.version }`. Also does not call `next()`.
+- `fs/observed` listener: `record(owner, key, version)`.
+
+An observed-state entry is the **prior-observation record**: a successful `read`, `write`, OR `edit` all emit `fs/observed` and record `{ version }`, so the entry's presence means "this owner has observed this target at this version", not narrowly "has read it". This is what lets a create-then-edit or edit-then-edit sequence work without an intervening re-read: the mutation refreshes the recorded version to its own result, so the next edit's basis is the version it just produced. `FS_NOT_OBSERVED` rejects only an edit with NO prior observation of any kind. The owner is derived structurally from `{ agent?: { session? } }`; disposal drops all state (HMR safety).
+
+`dsh-fs-policy` is now a pure policy/recording plugin with no service surface — it influences the world only through the event seam. That is what removes the method coupling from `dsh-tool-fs`.
+
+## Bare-provider behavior (no `dsh-fs-policy`)
+
+This is not the intended deployment stance — a config loading the fs tools is expected to also load `dsh-fs-policy`. It is the unconstrained provider floor that exists once the tool is no longer coupled to a policy method service. With `dsh-fs-policy` absent, every `fs/*` waterfall falls through to its `undefined` default and `fs/observed` has no listener:
+
+- **read** is identical (it never needed policy; it only emits a now-unheard `fs/observed`).
+- **write** unconditionally creates-or-overwrites: `expected` is `undefined`, so `writeText` writes whether or not the file exists and whatever its current version. No read-first requirement, no version check.
+- **edit** unconditionally replaces literal text in the file's current content: `expected` is `undefined`, so `editText` matches and rewrites without a version guard or a read-first requirement (`FS_EDIT_NOT_FOUND`/`FS_AMBIGUOUS_EDIT` still apply — those are about the literal match, not freshness). A missing target still reports `FS_STALE_VERSION`, matching the guarded edit path's "cannot edit this target now" code.
+
+Both mutations are still atomic (the backend's per-target lock is unconditional). What is simply *absent*, not lost, is the policy `dsh-fs-policy` would add: observed-state, read-before-edit, and version-guarded write/edit. Loading `dsh-fs-policy` layers those constraints on by having its listeners return guarded `expected` values instead of `undefined`; nothing in the bare provider changes.
+
+## Supersedes
+
+This amends — does not reverse — [the split-fs-seam Agent Note](../simplification/2026-06-26-fsspec-style-fs-seam.md). The four-layer split, the provider contract, and the freshness *policy* are all kept. What changes is the **coupling between the tool and the policy layer**: a mandatory method service became a plugin-owned event gate, and the fs I/O + read windowing moved from `fileContext` up into `dsh-tool-fs`. The split-fs-seam Agent Note's description of `dsh-tool-fs` injecting `fileContext` and of `fileContext` owning `read`/`write`/`edit` was updated to match in the same change.
+
+## Verification
+
+Tests pin both paths: without `dsh-fs-policy`, the root tool plugin boots against `dsh-fs-local`, and read, create, overwrite, and unread edit succeed; with the policy, unread edit returns `FS_NOT_OBSERVED` and unread overwrite is gated by `createIfAbsent`. A later intent listener is not reached after the policy decides. Stale edits fail through provider CAS while the policy performs no `stat`; the tool budgets remain one `stat` for read and zero for write or edit on either path. Model-facing schemas remain byte-for-byte unchanged, so snapshots do not change.
+
+## Alternatives considered
+
+- **Keep `ctx.fileContext` as an in-path method service** — the shape [the split-fs-seam Agent Note](../simplification/2026-06-26-fsspec-style-fs-seam.md) first landed; rejected because the tool could not run without the policy layer, making policy load-bearing for basic operation instead of an opt-in tightening.
+- **Policy-side version checking** (`dsh-fs-policy` stats and compares in its waterfall handler) — rejected for the TOCTOU gap between that check and the tool's actual write; the provider's mutation critical section is the only race-free place, so the policy only chooses the CAS basis and gates on prior observation.
+- **Per-tool `/read`/`/write`/`/edit` subpath plugins** — dropped on implementation: no consumer needed a single-tool deployment, and subpath publishing forced bespoke `tsdown`/`tsconfig`/`files`/workspace-constraint handling no sibling tool package carries; the per-tool registration helpers remain internal modules the root plugin composes.
+
+## Consequences
+
+- **Event indirection over a method call.** A waterfall + emit is less direct than `await ctx.fileContext.edit(...)`. The payoff is removing the tool-to-policy method dependency while keeping the default policy plugin; the cost is one more event vocabulary to learn. Mitigated by keeping the three events narrow and documenting the default-thunk semantics on each.
+- **Policy events in the storage seam.** `dsh-fs` gains two version-decision events plus a recording event though it is "just storage". This is the price of decoupling (the emitter cannot depend on the policy plugin). The events carry only `dsh-fs` vocabulary plus an opaque `object` actor and no model-facing concepts, so the seam stays free of line-window/observation policy types and of the agent/session owner structure.
+- **Single policy occupant, first-wins by convention.** The `fs/write-intent`/`fs/edit-intent` slots hold exactly one decider; the first-registered (or `prepend`ed) listener wins and the rest are short-circuited. `dsh-fs-policy` owning the slot is a deployment convention, not an event-enforced invariant — a second decider registered first would bypass it. This is acceptable because a second fs-version-policy decider is a misconfiguration, not a feature. If a future need for *layered* fs version policy appears, it is a new Agent Note (a composable value-passing seam), not a silent second listener on these events. Layered permission/audit/sandbox interception already has its home on `tools/execute`.
+- **Dropping the post-read confirming stat** makes a follow-up *guarded* edit occasionally fail-closed (`FS_STALE_VERSION` → re-read) under a read/write race. This is a UX nicety lost, never a correctness hole; the provider lock still prevents wrong-version writes.
+- **The bare provider does no read-before-write/edit and no version check.** A deployment without `dsh-fs-policy` lets the model overwrite or edit any existing file unconditionally. This is the deliberate meaning of keeping the tool independent of a policy service: the safety disciplines live in the `dsh-fs-policy` plugin. A deployment that omits it is opting into an unconstrained filesystem on purpose; that is not the intended stance for a config that ships the fs tools.

.agents/notes/implemented/architecture/2026-06-30-bash-stdin-env-trusted-plugin-surface.md

@@ -0,0 +1,31 @@
+# Agent Note: stdin + extra env on the bash seam
+
+Status: implemented
+
+## Problem
+
+The hooks subsystem runs external hook commands the way Claude Code and Codex do: a hook is a shell command that receives its event payload as **JSON on stdin** and reads context from a handful of **environment variables** (`CLAUDE_PROJECT_DIR`, `CLAUDE_PLUGIN_ROOT`, `PLUGIN_ROOT`, …). The harness already has a perfectly good command runner behind the `ctx.bash` capability seam ([dsh-bash](../../../../packages/bash/bash) → [dsh-bash-local](../../../../packages/bash/bash-local)), with process-group kills, output truncation/spill, and a credential scrub. Reusing it for hook execution means a hook bridge does not re-implement subprocess plumbing — but the seam had no way to write stdin or set extra env. This Agent Note adds those two inputs.
+
+`stdin` and `env` do not create a new model capability because ordinary shell syntax already supplies both. Ambient credentials are protected by `dsh-bash-local`'s child-environment scrub, not by hiding these seam fields; model tool arguments are static JSON and do not expand shell variables. The fields therefore serve trusted in-process callers, such as hook bridges, that need to pass structured input and `CLAUDE_*` variables without embedding them in model-visible shell text. See [defensive-patterns.md](../../../../docs/defensive-patterns.md) for the ambient-environment rule.
+
+## Decision
+
+Add `stdin?: string` and `env?: Record<string, string>` to **both** `BashExecRequest` (the model-/plugin-facing request) and `BashExecSpec` (the resolved spec `run`/`start` act on), and thread them through `dsh-bash-local`: `resolve()` carries them verbatim, `run()`/`start()` pass them to `runBash`, which writes the bytes to the child's stdin and merges the extra env.
+
+Three deliberate choices:
+
+1. **The model-facing tool omits `stdin` and `env`.** Shell syntax already covers those needs, so duplicate parameters would add surface without authority separation. The tool builds requests only from declared model arguments, signal, and owner; trusted in-process callers may set the seam fields directly. Harness-owned variables use the separate `dshEnv` channel from the [managed environment decision](../feature/2026-07-10-agent-session-identity-and-log-location.md), so ordinary `env` cannot replace them.
+
+2. **`env` merges AFTER the credential scrub, so an explicit caller entry wins even on a credential-shaped name.** The later managed-namespace decision reserves `DSH_*`: ambient entries are removed, ordinary `env` cannot set them, and trusted `dshEnv` merges last. The complete order is `scrub(process.env, including DSH_*)` → `ENV_OVERRIDES` → ordinary `env` → `dshEnv`.
+
+3. **`stdin`/`env` are required-absent-OK (plain optional) on the resolved spec, NOT required-but-nullable like `owner`.** `owner` is required-but-nullable because a *silently* missing owner yields an unowned, cross-session-readable task — a security footgun that a visible `undefined` guards against. `stdin`/`env` have no such hazard: a missing one means "no stdin / no extra env", which is the safe, ordinary case (every model-driven call). So they stay plain optionals, matching `signal`.
+
+`dsh-bash-local` creates a stdin pipe only when bytes are supplied; otherwise fd 0 remains `/dev/null`, preserving prior behavior. It writes the bytes and closes the pipe. `EPIPE` from a child that exits without reading is ignored because command exit and output determine the result.
+
+## Alternatives considered
+
+**Configurable ambient-secret scrub.** Rejected as speculative. Trusted callers can explicitly provide required values after the scrub without weakening the default ambient protection.
+
+## Consequences
+
+Hook bridges pass JSON payloads and hook-specific variables through the existing bash seam, retaining its process-group, truncation, and spill behavior. The model surface remains unchanged, and the bash tool remains the sole owner of model-call request construction. The vocabulary lives in [the bash data-structure reference](../../../../docs/core-data-structures/bash.md).

.agents/notes/implemented/architecture/2026-06-30-event-domain-semantics.md

@@ -0,0 +1,37 @@
+# Agent Note: Event-domain semantics — session is the fact log, agent is the live surface
+
+Status: implemented
+
+## Problem
+
+The harness extends the agent loop through a Cordis event taxonomy (see [the microkernel event-taxonomy Agent Note](2026-06-11-microkernel-event-taxonomy.md)). As that taxonomy grew, the line between the three event domains blurred:
+
+- `session/*` carries the durable, event-sourced log (`SessionEventMap`).
+- `agent/*` carries live runtime signals that hand a plugin the `Agent` handle.
+- `tools/*` carries the tool registry + execution seam.
+
+Two problems motivated pinning the semantics down. First, several turn/step boundaries existed BOTH as a durable `SessionEvent` (`turn/start`, `turn/end`, `step/start`, `step/end`) AND as a mirrored `agent/*` emit (`agent/turn-start`, `agent/turn-end`, `agent/step-start`, `agent/step-end`). A consumer had two sources of truth for the same fact, and every lifecycle change had to update both. Second, the upcoming Hooks subsystem needs ONE coherent, documented surface to subscribe to — a plugin author (and the Claude Code / Codex hook bridges built on top) must know, without reading the loop, whether to listen on a session event or an agent event, and why.
+
+This vocabulary is the foundation for interception decisions, the durable `hook/*` log, and the Claude Code and Codex bridges.
+
+## Decision
+
+**Three domains, one job each, with a single boundary rule.**
+
+- **`session/*` — the durable, replayable FACT log.** Owns `SessionEventMap`; every entry is JSON-only (no live objects). One `session/event` emit per append, plus the `session/flush` parallel durability checkpoint. It is also the live transcript feed: a consumer that wants to render or react to what happened subscribes here, so live rendering and `session/load` replay share one path.
+- **`agent/*` — the LIVE runtime surface.** Always carries the live `Agent`. Two shapes: INTERCEPTION waterfalls (`agent/request`, `agent/step-result`, `agent/turn-continuation`) that mutate or veto, and TRANSIENT emits (`agent/status`, `agent/error`, `agent/created`/`agent/disposed`, `agent/queued`) that notify with the `Agent` in hand. Turn and step BOUNDARIES are NOT here — they are durable session events read off `session/event`, and so are the token stream (`assistant/chunk`) and mid-turn steering (`steering/message`).
+- **`tools/*` — the tool registry + execution seam.**
+
+**The boundary rule:** a durable, replayable fact is a `SessionEvent`; a live interception or a transient/live-object signal is an `agent`/`tools` Cordis event. A turn or step boundary is a durable fact, so it lives in the session log and is read off the `session/event` feed — it is NOT mirrored as an `agent/*` emit.
+
+**Applying the rule to the boundary twins:** all four boundary mirrors — `agent/turn-start`, `agent/turn-end`, `agent/step-start`, `agent/step-end` — are **REMOVED**. No production consumer needs the live `Agent` at a boundary: the ACP bridge settles from `session/event` `turn/end` plus `agent/status`, and the only turn-mirror consumer (`dsh-ui-stdio`, a disposable test REPL) renders boundaries from `session/event` while retaining its live target object for the fixed `main` label. The step mirrors were removed first (they had no consumer at all); the turn mirrors followed once ui-stdio was migrated — see [the remove-boundary-mirror-events Agent Note](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md), which owns that decision. Removing the emits also simplifies the loop's `closeStep`/`closeTurn` (one append each, no paired emit).
+
+## Consequences
+
+- The loop no longer emits any boundary mirror; `closeStep` appends `step/end` only and `closeTurn` appends `turn/end` only. `Session.append` owns post-commit observer containment, so a throwing boundary observer cannot change the turn outcome or starve later consumers; an acceptance or internal validation failure still escapes before the boundary enters the log.
+- Tests that observed boundaries via the removed emits now observe the durable `turn/start`/`turn/end`/`step/start`/`step/end` session events — the behavior they pin (boundary ordering, step counting) is unchanged; only the feed they read moved to the canonical one. The tests that exercised a *throwing turn-boundary emit listener* were deleted, because that code path no longer exists (there is no emit to throw from). Per [AGENTS.md "tests document behavior, not golden truth"](../../../../AGENTS.md), the behavior and its test moved (or died) together.
+- The loop marks the step open (`stepOpen = true`) only after `append('step/start')` returns. Internal dispatch validation runs before the log push and may reject without opening a step; post-commit `session/event` observer failures are contained inside `Session.append`. The marker therefore represents exactly the committed boundary that owes a later `step/end`.
+- The full realization of this is [the simplification Agent Note "Stop mirroring durable boundaries as agent events"](../simplification/2026-06-20-remove-agent-boundary-mirror-events.md): all four boundary mirrors are removed and every consumer reads boundaries off `session/event`. `agent/steering` (not a boundary mirror) stayed outside that Agent Note's scope and was removed by its own follow-up, [Remove the `agent/steering` mirror emit](../simplification/2026-07-04-remove-agent-steering-mirror.md) — it mirrored the durable `steering/message`.
+- The cordis events catalog (`docs/cordis-catalog/events.md`) is regenerated to drop the mirror events.
+
+<!-- agent-note-format: alternatives-not-recorded (pre-format Agent Note) -->

.agents/notes/implemented/architecture/2026-07-02-fs-per-session-cwd.md

@@ -0,0 +1,32 @@
+# Agent Note: Resolve filesystem paths against the caller's session cwd
+
+Status: implemented
+
+## Problem
+
+The ACP bridge gives every session its own workspace: `session/new` records the editor's project directory as `SessionHeader.cwd`, and `dsh-tool-bash` defaults each bash call's `workdir` to the calling agent's `session.header.cwd` (see [the per-session cwd Agent Note work in `packages/ui/acp`](../../../../packages/ui/acp) and `resolveWorkdir` in `dsh-tool-bash`). So a bash command in session A runs in A's project, and in session B runs in B's — one server process, N workspaces.
+
+Filesystem resolution used one plugin-load cwd while bash used the session project directory. Relative paths therefore disagreed whenever the editor project differed from the server launch directory; snapshots hid the bug by making those paths identical.
+
+## Decision
+
+Thread the caller's session cwd into path resolution, exactly as `dsh-tool-bash` already does for `workdir`. The **caller** (the tool) supplies the cwd; the provider does not read a session or agent.
+
+- `FileSystem.resolve` accepts `resolve(path: string, opts?: { cwd?: string; signal?: AbortSignal }): Promise<FsTarget>`. `opts.cwd` is the base a RELATIVE `path` resolves against; an absolute `path` ignores it; omitting `opts.cwd` uses the backend's own default. `opts.signal` cancels resolution when the backend performs I/O. The options object keeps both caller-owned resolution controls together without positional growth.
+- `dsh-fs-local.resolve` uses `resolveLocalTarget(opts?.cwd ?? this.config.cwd, path)`. `config.cwd` stays the default for a caller that supplies none (non-ACP / no-session use, and the single-session stdio demo where `process.cwd()` IS the workspace).
+- `dsh-tool-fs`'s `read`/`write`/`edit` derive the session cwd through a shared `sessionCwd(exec)` helper (`exec.agent?.session.header.cwd`, mirroring bash's `resolveWorkdir`) and pass it to `resolve`. A non-agent / headerless caller yields `undefined`, so the backend applies its default.
+
+## Alternatives considered
+
+### Why the caller supplies the cwd (not the provider)
+
+The provider seam must not depend on `dsh-agent` / `dsh-session` — it is a text-storage backend that a sandboxed or remote implementation also satisfies, and those have no notion of an "agent session". The tool already receives the `ToolExecution` (`exec`), which carries the agent, so the tool is the right place to project `exec → cwd` and hand the provider a plain string. This is the "explicit > implicit at package seams" convention: the base directory arrives as an explicit argument the provider acts on, not smuggled in by having the provider reach into a session it should not know about. It also matches `dsh-tool-bash` one-to-one, so the two model-facing file surfaces resolve paths identically.
+
+The default lives in ONE place — the provider's `config.cwd`. `sessionCwd` returns `undefined` rather than `process.cwd()` when there is no session, so the tool never manufactures a base the provider would otherwise choose.
+
+## Consequences
+
+- In the ACP demo the fs tools and bash now agree on each session's workspace; an editor can open any project folder and both tool families act on it.
+- No change to `FsTarget` identity: `targetKey` is still the realpath of the resolved absolute path, so observed-state keying and symlink identity are unaffected — a correct per-session cwd produces the same key bash targets.
+- Backward compatible: every existing `resolve(path)` call (all in tests) keeps working; the new argument is optional.
+- The single-session stdio demo is unaffected: it supplies no session cwd (its agent's session has no `cwd`), so resolution falls back to `config.cwd = process.cwd()`, which is the workspace.

.agents/notes/implemented/architecture/2026-07-02-result-time-applied-hunk-diffs.md

@@ -0,0 +1,59 @@
+# Agent Note: Result-time applied-hunk diffs for file mutations
+
+Status: implemented
+
+## Problem
+
+The [tagged render-intent union](2026-07-02-tool-render-intent-union.md) gave `dsh-tool-fs` write/edit a `card:'diff'` at CALL time, derived purely from the tool's args: write ⇒ `{oldText:null, newText:content}` (the whole new file), edit ⇒ `{oldText:old_string, newText:new_string}` (the bare replaced snippet). An editor renders that as an inline diff, but it is a **context-free** diff — the bare `old_string`→`new_string` with no surrounding lines, and a `replace_all` that touched five scattered sites still renders as one snippet pair.
+
+Driving `claude-agent-acp`'s own ACP bridge shows what a full editor diff looks like: after the mutation applies, it emits a SECOND `tool_call_update` whose diff is the **applied hunk with ±3 context lines** (and one hunk per changed site for `replace_all`), reconstructed from the tool's `structuredPatch`. That result-time hunk is what makes Zed show the change *in place* in the file rather than as a floating snippet. Our tools stopped at the call-time snippet; the completed result carried only the plain "updated successfully" text, no diff.
+
+The obstacle is a seam boundary: `presentResult(args, result)` is a **pure function of `args` + the model-facing `result` (`{content, isError}`)** — it runs on live streaming AND on session-log replay, so it must be replay-deterministic and cannot do I/O. It never sees the file's before/after content, and `FsEditOutcome`/`FsWriteOutcome` carried only a replacement count + version, not the text. So there was no way to compute — or even carry — an applied hunk to the presenter.
+
+## Decision
+
+Add a **persisted, tool-private presentation channel** so a tool's `execute` can attach a result-time render payload that survives replay, and use it to carry the applied-hunk diff.
+
+### 1. A `meta` channel on the tool result (core)
+
+`ToolDefinition.execute` may now return either its model-facing `ContentBlock[]` (unchanged, the common case) OR `{ content: ContentBlock[]; meta?: unknown }`:
+
+```ts ignore-check
+type ToolExecuteReturn = ContentBlock[] | { content: ContentBlock[]; meta?: unknown }
+```
+
+`meta` is tool-owned `unknown` that the core persists without interpretation. `Session.append` rejects non-JSON values, and replay passes the stored payload back to `presentResult`; presentations therefore reproduce without I/O or recomputation. Runtime validation avoids adding a shared serializable-value dependency to the tools core.
+
+This is the general shape ("a tool attaches durable result presentation"), not an fs-specific one — any tool can use it.
+
+### 2. The tool computes the hunk; the backend returns before/after (fs)
+
+Per the [capability-seam split](2026-06-13-capability-seams.md), the storage backend returns only **storage facts** and the model-facing tool owns **presentation**:
+
+- `dsh-fs` widens `FsEditOutcome` with `{ before: string; after: string }` and `FsWriteOutcome` with `{ before: string | null; after: string }` (`before: null` ⇒ a create, or an existing-but-undiffable binary/non-UTF-8 file). The local backend already holds both texts at write time; it returns them as raw LF-normalized text, with **no diff/UI concept** entering the seam.
+- `dsh-tool-fs` stores contextual hunks in `meta: { diffs: FileDiff[] }`. Successful mutations always complete with a diff card because ACP result content replaces the pending card: creates or unchanged overwrites fall back to an args-derived whole-file diff, while edits use applied hunks. Failed mutations carry no diff metadata and render their error normally.
+
+### 3. The bridge renders a `diff` result card
+
+`ToolResultView` gains a `DiffResultView { card:'diff'; title?; diffs: FileDiff[] }`; the bridge's result-side `switch (view.card)` gets a `diff` arm emitting the `{type:'diff'}` `ToolCallContent` blocks (mirroring the call-side arm). An ACP `tool_call_update.content` REPLACES the call's content in an editor, so the result diff **supersedes** the call-time snippet (and keeps the model-facing result text from clobbering it) — the two-update sequence (call snippet, then result diff) matches `claude-agent-acp` exactly.
+
+## Alternatives considered
+
+**Hand-rolling or vendoring the diff algorithm.** Contextual hunks have established edge cases, so `dsh-tool-fs` uses the typed [`diff`](https://www.npmjs.com/package/diff) package and normalizes `structuredPatch` output in one module. The repository's vendoring policy applies to its framework source, not every leaf utility.
+
+## Consequences
+
+`tool/result` events may now carry a tool-private `meta` payload — part of the on-disk vocabulary, runtime-gated to JSON by `Session.append` — and any tool can attach durable result presentation without another core change. The diff card reproduces on session reload and snapshot replay for free: it is read back from the log, never recomputed. The costs: an overwrite holds both the prior and new text in memory to compute a UI-only hunk (`TODO(overwrite-diff-bound)`), and `dsh-tool-fs` carries a small, well-known runtime dependency.
+
+## Non-goals
+
+- **Live incremental diff streaming.** The hunk is computed once, after the mutation completes; there is no per-keystroke diff.
+- **Diffing a binary/non-UTF-8 overwrite.** `before` is `null` for such a file (it has no text diff basis); the write still succeeds and the result renders a whole-file diff (`oldText: null`) rather than a contextual hunk.
+- **Rename/move diffs.** Only content diffs of a single resolved path.
+- **Bounding the overwrite diff basis.** An overwrite reads the whole prior file into memory to compute the contextual hunk (on top of the new content already held), so a very large text overwrite allocates both texts for a UI-only diff. A future refinement can bound the pre-read and fall back to a whole-file / no contextual diff above a size threshold; tracked as `TODO(overwrite-diff-bound)` at the read site.
+
+## Related
+
+- Completes the one remaining representation difference named as a non-goal in [Tagged render-intent union](2026-07-02-tool-render-intent-union.md) — that Agent Note's Non-goals section is updated to record that applied-hunk diffs shipped here.
+- Builds on the [filesystem capability seam](2026-06-17-filesystem-capability-seam.md) (the before/after are storage facts the backend returns) and [event-sourced sessions](2026-06-11-event-sourced-sessions.md) (the `meta` payload persists on the `tool/result` event, so replay reproduces the card).
+- The `meta` channel is deliberately generic: a future tool (a structured search, a data-table result) can attach its own durable result presentation without another core change.

.agents/notes/implemented/architecture/2026-07-02-tool-render-intent-union.md

@@ -0,0 +1,80 @@
+# Agent Note: Tagged render-intent union for tool-call presentation
+
+Status: implemented
+
+## Problem
+
+A tool declares how its calls render in a UI (an editor's tool-call card) through two callbacks, `presentCall`/`presentResult` on `ToolDefinition`, returning `ToolCallPresentation` / `ToolResultPresentation` with an optional `ToolTerminal` sub-shape. These grew incrementally into a **bag of optional fields**: `title`, `kind`, `rawInput`, `content`, `locations`, `terminal` on the call; `title`, `content`, `terminal` on the result; `cwd`/`output`/`exitCode`/`signal` on `ToolTerminal`. The split of responsibility is muddy:
+
+- The call-side and result-side `terminal` fields overlap, and the bridge reconciles a `content` block AND a `terminal` block AND `rawInput` per call, stitching them together with ad-hoc conditionals.
+- Which combinations are *valid* is unwritten: a `terminal` call that also sets `content` means "description above the card"; a generic call that sets `terminal` is meaningless but representable. The type permits nonsense.
+- There is no way to express the one file-tool affordance an editor most wants — a **diff card** (`{path, oldText, newText}`, which Zed renders as an inline diff / new-file preview). `ToolCallPresentation.content` is the *LLM* `ContentBlock[]` vocabulary (text/image), so a tool literally cannot ask for a diff.
+
+The existing `FIXME(tool-presentation)` in `packages/core/tools/src/index.ts` named the fix: "redesign the type so a tool declares its render INTENT once (e.g. a tagged union over card kinds) rather than a bag of optional fields the bridge stitches together." The rejected Agent Note [Collapse tool-owned UI presentation](../../rejected/simplification/2026-06-20-generic-tool-rendering.md) deferred it explicitly: rich rendering "should return later as a tagged render-intent union after there are at least two real tools and two real consumers to validate the vocabulary." That bar is now met — two producer families (`dsh-tool-bash`, `dsh-tool-fs`) and two consumers (the ACP bridge live path + the snapshot replay path).
+
+## Decision
+
+Replace the optional-field bag with a **`card`-tagged discriminated union**. A tool declares one render intent per call/result; the bridge switches on the tag.
+
+```ts ignore-check
+type FileLocation = { path: string; line?: number }
+type FileDiff = { path: string; oldText: string | null; newText: string } // oldText null ⇒ new file
+
+// presentCall → ToolCallView
+type ToolCallView = GenericCallView | TerminalCallView | DiffCallView
+interface GenericCallView { card: 'generic'; title: string; kind?: ToolCallKind; rawInput?: unknown; content?: ContentBlock[]; locations?: FileLocation[] }
+interface TerminalCallView { card: 'terminal'; title: string; description?: string; cwd?: string }
+interface DiffCallView { card: 'diff'; title: string; diffs: FileDiff[]; locations?: FileLocation[] }
+
+// presentResult → ToolResultView
+type ToolResultView = GenericResultView | TerminalResultView
+interface GenericResultView { card: 'generic'; title?: string; content?: ContentBlock[] }
+interface TerminalResultView { card: 'terminal'; title?: string; output?: string; exitCode?: number; signal?: string }
+```
+
+`card` is **required** on every variant — a real discriminant, not an optional default. The bridge does `switch (view.card) { case 'generic': … case 'terminal': … case 'diff': … default: assertNever(view) }`. The union is **closed** (per the [switch-exhaustiveness convention](../../../../AGENTS.md)): a fourth render intent (a table, a chart) needs new bridge code to render it anyway, so a plugin-added variant that the bridge silently drops would be worse than a compile error. Adding a variant breaks compilation at the bridge switch — exactly the signal we want.
+
+### Why a tagged union beats the field-bag
+
+- **Invalid states become unrepresentable.** A generic card cannot carry terminal output; a terminal card cannot carry a diff. The old bag permitted all of these.
+- **The bridge switches instead of stitching.** One arm per card kind, each producing exactly the wire shape that card needs, rather than reconciling five optional fields whose interactions are undocumented.
+- **`diff` is a first-class intent.** `dsh-tool-fs` write/edit declare `card:'diff'`; the bridge emits an ACP `{type:'diff', path, oldText, newText}` `ToolCallContent` (already in the SDK's `ToolCallContent` union, previously unused by the bridge). This is the affordance the redesign unlocks.
+
+### Producer mapping
+
+- `dsh-tool-fs` read → `generic` (`kind:'read'`, a follow-along `location`); write → `diff` (`oldText:null`); edit → `diff` (`oldText:old_string || null`, `newText:new_string ?? ''`). This mirrors `claude-agent-acp`'s `toolInfoFromToolUse` Read/Write/Edit arms field-for-field.
+- `dsh-tool-bash` foreground → `terminal` call + `terminal` result; `run_in_background` → `generic`. The generic `task_*` controls own their own generic cards.
+- `dsh-tool-todo` → `generic`.
+
+### Terminal fallback ownership
+
+`TerminalResultView` carries only `output`/`exitCode`/`signal`. A UI without the terminal capability needs a fenced ` ```console ` text fallback; that derivation moves to the **bridge** (it wraps `output` in a fenced block on the no-capability path), rather than the tool double-encoding it. This keeps the bash tool's result a single structured shape and preserves the existing capability-gated behavior byte-for-byte.
+
+### Purity preserved
+
+`presentCall`/`presentResult` remain pure functions of `args` (+ the result for `presentResult`) — they run on live streaming AND session-log replay, so they must be replay-deterministic. Every view is derived from args alone: write's diff is new-file style (`oldText:null`) because the tool has no old content at call time; edit's diff is `old_string`→`new_string`.
+
+## Relative-path display titles
+
+`claude-agent-acp` relativizes a file card's title path against the session cwd (`toDisplayPath`) — `Read src/foo.ts`, not `/abs/proj/src/foo.ts` — while keeping `locations[]`/`diff.path` **raw** (the editor opens the real path). Our `presentCall` is pure/args-only and cannot see the session cwd, so this relativization happens at the **bridge**, which already threads the session cwd into tool-call rendering (the same cwd it uses to resolve a terminal card's header). The bridge relativizes the title only, by an exact structured replace of the known `locations[0].path`/`diffs[0].path` substring — generic over the file-card kinds, never special-casing tool names.
+
+## Alternatives considered
+
+- **Delete tool-owned presentation entirely** — [the rejected collapse proposal](../../rejected/simplification/2026-06-20-generic-tool-rendering.md); its own verdict deferred to exactly this union once two real tools and two real consumers existed, and that bar is now met.
+- **A merge-extensible union** (the `ContentBlockMap` pattern) — rejected: a new render intent needs new bridge code to render it anyway, so a plugin-added variant the bridge silently drops would be worse than the compile error the closed union raises at the bridge's `assertNever` switch.
+- **Keeping the optional-field bag** — the status quo the Problem dissects: invalid states representable, undocumented field interactions, and no way to ask for a diff card at all.
+
+## Consequences
+
+A new render intent is a compile-breaking change at the bridge switch — deliberately: rendering code must exist before a card kind does. Invalid card/field combinations are now unrepresentable, and the bash fallback derivation lives in the bridge, so a tool returns one structured shape. The bar for a fourth card (a table, a chart) is writing its bridge arm in the same change.
+
+## Non-goals
+
+- **Live incremental `terminal_output_delta` streaming** and **command classification** — the terminal-rendering Agent Note's own deferred follow-ups, untouched here.
+
+## Related
+
+- Supersedes the deferral in [Collapse tool-owned UI presentation](../../rejected/simplification/2026-06-20-generic-tool-rendering.md) (rejected — "wait for two real tools and two real consumers, then a tagged render-intent union"). That bar is now met; this is that union.
+- Extended by [Result-time applied-hunk diffs](2026-07-02-result-time-applied-hunk-diffs.md), which adds a persisted `meta` channel so write/edit emit a result-time `DiffResultView` — the applied change (a contextual hunk with context lines / one per `replace_all` site, or a whole-file diff for a create) — on top of this union's call-time diff card.
+- Folds `ToolTerminal` into the `terminal` views described by [ACP terminal and tool-call rendering](../feature/2026-06-18-acp-terminal-and-tool-rendering.md) (the `_meta` terminal-card convention and capability gate are unchanged; only the harness-side presentation type changes).
+- The ACP SDK's `Diff` / `ToolCallContent` types back the new `diff` card.

.agents/notes/implemented/architecture/2026-07-03-filesystem-directory-listing-seam.md

@@ -0,0 +1,51 @@
+# Agent Note: Add direct directory listing to the filesystem seam
+
+Status: implemented
+
+## Problem
+
+`@deepseek-ai/dsh-fs` is the provider seam for filesystem access, with local and future non-local backends behind the same `ctx.fs` contract. Before this change it could resolve paths, stat targets, read text, stream text, write text, and edit text. That was enough for model-facing file tools, but not for non-model-facing consumers that need to enumerate directories without importing `node:fs`.
+
+The immediate pressure came from skill loading: reading an individual `SKILL.md` can already go through `ctx.get('fs')`, but discovering which skill roots contain `<name>/SKILL.md` or `<name>.md` still needs directory enumeration. Adding directory listing only in `dsh-skill` would either keep a direct Node dependency there or invent a one-off local helper outside the filesystem provider stack.
+
+This decision adds the provider capability without a model-facing `ls`/`list` tool or skill-discovery change. Those consumers require separate UX, prompt, and policy decisions.
+
+## Decision
+
+Add `FileSystem.listDir(target, signal?)` to `@deepseek-ai/dsh-fs`.
+
+`listDir` lists one directory level only. It returns direct children in stable name order and includes:
+
+- `name`: the child basename.
+- `type`: `file`, `directory`, or `other`.
+- `target`: the resolved child `FsTarget`.
+- `version`: cheap metadata when available.
+- `size`: regular-file size when available.
+
+It never reads file contents. Recursive traversal, globbing, pagination, search, file watching, and model-facing rendering are intentionally out of scope.
+
+The local backend implements this through `readdir({ withFileTypes: true })`, `resolveLocalTarget`, and metadata `stat`/`realpath` probes. The result order is deterministic (`name.localeCompare`) to keep prompt/listing output stable for future consumers and improve prefix-cache reuse.
+
+Broken or disappeared children may be represented as `type: 'other'` without `version`/`size`; they do not abort the whole listing. Permission or backend I/O failures while listing the directory or resolving/probing child metadata fail the whole listing with structured `FsError` codes:
+
+- `FS_NOT_FOUND` for missing targets.
+- `FS_NOT_DIRECTORY` for existing non-directory targets.
+- `FS_PERMISSION_DENIED` for permission failures.
+- `FS_IO_ERROR` for other backend I/O failures.
+- `FS_ABORTED` for aborted calls.
+
+## Alternatives considered
+
+**Add a model-facing list tool with the seam.** Rejected because its prompt, schema, and rendering contracts are independent of the provider primitive.
+
+**Keep directory enumeration in each consumer.** Rejected. That would bind product packages such as `dsh-skill` to Node/local filesystem behavior and bypass policy/remote/sandboxed backends.
+
+**Make `listDir` recursive or glob-shaped.** Rejected for now. Skill-root discovery only needs direct children, and a simple direct listing is the smallest backend contract future consumers can safely compose.
+
+**Skip children that fail metadata resolution.** Rejected. The API promises resolved child targets, so permission/IO failures while resolving a child are contract failures. Broken or disappeared children are the exception because they can still be represented without claiming a live resolved file.
+
+## Consequences
+
+Every filesystem backend must now implement one additional provider primitive. That is deliberate foundation work while the harness is still unreleased, but it does mean future sandboxed/remote backends need to define equivalent direct-child listing behavior.
+
+The capability remains provider-facing. Until a consumer lands, ACP/model sessions will still need existing tools such as `bash` for directory listing. The absence of a model-facing `listdir` tool is expected, not a wiring failure.

.agents/notes/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md

@@ -0,0 +1,70 @@
+# Agent Note: Prompt variables and tool-guidance ownership
+
+Status: implemented
+
+## Problem
+
+The assembled system prompt had four defects, all of one family: facts the harness already knows were restated by hand somewhere else, and drifted.
+
+**The model could not know its own name.** `AgentOptions.model` drives every request, but no prompt text carried it — and nothing COULD carry it: sections in `dsh-system-prompt` were context-global while the model name is per-agent, and `assemble()` took no per-agent input at all.
+
+**Tool guidance was hand-written prose in leaf YAML.** The bash/subagent/todo_write usage guidance lived in the `systemPrompt` strings of `examples/repl-agent/cordis.yml` and `examples/acp-agent/cordis.yml` — two drifting copies (the ACP one was already abridged) — while `dsh-tool-fs` and `dsh-tool-web` owned their guidance as `ctx.systemPrompt.section()` contributions. Loading or dropping a tool plugin meant editing every deployment's persona by hand; both YAMLs carried a `FIXME(config-comments)` apologizing for a symptom of the split, and the stdio welcome banner hand-enumerated the tool set too.
+
+**The persona rendered after tool guidance.** The loop string-joined `agent.options.systemPrompt` AFTER the assembled sections, so the model read "Use the read tool…" before "You are a coding agent" — backwards relative to the identity-first convention (Claude Code, Codex) and a second composition path besides the section pipeline.
+
+**The fork tool's description was false.** `dsh-tool-subagent` hardcoded one description written for spawn semantics — "a separate agent that works in its own context … it does not see this conversation" — and the `subagent_fork` instance (whose child inherits the parent's completed turns) got the same words; the YAML prose corrected the lie out-of-band. Minor kin: `PromptSection.name` was documented "(diagnostics / dedup)" but duplicates were silently accepted.
+
+## Decision
+
+**One principle: every fact in the prompt has exactly one owner.** The model name and workspace are config/session facts → the harness exposes them as variables and the persona references them. Per-tool semantics and when-to-use → the tool's `description`. Cross-call habits a description cannot carry → the tool package's prompt section. Harness provenance → the static `harness:identity` section. Deployment role and behavior → the deployment's persona.
+
+### Assemble context
+
+`SystemPrompt.assemble(context)` takes a merge-extensible `AssembleContext`. `dsh-system-prompt` declares the optional `scope` selector used for scoped routing, while `dsh-agent` declaration-merges the optional typed `agent` field onto it (a type-level edge `agent → system-prompt`, with no runtime dependency cycle). The loop calls `assembleContextFor(agent)` each step so both fields identify the same agent; section text providers may read that context, and the `system-prompt/assemble` waterfall receives it so a listener can filter or extend per agent.
+
+### Prompt variables
+
+Plugins register `{{name}}` values through `ctx.systemPrompt.variable(name, provider)`. Assembly resolves them into the waterfall-visible variable map. Rendering rejects unknown own-property references, registered providers that return `undefined`, malformed complete references, and unbalanced references that still contain a closing `}}`; a lone unmatched `{{` remains prose, and substituted values are not rescanned. Registration rejects invalid or duplicate variable names, and section names are unique.
+
+`dsh-agent-loop` registers the two built-ins, both pure projections of the context agent: `model` (= `options.model`) and `cwd` (= `session.header.cwd`). The example personas write `powered by the {{model}} model` — the model name is stated once, in the `model:` config key. `{{cwd}}` is demonstrated in the ACP example only: every ACP session carries the client's cwd, while config-pre-created stdio agents have none (a persona claiming `{{cwd}}` there fails the turn — by design). The variables stay on the loop plugin (unlike the sections below): they are runtime facts of the agents THIS loop drives, and a replacement loop supplies its own.
+
+### Persona as the order-0 section
+
+`dsh-system-prompt` owns `harness:identity` at order `-100` and the configured `deployment:persona` at order 0, so both survive a replacement loop. Prompt rendering has one path, `renderPrompt(assembly)`, and the routed request header therefore records the exact prompt later replayed by `ctx.tokenMeter` for compaction pressure. An agent-scoped `deployment:persona` shadows the global default and lets subagent providers install a persona before publication. The conventional order bands are identity `-100`, persona `0`, and tool guidance `100–199`.
+
+### Tool guidance ownership
+
+Per-tool semantics and selection guidance live in tool descriptions. Prompt sections carry only cross-call habits, such as checking bash exit markers or preferring filesystem tools over shell commands. `todo_write` and subagent tools need no section because their descriptions contain the full contract. Deployment personas contain only role and behavior.
+
+### The subagent conversation-history descriptor
+
+`SubagentProvider.inheritsParentContext` describes conversation seeding, not scope, services, tools, or authority. Spawn and ACP set it to `false`; fork sets it to `true`. `dsh-tool-subagent` derives its tool and prompt-parameter descriptions from the flag, including that fork inherits completed turns but not the in-flight turn. Provider lifecycle events keep that wording synchronized with reactive provider registration; their rationale lives in the [provider-lifecycle-events Agent Note](2026-07-05-subagent-provider-lifecycle-events.md).
+
+## Alternatives considered
+
+- **The loop composes an identity line itself** — hardcodes model-facing prose in the one package that must stay thin ("plugins, not loop changes"), and outside the section pipeline it would be a second composition path. (The identity DOES ship as a code literal — but as an ordinary section registered by `dsh-system-prompt`, whose `system-prompt/assemble` waterfall remains the escape valve for a deployment that must drop it.)
+- **Inject the model name via the `agent/request` waterfall** — prompt text would be composed in two places and the earlier rendered persona could disagree with the final routed header. The request plugin that owns late routing must also own any earlier prompt claim about that model.
+- **Hand-write the model name in each persona** — duplicates the `model:` key one line above and silently lies after a config edit; the exact disease this Agent Note cures.
+- **Lenient interpolation (leave unknown refs verbatim, or substitute empty)** — a typo ships `{{modle}}` (or a hole) to the model and nobody notices until transcript review.
+- **Per-instance subagent wording in config** — returns model-facing prose to every deployment × instance, the P2 disease again. **Keying wording off the provider NAME** — `providerName` is itself config, so a renamed provider silently gets the wrong words.
+- **Resolving the provider at `apply` time (a load-order requirement)** and **section-only subagent wording (lazily resolved at assemble)** — the alternatives to the provider-lifecycle events; both rejected in [the provider-lifecycle-events Agent Note](2026-07-05-subagent-provider-lifecycle-events.md).
+
+## Out of scope
+
+- Further variables (`date`, platform, git state) — the registry makes each a one-line contribution by whichever plugin owns the fact; none is claimed here.
+- A config `cwd` for pre-created stdio agents (would let the stdio persona use `{{cwd}}` and partition persistence by real path) — deferred until the session-cwd story is revisited.
+
+## Shipped invariants
+
+- The repl-agent prompt renders identity, persona with the interpolated model, then fs/bash/web guidance through one assembly path.
+- Fork and fresh subagent descriptions reflect whether the provider inherits completed conversation turns; the tool appears, disappears, and is reworded with provider lifecycle changes.
+- Unknown, valueless, malformed, or unbalanced variable references name the section and throw; duplicate section, variable, and tool registrations also throw.
+- Snapshot replay is prompt-independent: it keys recorded chunk streams by turn and step without comparing the outgoing request.
+
+## Consequences
+
+- Every fact in the assembled prompt now has exactly one owner, and the hand-maintained tool prose in leaf YAML is gone: loading or dropping a tool plugin no longer means editing any deployment's persona.
+- `{{model}}` reflects `AgentOptions.model` at assembly time. A plugin that switches models in the `agent/request` waterfall makes the prompt's claim stale for that step, and one that SUPPLIES the model there (options.model unset — the loop's documented fallback) leaves the variable valueless at render, failing a `{{model}}` persona before the waterfall runs. Both have the same remedy, and it is the ownership rule itself: the plugin that owns the late-bound model fact states it early on the `system-prompt/assemble` waterfall (`assembly.variables['model'] = …`) — one owner, both statements; a loop test pins the supply path end-to-end. Accepted.
+- While a bound provider is absent (not yet activated, unloaded, mid-HMR-reload), the subagent tool does not exist and a model request in that window simply lacks it. That is the honest state — the alternative was a registered tool whose description or execution could not be trusted.
+- Strictness means a persona can fail a turn at render (e.g. `{{cwd}}` on a cwd-less session). The failure is contained — the turn ends `error`, the loop survives — and it is an authoring error we WANT loud.
+- No escape syntax for a literal `{{name}}` in prompt prose yet; add one if a real prompt ever needs it.

.agents/notes/implemented/architecture/2026-07-05-reconstructable-requests.md

@@ -0,0 +1,54 @@
+# Agent Note: Every LLM request is reconstructable from the session log
+
+Status: implemented
+
+## Problem
+
+The request pipeline did not guarantee prefix stability for provider caching, and the session log could not reconstruct what the model saw. It omitted model, system prompt, and tool schemas while allowing per-call request rewrites. Cache behavior and replay equivalence therefore depended on whichever plugins happened to be loaded.
+
+The reference shape for the happy path is MiniCode's `LLMClient`: a stateful conversation client, appended to — never rebuilt — as the conversation advances, resetting only when the system prompt, tool set, or compaction genuinely changes what the model must see. The design question this Agent Note answers is how to get that discipline without giving up event-sourcing.
+
+## Decision
+
+### The principle
+
+**Model-visible ⟺ logged.** Anything that reaches a model request must be recorded in the session log. The checkable consequence: **every conversation request the loop sends is a pure function of the session log** — anyone holding the log reconstructs it byte-for-byte. Scope, stated precisely: the guarantee covers the loop-built `GenerateOptions`; provider wire bytes follow from it because both adapters' serialization is a pure per-message function at a pinned code version; direct one-shots (compaction's summarize call) log their envelope scalars (`compact/summary.{provider, model, maxTokens}`) and their input is deterministic code over the logged region — reconstructable from log + code, outside the invariant by the unfrozen-request marker.
+
+Prefix-cache stability is corollary #1, not the headline: an append-only log projected by a per-node pure function yields requests that are append-extensions of their predecessors whenever the header is unchanged — stability is emergent, not managed. Byte-exact audit/replay is corollary #2; resume and fork with *attributable* drift is corollary #3.
+
+### The mechanism
+
+**Messages.** `Session.deriveMessages()` is cached: each surface entry is projected exactly once, when first seen, through the public per-event function `deriveEventMessage(event)`; a surface rewrite (a compaction `replace` — `SurfaceManager.replaceGeneration`) rebuilds. Callers get a fresh array per call over shared, deep-frozen messages: mutating logged history through a projection is unrepresentable (it throws), replacing the old clone-per-call isolation. External reconstructors fold the same public function over a log prefix, so no two paths can disagree.
+
+`EpochHeader` records the request's non-history state: call config, rendered system prompt, tool schemas, and session prefix, with empty values canonicalized to absence. `request/header` always writes a full snapshot: the first loop instance uses reason `initial`, later instances use `resume`, and an in-instance change uses `change`. `foldRequestHeader` selects the latest snapshot. Legacy `request/header-delta` events and the removed `fallback` reason are rejected when appended or loaded.
+
+Each step rebuilds prompt assembly. On the instance's first step, `agent/session-prefix` extends a frozen empty seed with request-only opener messages; the result is frozen and cached for that loop instance before the generic `agent/pre-step` checkpoint and boundary snapshot. The first call config starts from explicit `AgentOptions`, preserving fork overrides and resume reconfiguration; later calls start from the folded header. `agent/request` may replace only that frozen config seed, while model-visible content enters through logged channels. The loop records the owed header event—the prefix's only durable home—builds `GenerateOptions` from prefix, snapshot, and header, and deep-freezes it while leaving `AbortSignal` live. Per-instance state is only the cached prefix and whether its anchoring snapshot has been written.
+
+**`step/start` is the reconstruction boundary.** A step derives messages from events before that sequence. Injection after the snapshot joins the next request, and reentrant appends are rejected during event publication. `agent/pre-step(agent, turn, step, signal)` remains the generic seam for content needed by the current request. Header reconstruction selects the step's `request/header`, or carries the prior snapshot when no new header is written.
+
+**Enforcement.** In development, `dsh-invariants` independently rebuilds each loop request through a fresh `Session`, so the live cache cannot vouch for itself, then compares messages and folded header fields at `llm/stream`. Loop requests are identified by their frozen shape and session id; direct one-shots are excluded. Correctness depends on sequence-bounded reconstruction rather than listener order. A with-key e2e requires positive cache-read tokens after the first request; per-step usage is the production signal, and a header change or compaction appears as a cache-read drop on the next step.
+
+### The MiniCode shape: adopted, with the provenance arrow inverted
+
+Like MiniCode, the conversation advances append-only and resets only when model-visible state changes. Unlike MiniCode, the event log remains the source of truth because it also owns persistence, recovery, boundaries, tool pairing, and provenance. `Session` caches message and header folds derived from that log, making every request independently checkable.
+
+## Alternatives considered
+
+- **Client as source of truth** (literal MiniCode): a second operative truth beside the log — the two drift and nothing notices; see the section above.
+- **A stateful transmission client mirroring the log** — duplicates conversation state, needs rollback around listeners, leaves an unlogged edit surface, and still cannot reconstruct request headers. Session-owned caches plus logged headers avoid those split truths.
+- **Per-call request scalars** (a freely mutable config handed to each `agent/request` dispatch): a listener flips the model per call with zero accounting, silently abandoning the provider cache this design exists to protect. Config is per-conversation logged state; the waterfall proposes, the log records.
+- **Detect-and-report** (compare consecutive requests, warn on divergence): catches violations after the fact; a violating request is still constructible and ships. Rejected for interface-level unrepresentability.
+- **Event-driven assembly** (re-render only on change signals): a missed-signal bug class — a tool registered mid-session emits `tools/change`, not `system-prompt/change`, and a third-party provider may emit nothing. Per-step render + value compare is robust with zero signal discipline.
+- **A custom header-delta codec** (system line edits, name-keyed tool edits, whole config/prefix replacements): reduced repeated bytes but duplicated the representation and its diff/apply/fallback machinery. Full snapshots retain one replay representation.
+- **Narrative changed-field lists on header snapshots**: derivable by comparing consecutive snapshots. The `reason` remains because an instance boundary is not derivable from the snapshot values.
+
+## Consequences
+
+- A request that is not explained by the log cannot be constructed by accident — not by the loop, not by a listener; mutating a built request throws; every header change is a durable, diffable log event.
+- Choosing between the advisory channels is a change-frequency decision, and the design makes the stable one structural: an `agent/session-prefix` contribution is composed once per loop instance and reused verbatim, so it extends the cacheable prefix at zero marginal cost and CANNOT bust the provider cache mid-session; content that changes mid-session flows through the append-only history channels — `agent.inject()` and tool/prompt-submit `additionalContexts` — each a durable `context/message` paid once and prefix-cached thereafter, at the price of accumulating in history and the log. Route session-frozen openers to the prefix and change notices to the history channels; a per-step request-only tail slot was deliberately dropped (no consumer, and a durable append covers every current update pattern).
+- What still costs full price at the provider is inherent and logged: compaction (its `compact/*` events and replacement entry), a real prompt, tool, or config change (`request/header` with reason `change`), or a process boundary with drift (a differing `resume` snapshot). The provider's own reasoning-content exclusion is managed server-side.
+- The `step/start`-listener behavior change (above) is the one observable semantics change for plugins; `agent/pre-step` is the current-request seam.
+- Tool-result trimming (planned) needs no new mechanism: a logged single-entry surface replace (`start === end`) carrying a trimmed `tool/result` under the same `callId` — compaction-family, replay-correct, cache-bust batched by the same pressure logic.
+- Session logs grow one `request/header` snapshot per loop instance plus snapshots on real changes. This is larger than a delta codec but small beside chunk-heavy logs and retains one replay representation. `SESSION_FORMAT_VERSION` stays `0`; legacy delta events are rejected rather than migrated.
+- Snapshot expected outputs changed once (every transcript gains its header events); the fs-writing fixtures are stored in the normalized authored form with cwd-relative tool arguments, because replay only round-trips cwd-independent argument paths.
+- FIXME(call-config-shape): revisit `LlmCallConfig`'s exact field set — which fields are genuinely epoch-level for cache purposes (`model` certainly; the sampling scalars sit there out of caution), and where provider-specific extras (reasoning options, extra body params) belong when an adapter needs them.

.agents/notes/implemented/architecture/2026-07-05-subagent-provider-lifecycle-events.md

@@ -0,0 +1,34 @@
+# Agent Note: Subagent provider-lifecycle events — `subagent/provider-added` / `subagent/provider-removed`
+
+Status: implemented
+
+## Problem
+
+[The prompt-variables Agent Note](2026-07-05-prompt-variables-and-tool-guidance-ownership.md) makes `dsh-tool-subagent` DERIVE its model-facing wording from its provider: `SubagentProvider.inheritsParentContext` (spawn/ACP `false`, fork `true`) drives both the tool description and the `prompt` parameter description, so the fork tool stops lying about context inheritance. That fix created a cross-fiber data dependency: a tool's description is fixed at TOOL REGISTRATION (deliberately — the description is where tool-choice guidance lives), but the provider arrives on its own plugin fiber, on no particular schedule.
+
+Resolving the provider at the tool plugin's `apply` time creates an implicit load-order requirement ("list the backend before the tool in cordis.yml"). That requirement fails because the Cordis Loader starts sibling entries concurrently and `Entry.init()` does not await activation: a delayed backend can leave the tool fiber failed even when listed first. The Loader offers no sibling-order guarantee — "async state is not synchronous state" ([defensive patterns](../../../../docs/defensive-patterns.md)).
+
+## Decision
+
+The registry announces provider membership as typed events, and the consumer mirrors them instead of assuming order:
+
+- **`subagent/provider-added(provider)`** — a provider became resolvable in the `ctx.subagents` registry. Emitted on registration.
+- **`subagent/provider-removed(name)`** — a provider left the registry (its plugin's fiber was disposed — an unload or an HMR reload). Emitted from the registration's disposer.
+
+`dsh-tool-subagent` mirrors its named provider's lifecycle: it registers the tool when the provider is (or becomes) available — deriving the wording from that provider at that moment — unregisters the tool when the provider goes away, and re-derives on re-registration (HMR reload). While the provider is absent the tool does not exist, which cannot lie to the model. There is deliberately NO load-order requirement left to document: the events make the ordering question disappear instead of pinning it.
+
+The events also complete the seam's vocabulary: `ctx.subagents` is a named registry on which multiple delegation backends coexist (`spawn`, `fork`, `acp`), and a registry whose contents other plugins derive state from should announce membership changes as typed events rather than requiring polling or load-order faith.
+
+## Alternatives considered
+
+- **Resolve the provider at `apply` time and throw when absent** — rejected because "list backends first" would claim a Loader ordering guarantee that does not exist.
+- **Retrying the lookup (poll until the provider appears)** — converges eventually but invents a private readiness protocol beside the one the framework already has (effect registration + disposal); it also cannot notice a provider LEAVING, so HMR would strand a tool whose wording describes a disposed backend.
+- **Section-only subagent wording, lazily resolved at assemble time** — tolerates any load order too, but moves tool-choice guidance out of the DESCRIPTION, contradicting the ownership rule the prompt-variables Agent Note establishes (per-tool semantics and when-to-use belong in the description). Reactive registration keeps the description authoritative AND order-free.
+- **Keying wording off the provider NAME instead of the provider object** — `providerName` is itself config, so a renamed provider silently gets the wrong words; deriving from the resolved provider's own `inheritsParentContext` cannot drift.
+
+## Consequences
+
+- Consumers deriving state from a named provider react to `subagent/provider-added`/`-removed` instead of reading the registry at `apply` time; `dsh-tool-subagent` is the reference implementation.
+- **Addition fails loud; removal is contained per listener.** An addition listener may unwind registration. Removal runs during disposal, so one throwing listener is logged without starving later mirrors or disrupting teardown. `start()` still resolves the provider by name for every run, preventing stale tools from calling a removed backend. See the [events catalog](../../../../docs/cordis-catalog/events.md) and [producer/consumer map](../../../../docs/event-producer-consumer.md).
+- **A window where the tool is absent.** Between backend disposal and re-registration (an HMR reload), the model sees no subagent tool. This is the honest state — the alternative is a tool that dispatches into nothing — and the tool registry's `tools/change` emit keeps prompt assembly current.
+- **Two waiting fibers sharing a `toolName` is an invalid config caught late.** If two loads of `dsh-tool-subagent` name different providers but the same `toolName`, both wait, and whichever provider arrives first registers; the second registration throws only when ITS provider arrives. `TODO(subagent-dup-toolname)` in the plugin records this blast radius; the tool registry's duplicate-name rejection remains the backstop.

.agents/notes/implemented/architecture/2026-07-06-timeout-deadline-library.md

@@ -0,0 +1,96 @@
+# Agent Note: A shared timeout/deadline primitive, with hard-kill left to each capability
+
+Status: implemented
+
+## Problem
+
+Timeout handling was drifting apart across the tool-bearing capabilities, and the divergence was not superficial — it was the same logic re-implemented three ways, each with its own subtle correctness burden.
+
+- **bash** ([packages/bash/bash-local/src/run.ts](../../../../packages/bash/bash-local/src/run.ts)) had a full, correct timeout inside the process plumbing: a config-clamped `timeoutMs`, two independent triggers — a `killTimer` for the timeout and an `onAbort` listener for upstream cancellation — each calling one `kill()` closure that escalates SIGTERM→grace→SIGKILL on the process group, and two orthogonal outcome booleans (`timedOut`, `aborted`) latched independently.
+- **web_fetch** ([packages/web/web-fetch-local/src/provider.ts](../../../../packages/web/web-fetch-local/src/provider.ts)) had a correct but *hand-rolled* timeout: it constructed an `AbortController`, wired `setTimeout(() => controller.abort(new WebError(…, 'WEB_FETCH_TIMEOUT')))`, manually added and removed the upstream-signal listener, cleared the timer in a `finally`, and recovered the timeout reason from `signal.reason` in a `translateAbortOrNetwork` helper because the reader surfaces a bare `AbortError`.
+- **web_search** ([packages/web/tool-web/src/search.ts](../../../../packages/web/tool-web/src/search.ts)) had **no timeout at all**: `WebSearchRequest` ([packages/web/web/src/types.ts](../../../../packages/web/web/src/types.ts)) carries no `timeoutMs` field, and each provider's `search()` only forwards `exec.signal`. (web_search stays untimed here — see Consequences.)
+
+Each new external-process or network tool re-derived the same four things — clamp the requested value, start a timer, fuse the timeout with upstream cancellation, and distinguish "timed out" from "cancelled" on the way out — and the fusion and reason-recovery are exactly the parts that are easy to get subtly wrong (web_fetch's `signal.reason` dance is evidence). At the same time, the *termination* each performs is irreducibly different: bash kills an OS process group (work runs in a child process, outside this runtime, reachable only by signal), while web aborts an in-process `fetch` (undici tears down the socket). There is no single mechanism that can stop all of them.
+
+## Decision
+
+`@deepseek-ai/dsh-timeout` lives under `packages/util/` (peer to `dsh-brand`) and owns the *timing and classification* half of timeout; the *termination* half — the hard kill — stays in each capability's implementation. It is a library of pure functions, **not** a cordis service or plugin: it takes no `ctx`, registers nothing, holds no cross-call state, and emits no events. There is deliberately no central "timeout service" that would have to know how to stop every capability's work — that knowledge is exactly what a microkernel keeps out of shared layers, and what Codex's exec-only `ExecExpiration` scope demonstrates.
+
+### The library surface
+
+Three functions plus one reason type:
+
+```ts ignore-check
+/** The internal reason attached to a timeout abort, so consumers can classify it after the fact. */
+export class TimeoutReason extends Error {
+  override name = 'TimeoutReason'
+
+  constructor(readonly code: string, readonly timeoutMs: number) {
+    super(`${code} after ${timeoutMs}ms`)
+  }
+}
+
+/** Validate/fill a caller's optional positive hint from the backend's default, then cap at its max. */
+export function clampTimeout(
+  requested: number | undefined,
+  def: number,
+  max: number,
+  name = 'timeoutMs',
+): number
+
+/**
+ * Build a deadline signal that aborts on upstream cancellation OR on timeout,
+ * with the timeout carrying a `TimeoutReason`. `timeoutMs <= 0` means "no
+ * timeout" (background tasks): forward only the upstream signal, arm no timer.
+ * The returned object's `[Symbol.dispose]` clears the timer — `using` for a
+ * scope-lifetime consumer, a manual call for an event-lifetime one.
+ */
+export function deadline(
+  upstream: AbortSignal | undefined,
+  timeoutMs: number,
+  code: string,
+): { signal: AbortSignal; [Symbol.dispose](): void }
+
+/** Recover the TimeoutReason from an aborted signal (or error); `code` scopes the match to this deadline's timer. */
+export function timeoutOf(x: AbortSignal | { reason?: unknown }, code?: string): TimeoutReason | undefined
+```
+
+`deadline` fuses an upstream signal with a timer through `AbortSignal.any`, adds a typed `TimeoutReason`, and exposes disposable timer cleanup. Non-positive timeouts are an internal no-timeout sentinel for backend-owned background work; external hints pass through `clampTimeout` and must be positive and finite. Without a timer or upstream signal, the function returns a never-aborting signal with the same disposal shape. Providers translate timeout reasons into seam-specific results. `timeoutOf(signal, code)` scopes classification so an outer nested deadline is treated as upstream cancellation rather than the inner capability's timeout.
+
+### The division of labor
+
+| Concern | Owner |
+|---|---|
+| Validate request hint and clamp default/max | `dsh-timeout` (`clampTimeout`) — pure arithmetic plus the shared positive-finite request contract |
+| Arm timer, abort on deadline, carry reason, fuse with upstream cancel | `dsh-timeout` (`deadline`) |
+| Clear the timer | `dsh-timeout` (`[Symbol.dispose]`) |
+| Classify the first abort reason after abort | `dsh-timeout` (`timeoutOf`) |
+| **Actually terminate the work** | the capability's implementation |
+| The default/max *values* | the capability's config |
+| The timeout `code` string | the capability (`WEB_FETCH_TIMEOUT` ≠ `BASH_TIMEOUT`) |
+
+The signal only *notifies*; termination is always the listener's job, and the listener differs by capability. bash writes its own `addEventListener('abort', kill)` because the OS process lives outside this runtime and nothing else will kill it; web hands `d.signal` to `fetch` and undici tears down the socket. This is why file read/write/edit take **no** `timeoutMs`: a local syscall is best-effort-abortable at most, a timeout could not force `fsync`/`rename` to stop, and adding one would be an implicit default that violates explicit-over-implicit. Both reference agents leave file I/O untimed for the same reason.
+
+### How each capability consumes it
+
+- **web_fetch** — the tool stays validate-and-forward; the provider's hand-rolled controller + `setTimeout` + manual listener + `finally` + `signal.reason` recovery is replaced by provider-owned `deadline`/`timeoutOf`. A pre-aborted upstream signal still throws `WEB_ABORTED` up front; otherwise `fetch` runs against the fused `d.signal`, and `translateAbortOrNetwork` classifies a thrown error by the signal (`timeoutOf` → `WEB_FETCH_TIMEOUT`, else aborted → `WEB_ABORTED`, else network → `WEB_PROVIDER_ERROR`). The public error-code contract is unchanged, and `TimeoutReason` never crosses the web seam as the public error.
+- **bash** — `resolve()` clamps the request into an explicit spec. Foreground `run()` creates the deadline and passes its signal to process execution, whose existing abort listener performs the process-group kill. The executor classifies the first abort as timeout or cancellation. Background starts remain timeout-free and forward only upstream cancellation.
+
+## Consequences
+
+- `runBash`'s outcome no longer independently latches `timedOut` and `aborted`; a timeout and a user abort racing before process close now report a single first-abort cause instead of both being true. The uniform SIGTERM→grace→SIGKILL kill is unchanged, and the seam type `BashRunResult` keeps both booleans (now mutually exclusive), so `dsh-tool-bash`'s result rendering is untouched.
+- `SpawnSpec.timeoutMs` and `SpawnOutcome.timedOut`/`aborted` were removed rather than kept as always-zero/always-false vestiges: with `runBash` owning no timer and the executor owning classification, they were read nowhere. This is the one deviation from the literal proposal shape (which passed `timeoutMs: 0` into `runBash`); an always-0 field read by nothing is dead weight under the per-file coverage gate.
+- web_fetch shed its bespoke controller/timer/listener/reason-recovery; the classifier now keys off the deadline signal (`timeoutOf` + `aborted`) rather than the thrown error's shape, which is robust across both the request-phase reject-with-reason and the read-phase bare-`AbortError`.
+- `AbortSignal.any` and `using`/`Symbol.dispose` enter the repo for the first time here (Node ≥ 24 baseline, already met).
+
+Out of scope, named to mark the boundary: `web_search` can gain an optional model-facing `timeout_ms` once its tool-schema/snapshot coverage is planned; future ripgrep-backed fs discovery tools can consume the same provider-owned deadline shape once they exist; a `tools/execute` waterfall middleware could arm a default deadline for every tool call by driving `exec.signal` — that would be a plugin that *consumes* this library and still only notifies, the hard kill remaining each capability's job.
+
+## Alternatives considered
+
+**A unified timeout *plugin* / `ctx.timeout` service.** Rejected on microkernel grounds. A service that could stop any tool's work would have to understand every capability's termination mechanism (process-group SIGKILL, socket teardown, syscall-boundary checks) — the "kernel knows too much" the architecture forbids. Codex's `ExecExpiration` is scoped to the exec family precisely because the kill it drives (`killpg`) is process-family-specific; MCP and model-stream keep their own. There is no coherent middle layer that owns termination for everything, so the shared piece can only be the pure timing/classification half — a library, not a service.
+
+**Per-tool ad-hoc timeout, no shared code (the prior status quo, and Claude Code's choice).** Rejected because it was already producing divergence and duplicated correctness burden: web_fetch hand-rolled the exact controller/reason logic that future network/process-backed tools would each have to re-derive, and the fusion + `signal.reason` recovery are the error-prone parts. Claude Code tolerates full duplication; this repo has a single shared abort channel (`exec.signal` on every `execute`) that makes a small shared primitive strictly cleaner, so the cost/benefit differs.
+
+**A `withTimeout(promise, ms)` wrapper instead of a signal factory.** Rejected because racing a promise against a timer resolves the *tool-call* promise on deadline without stopping the underlying work — the child process or fetch socket leaks on. Handing out a signal and requiring the capability to listen is what forces a real termination path to exist. This mirrors the "dispose must reach quiescence, not just request it" defensive rule.
+
+**Keep separate bash timeout and cancellation triggers.** Rejected because one deadline signal removes the bespoke timer and standardizes classification. Racing causes report whichever abort arrived first, while the existing SIGTERM-to-SIGKILL termination path remains unchanged.

.agents/notes/implemented/architecture/2026-07-06-tool-result-retention-library.md

@@ -0,0 +1,155 @@
+# Agent Note: Tool result retention library
+
+Status: implemented
+
+## Problem
+
+Several model-facing tools already bound the amount of context they return, but each one owns a different local mechanism and vocabulary: bash keeps a tail plus spill files, web search caps source lists, web fetch caps body content, and `glob` / `grep` discovery needs an inline first page while keeping exact omission metadata for the full result set. A single `truncate(text)` helper cannot cover those cases: item tools need item counts and grouping outside the primitive, while text tools need byte budgets and UTF-8-safe head/tail cuts.
+
+The shared abstraction the tools need is **retention**, not generic collection. A caller feeds items or text chunks into a bounded object and later receives the retained content plus exact omission metadata. Tool-specific code still owns business semantics: file grouping, line numbering, exit codes, provider error states, spill files, and model-facing prose. The common library owns only the mechanical question "what did we keep, and what did we omit?"
+
+## Decision
+
+`@deepseek-ai/dsh-retention` lives under `packages/util/` (peer to `dsh-brand` and `dsh-timeout`) and owns bounded model-facing output. It is a library of pure classes and functions, **not** a Cordis service or plugin: it takes no `ctx`, registers nothing, holds no cross-call state, and emits no events. Tool packages import it directly when they need bounded output.
+
+The library has two independent retainers:
+
+- `ItemRetainer<T>` handles ordered logical units such as paths, grep matches, or search sources. It supports `head` retention only in v1, while keeping the retainer shape open to additional retention strategies later.
+- `TextRetainer` handles byte-oriented text streams such as bash stdout/stderr or web response bodies. It supports `head`, `tail`, and `headTail` retention while preserving UTF-8 boundaries at `finish()`.
+
+Both retainers return a small `PushDecision` after each `push()` so callers can tell whether that unit/chunk was fully retained and whether the accumulated result is now truncated. Omission counts are exact because callers keep feeding every observed item/chunk.
+
+```ts ignore-check
+/**
+ * How much content the retainer omitted.
+ *
+ * `unknown` is reserved for callers that omit without a count; the retainers
+ * themselves return `none` or `exact`.
+ */
+type Omitted =
+  | { kind: 'none' }
+  | { kind: 'exact'; count: number }
+  | { kind: 'unknown' }
+
+interface PushDecision {
+  kept: boolean
+  truncated: boolean
+}
+
+/**
+ * Final result for ordered logical units.
+ */
+interface RetainedItems<T> {
+  items: T[]
+  truncated: boolean
+  seen: number
+  kept: number
+  omitted: Omitted
+}
+
+/**
+ * Final result for text streams.
+ *
+ * The returned `text` is safe to send to a formatter; the retainer does not add
+ * tool-specific headers, exit markers, XML tags, or recovery instructions.
+ */
+interface RetainedText {
+  text: string
+  truncated: boolean
+  omittedBytes: Omitted
+}
+```
+
+### Strategies
+
+Item retention supports a head window. Text retention supports head, tail, and headTail byte windows.
+
+```ts ignore-check
+type ItemRetentionStrategy =
+  | {
+      /** Keep the first `maxItems` units. Use for `glob`, `grep`, and web sources. */
+      kind: 'head'
+      maxItems: number
+    }
+
+type TextRetentionStrategy =
+  | {
+      /** Keep the first `maxBytes` bytes. */
+      kind: 'head'
+      maxBytes: number
+    }
+  | {
+      /** Keep the final `maxBytes` bytes. Requires reading to the end. */
+      kind: 'tail'
+      maxBytes: number
+    }
+  | {
+      /** Keep a stable prefix and suffix, omitting the middle. Requires reading to the end. */
+      kind: 'headTail'
+      headBytes: number
+      tailBytes: number
+    }
+```
+
+### Tool mapping
+
+`read` is intentionally outside the v1 retention library. Its `read-render` helper owns a file-specific pagination contract: `offset` / `limit`, line numbers, `totalLines`, offset-out-of-range errors, per-line preview truncation, and a selected-output byte cap that can stop scanning mid-window. That is a line-window renderer, not a generic retention primitive. It may share future neutral notice helpers, but it should not pass its already-selected window through `ItemRetainer`.
+
+`FsGlobEntry` and `FlatGrepMatch` below are the intended discovery-tool item shapes, not existing retention-library exports. `FsGlobEntry` is one backend-derived path, and `FlatGrepMatch` is one ungrouped grep match before the backend groups retained matches by file.
+
+`glob` uses `ItemRetainer<FsGlobEntry>` with `{ kind: 'head', maxItems: globMaxResults }` after collecting the full sorted path list. The tool keeps the retained first page inline and may save the full list through the spill seam. Path mapping, skipped candidates, and `incomplete` stay outside the retainer.
+
+`grep` uses `ItemRetainer<FlatGrepMatch>` with `{ kind: 'head', maxItems: grepMaxMatches }` before grouping. The executor parses ripgrep output, maps paths, applies per-line preview truncation, and pushes flat matches. After `finish()`, the tool groups retained matches by file and can save the full match list through the spill seam when the inline result is capped. Grouping is not part of the retainer because the cap is total matches, not files; per-match preview truncation and `incomplete` are also separate from result-level retention.
+
+`bash` can use `TextRetainer` with `tail` or `headTail` and reads to process completion. The bash executor still owns spill files, exit status, signal, timeout, and background-task behavior; the retention helper only replaces ad hoc in-memory head/tail accounting where that behavior is desired. Long-running task ownership remains orthogonal to the [generic long-running tool runtime](2026-06-20-generic-long-running-tool-runtime.md).
+
+`web_fetch` can use `TextRetainer` with `head` or `headTail`, or keep provider-owned body caps when the provider must read and decode internally. Either way, the fetch result's `truncated` remains a provider/tool fact, and the library only supplies retained text and omission metadata.
+
+`web_search` can use `ItemRetainer<WebSearchSource>` with `head`. Current providers often return an array, so this is post-hoc but still standardizes notices.
+
+### Notices
+
+The library exposes a neutral notice shape and a tiny formatter hook, but tools provide the user-facing words. A grep footer says "Narrow the pattern, path, or include"; a web fetch footer says "Fetch a more specific URL or section"; bash may point to a spill file. The retainer cannot know those recovery actions.
+
+```ts ignore-check
+interface RetentionNotice {
+  scope: string
+  strategy: 'head' | 'tail' | 'headTail'
+  unit: 'items' | 'bytes' | 'chars' | 'lines'
+  limit: number | { head: number; tail: number }
+  kept: number
+  omitted: Omitted
+}
+
+const formatGrepNotice = (notice: RetentionNotice): string =>
+  formatRetentionNotice(
+    notice,
+    ({ kept }) => `Results capped at ${kept}. Narrow the pattern, path, or include to see more.`,
+  )
+```
+
+The formatter hook is deliberately small: a tool turns a `RetentionNotice` into its own footer text. The helper may standardize omission wording, but it does not own recovery guidance.
+
+`truncated` means the retainer omitted otherwise-available content because of a budget. It does not mean the upstream was incomplete. Tools keep separate fields for permission failures, skipped binary files, provider partial failures, unreadable candidates, invalid UTF-8, and any other "could not inspect" condition.
+
+## Consequences
+
+**What shipped.** `@deepseek-ai/dsh-retention` exports `ItemRetainer`, `TextRetainer`, the result types (`RetainedItems`, `RetainedText`), the strategy types (`ItemRetentionStrategy`, `TextRetentionStrategy`), `Omitted`, `PushDecision`, `RetentionNotice`, and the neutral notice helpers `describeOmitted` / `formatRetentionNotice` — with no dependency on Cordis or any tool package. Unit tests cover item-head retention with exact omission counts, text-head retention, text-tail retention, head-tail byte retention, zero budgets, UTF-8 boundary handling (2-, 3-, and 4-byte codepoints and invalid lead bytes at each cut), and unknown omission wording.
+
+**What is documented but not yet migrated.** `glob`, `grep`, `bash`, `web_fetch`, and `web_search` have their mappings documented in the [package README](../../../../packages/util/retention/README.md), but not every tool has been migrated onto the library in this change; migration is deliberately separate follow-up work. `read` is documented as intentionally out of scope: its `read-render` line-window contract (`offset`/`limit`, `totalLines`, offset-range errors, per-line preview truncation, a byte cap over the selected window) is not generic retention, and one `Omitted` count cannot represent both sides of a line window.
+
+**Boundaries the library holds.** `truncated` means the retainer omitted otherwise-available content because of a budget; it never means the upstream was incomplete. Tool-specific states — `incomplete`, permission failures, provider partial failures, binary skips, bash spill-path recovery, invalid UTF-8 — stay in tool-domain fields, outside the retainer. When a future change migrates a tool, that package's README and tests must prove the model-facing result text is unchanged except for deliberate notice wording.
+
+**Tradeoffs accepted.** The v1 surface deliberately supports only item `head` retention and text `head` / `tail` / `headTail`; windows, grouped budgets, sort-aware caps, and upstream-stop control wait until a second consumer proves the need. Text retention counts bytes for process/body safety, leaving character- and line-level preview budgets as separate tool-owned concerns.
+
+## Alternatives considered
+
+**Post-hoc `truncate(text)` only.** Rejected: it matches Codex's history/tool-output truncation use case but loses item counts, grouping boundaries, UTF-8-safe byte windows, and exact omission metadata.
+
+**One generic `Collector<T>` with pluggable callbacks.** Rejected for v1: it hides the two important resource modes. Logical item retention counts items; text retention counts bytes and preserves UTF-8 boundaries. Separate `ItemRetainer` and `TextRetainer` names make that difference explicit while keeping the API small.
+
+**Put `read` windowing behind `ItemRetainer`.** Rejected for v1: `read` is the only current window consumer, and its semantics are file pagination rather than generic retention. A single `Omitted` count cannot represent both sides of a line window, and `read` also carries `totalLines`, offset-range errors, per-line preview truncation, and a byte cap over selected output. Keeping `read-render` tool-owned avoids growing the shared library around one special case.
+
+**Make truncation part of `ToolExecutionResult`.** Rejected: the tool registry would have to understand tool-specific recovery guidance, grouping, line numbering, exit status, and provider semantics. Retention is a library used before a tool returns `ContentBlock[]`; the model-facing result remains tool-owned.
+
+**Expose limits in every model-facing tool schema.** Rejected as the default: Claude Code's grep exposes `head_limit` / `offset`, but this harness keeps routine budgets as deployment config unless the model genuinely needs pagination control. A future read-like continuation field can be added per tool; it does not belong in the shared retention primitive.

.agents/notes/implemented/architecture/2026-07-07-tool-call-timeout-policy.md

@@ -0,0 +1,109 @@
+# Agent Note: Tool-call timeout policy as a plugin
+
+Status: implemented
+
+## Problem
+
+The [timeout/deadline Agent Note](2026-07-06-timeout-deadline-library.md) extracted the timing-and-classification primitive into `@deepseek-ai/dsh-timeout`, but timeout policy was still attached to individual capabilities and model-facing schemas. `bash` exposed `timeoutMs`; `web_fetch` exposed `timeout_ms`; `web_search` had no model-facing timeout even though providers already honor `exec.signal`; a future grep/glob tool would either import the timeout library directly or invent its own timeout policy. That is the wrong authoring shape for a plugin SDK: a tool author should normally forward `exec.signal` to the implementation it calls, and deployment policy should decide the budget.
+
+At the same time, not every timeout in the repo is a model-facing tool-call budget. Hooks execute command hooks by calling `ctx.bash` directly, not through `ctx.tools.execute()`, and the `bash` model tool multiplexes foreground execution, background start, background polling, and hook reuse through the same backend. Moving every timeout into a tool plugin in one step would conflate those paths and risk breaking hook timeout semantics.
+
+## Decision
+
+Tool-call timeout is a policy that applies only to model-facing tool execution, in three parts:
+
+- `@deepseek-ai/dsh-timeout` remains the shared library that owns `deadline()` and `timeoutOf()`.
+- `@deepseek-ai/dsh-tools` has an around-dispatch waterfall, `tools/execute`, between `tools/pre-execute` and `tools/post-execute`.
+- `@deepseek-ai/dsh-timeout-policy` reads each tool's declared `timeoutMs` from the registry and wraps a call that has one by deriving a new `exec.signal`.
+
+The execution pipeline is:
+
+```text
+ctx.tools.execute(exec)
+  -> tools/pre-execute
+  -> tools/execute
+       -> registry dispatch (the base next())
+            -> tool.execute(args, exec)
+            -> thrown tool errors normalize to ToolExecutionResult
+  -> tools/post-execute
+```
+
+The default behavior is conservative: a tool that declares no `timeoutMs` receives no `TOOL_TIMEOUT` deadline from the plugin.
+
+### The `tools/execute` around seam
+
+`@deepseek-ai/dsh-tools` declares a `tools/execute` waterfall whose base `next()` is the dispatch-with-normalization thunk — the same inner `try`/`catch` that turns a thrown tool (or unknown tool) into an `isError` `ToolExecutionResult`. A listener receives `(exec, next)`: it calls `next()` to delegate to dispatch (returning its result, optionally wrapped) or returns a replacement result to short-circuit dispatch. The whole pipeline still sits inside `execute`'s outer try/catch, so a throwing listener becomes an `isError` result, never a turn failure.
+
+That the catch is the base `next` — not something outside the waterfall — is load-bearing: when a provider sees the timeout signal and throws its own upstream-abort error, registry dispatch first converts it to a normal error result, and only then can `timeout-policy` replace the final result with `TOOL_TIMEOUT`.
+
+### The `timeout-policy` plugin
+
+The plugin is `@deepseek-ai/dsh-timeout-policy`, a zero-config function/namespace plugin (`name` / `inject` / `apply`) in the `packages/timeout/` group. The per-tool budget is DECLARED on the tool, not on this plugin: a `ToolDefinition` carries an optional `timeoutMs`, which the owning tool plugin sets from its own config. `dsh-tool-web`, for example, resolves `fetchTimeoutMs` / `searchTimeoutMs` (default 30000) onto the `web_fetch` / `web_search` definitions:
+
+```yaml
+- id: timeout-policy
+  name: '@deepseek-ai/dsh-timeout-policy'
+- id: tool-web
+  name: '@deepseek-ai/dsh-tool-web'
+  config:
+    fetchTimeoutMs: 30000
+    searchTimeoutMs: 30000
+```
+
+Timeouts live on tool definitions rather than a free-text name map, eliminating misspelled unused policy. `defineTool` validates a positive finite budget. During dispatch the enforcer derives a deadline signal, restores the caller signal afterward, and converts its own expiry into `TOOL_TIMEOUT`; tools without a budget pass through unchanged.
+
+Signal replacement is by **in-place mutation of `exec.signal`**, not by passing a new object to `next()`. Cordis's waterfall `next()` ignores any arguments handed to it and re-invokes downstream listeners with the shared payload array (`vendor/cordis/src/events.ts`), so the documented cordis idiom — mutate the shared object, then delegate — is the only mechanism that reaches dispatch. The plugin restores `exec.signal` to the caller's original in a `finally` so `tools/post-execute` never sees this plugin's (possibly already-aborted) deadline signal.
+
+`timeout-policy` owns both uses of the `TOOL_TIMEOUT` code: the internal deadline code passed to `deadline()`/`timeoutOf()` (scoped so a nested outer deadline reads as an ordinary cancel) and the structured tool-result error code. Its replacement result is:
+
+```ts ignore-check
+function toolTimeoutResult(timeoutMs: number): ToolExecutionResult {
+  return {
+    content: [{ type: 'text', text: `Error: tool call timed out after ${timeoutMs}ms` }],
+    isError: true,
+    error: { name: 'ToolTimeoutError', code: 'TOOL_TIMEOUT' },
+  }
+}
+```
+
+This is a cooperative deadline. It does not kill arbitrary work by racing the tool promise; the tool or the capability it calls must honor `exec.signal` and reach quiescence. Declaring `timeoutMs` therefore MEANS "this tool is cooperative with `exec.signal`", which the plugin README states as its contract.
+
+No new session event is needed for reconstructability: `TOOL_TIMEOUT` is the final model-facing `tool/result` for that call, so the existing session log already records the content and structured `{ name, code }` error the next model request sees.
+
+### Existing tool adaptation
+
+`web_fetch` and `web_search` are migrated. `dsh-tool-web` keeps ownership of their model-facing schemas, and those schemas expose no timeout knob: `web_fetch` dropped its `timeout_ms` parameter to match the reference-agent shape, and `web_search` stays query-only. The tool bodies do not import `@deepseek-ai/dsh-timeout`; they forward `exec.signal` to `ctx.web`.
+
+`dsh-web-fetch-local` keeps one configured provider-level `timeoutMs` as a large resource backstop for direct `ctx.web.fetch()` callers and misconfigured deployments; it owns no model-facing timeout. When a `TOOL_TIMEOUT` signal reaches the fetch provider first, provider-scoped classification treats it as upstream `WEB_ABORTED`, and the outer `tools/execute` wrapper replaces the final tool result with `TOOL_TIMEOUT`. A shipped web-tool deployment configures the provider backstop above the `timeout-policy` budget so the tool-call policy normally wins for model calls.
+
+`bash` stays on the current backend timeout path. `dsh-tool-bash` continues to expose `timeoutMs` and `run_in_background`; `dsh-bash-local` continues to use `@deepseek-ai/dsh-timeout` for `BASH_TIMEOUT`; hook bridges continue to call `runHook()` and pass `timeoutMs` through `ctx.bash`. This keeps foreground/background/hook behavior stable.
+
+`read`, `write`, `edit`, `todo_write`, `task_list`, and `task_kill` do not opt into tool-call timeout. `task_output` owns its bounded wait because a wait timeout is a successful live-status result, not a tool failure.
+
+A future model-facing grep/glob tool can be implemented on top of `ctx.bash` without importing `@deepseek-ai/dsh-timeout`: it forwards `exec.signal` to `ctx.bash`, and declares its own `timeoutMs` (from its plugin's config) for the enforcer to apply. If bash-local's backend timeout becomes a problem for such a tool, the bash seam can later add a caller-owned-deadline mode; that is outside this cut.
+
+## Alternatives considered
+
+**Name the plugin `tool-timeout`.** The literal Agent Note name matched the `gen-tool-catalog` completeness guard's `packages/*/tool-*` glob, which requires every match to register a model-facing tool. This plugin registers none — it is a `tools/execute` wrapper — so a `tool-*` name would either fail `verify-tool-catalog` or force a misleading boot entry. The package is `@deepseek-ai/dsh-timeout-policy` in a new `packages/timeout/` group; the cordis.yml `id` can still be `timeout-policy`.
+
+**Keep per-tool timeout handling only.** This was the shape for `bash` and `web_fetch`, and it matches Claude Code and Codex for shell commands. It loses for web-style tools because every new timeout-capable tool must choose validation, cap semantics, docs, snapshots, and classification. The plugin centralizes policy and classification while leaving each tool's schema focused on business input.
+
+**Move all timeout policy out of bash-local immediately.** Cleaner long-term — bash-local would become a pure subprocess executor and all callers would own their deadlines. It loses as the first step because hooks call `ctx.bash` directly and the bash model tool has foreground/background semantics that are not the same tool-call lifetime. Keeping `BASH_TIMEOUT` preserves those paths while tool-call timeout proves itself on simpler tools.
+
+**Use a global default budget for every tool.** Convenient, but it surprises tool authors: any tool that accidentally runs longer than the global budget would start failing once the plugin loads. A per-tool declared budget makes adoption deliberate.
+
+**Expose a model-facing `timeout_ms` override.** Claude Code's `WebFetch`/`WebSearch` and Codex's web tools keep timeout out of the model-call shape. A model override would make timeout part of prompt semantics and force schema/argument-stripping rules into `timeout-policy`. Web timeout stays deployment policy only.
+
+**Let `timeout-policy` match tool arguments itself.** A rule engine such as "disable timeout when `bash.run_in_background` is true" would make the policy plugin know tool-specific argument semantics. Avoided by not migrating bash to tool-call timeout.
+
+**Use `tools/pre-execute` plus `tools/post-execute` instead of a new around seam.** A pre listener could arm a deadline and mutate `exec.signal`; a post listener could classify and replace. That loses because the deadline lifetime would cross two independent waterfalls: a call-id map, cleanup on every pre-deny/tool-throw/post-throw/dispose path, and ordering rules with every other listener. `tools/pre-execute` is also the allow/deny gate, not an execution wrapper. `tools/execute` gives the timeout one lexical scope: arm, delegate, classify, dispose.
+
+**Use `Promise.race` to enforce timeouts for non-cooperative tools.** Rejected for the same reason as the timeout-library Agent Note: it returns control to the caller while the underlying process, fetch, or provider operation may still be running. The plugin only sends a signal; termination remains the implementation's responsibility.
+
+## Consequences
+
+- `@deepseek-ai/dsh-tools` gains an around-dispatch surface after the interception seams deliberately split pre/post tool hooks. Its contract is narrow — wrap registry dispatch, not replace the pre-gate or post-result policy — and the base `next()` is dispatch-with-normalization so a wrapper never sees a raw tool throw.
+- Multiple `tools/execute` listeners compose by ordinary Cordis waterfall order: a listener that calls `next()` wraps downstream listeners plus dispatch; one that returns without `next()` short-circuits them. A deployment combining timeout with a future retry/sandbox/metrics wrapper chooses semantics by registration order ("timeout covers the whole retry" vs "timeout covers each attempt").
+- Opt-in by declaration is a deliberate misconfiguration risk: a tool can declare a `timeoutMs` without honoring `exec.signal`, and that tool will not stop on timeout. The plugin contract states that declaring a budget means cooperative; the web tools prove the pattern on tools that already forward the signal.
+- During the transition `bash` and the migrated web tools use different timeout paths on purpose: `TOOL_TIMEOUT` is the model-facing tool-call budget, while `BASH_TIMEOUT` remains the bash backend timeout used by bash and hooks.
+- Deviation from the literal proposal, recorded per the implemented-Agent Note rule: the plugin package is `@deepseek-ai/dsh-timeout-policy` (not `tool-timeout`), signal replacement is in-place `exec.signal` mutation before `next()` (not `next({ ...exec, signal })`, which cordis ignores), and the per-tool budget is declared on the `ToolDefinition` (`timeoutMs`, set by the owning tool plugin from its config) rather than mapped by tool name in this plugin's config — so the enforcer is zero-config and a mistyped tool name is impossible. All three are described in `## Decision` above.

.agents/notes/implemented/architecture/2026-07-08-agent-scope-contexts.md

@@ -0,0 +1,169 @@
+# Agent Note: The agent is a registration scope
+
+Status: implemented
+
+## Problem
+
+One application needs to share infrastructure across many agents while letting each agent have its own tools, prompt contributions, policies, and listeners. Shared adapters, persistence, and user interfaces belong to the deployment; a persona, tool variant, or listener often belongs to one agent.
+
+A separate service graph per agent duplicates shared infrastructure. One global registration graph has the opposite failure: an agent-specific contribution can leak into unrelated agents. Contributors need one ordinary registration mechanism that determines both who can see a contribution and when it is cleaned up.
+
+The mechanism also needs a publication boundary. An agent must not become visible before its local world is complete, and teardown must retain that world until final work has stopped.
+
+## Decision
+
+Every live agent owns one flat registration layer exposed as `agent.ctx`. Code registers through the context that owns a contribution; scope-aware services combine deployment-global registrations with exactly one matching agent layer; operations choose that layer from their real agent; and the layer exists for the agent's complete published lifetime.
+
+Cordis is the plugin framework underneath the SDK. A Cordis **context** is the object plugins use to access services and register effects whose cleanup follows that context. The [Cordis primer](../../../../docs/cordis-primer.md) explains the framework in more detail.
+
+For most contributors, the complete contract is four rules:
+
+| Question | Rule |
+|---|---|
+| Where do I register behavior for one agent? | Call the ordinary registration API through `agent.ctx` |
+| What does an operation for an agent see? | Deployment globals plus that agent's layer, using the owning service's merge rules |
+| Which scoped listeners run? | Unscoped listeners plus listeners registered for the operation's agent |
+| How long does the layer exist? | Setup completes before publication; disposal keeps it until work reaches quiescence |
+
+The scope is flat. Resolution never walks parent or sibling scopes, and lifetime ownership does not imply registration inheritance.
+
+```mermaid
+flowchart LR
+  plain["Plain plugin context<br/>cleanup follows the plugin"] -->|"registers into"| globalLayer["Deployment-global layer"]
+  agentAContext["agentA.ctx<br/>cleanup follows Agent A"] -->|"registers into"| agentALayer["Agent A layer"]
+  agentBContext["agentB.ctx<br/>cleanup follows Agent B"] -->|"registers into"| agentBLayer["Agent B layer"]
+
+  operationA["Operation for Agent A"] -->|"selects"| agentAView["Agent A view<br/>globals plus A local"]
+  globalLayer --> agentAView
+  agentALayer --> agentAView
+  operationB["Operation for Agent B"] -->|"selects"| agentBView["Agent B view<br/>globals plus B local"]
+  globalLayer --> agentBView
+  agentBLayer --> agentBView
+```
+
+The missing cross-edges are the isolation rule: Agent A's local registrations do not enter Agent B's view, and a parent's registrations do not enter a child merely because the parent owns the child's lifetime.
+
+The companion [runtime-design Agent Note](2026-07-12-agent-scope-runtime-design.md) explains the implementation and correctness reasoning. The [subagent composition-controls Agent Note](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns the separate `persona`, `toolFilter`, and `maxDepth` feature.
+
+### Registration origin chooses visibility and cleanup
+
+A registration made through a plain plugin context is deployment-global and is disposed with that plugin. The same method called through `agent.ctx` contributes to one agent and is disposed with that agent's scope.
+
+| Registration origin | Default visibility | Disposed with |
+|---|---|---|
+| Plain plugin context | Every eligible agent view | Registering plugin |
+| `agent.ctx` | Exactly that agent's view | Agent scope |
+
+Tools, prompt sections and variables, tool restrictions, guards, and scoped event listeners adopt this contract. Named local values ordinarily shadow a same-named global value for that agent; each owning service documents exceptions and merge behavior.
+
+The ordinary contributor pattern is to register the complete local world during agent setup:
+
+```js
+const handle = await ctx.agents.create({
+  sessionId: SessionId('reviewer'),
+  agentOptions: { model: 'model-name' },
+  setup(agentCtx) {
+    agentCtx.systemPrompt.section({
+      name: 'deployment:persona',
+      order: 0,
+      text: 'Review code, but do not modify files.',
+    })
+    agentCtx.tools.register({
+      name: 'review_summary',
+      description: 'Return the review summary.',
+      parameters: { type: 'object', properties: {} },
+      async execute() {
+        return [{ type: 'text', text: 'review complete' }]
+      },
+    })
+  },
+})
+
+ctx.tools.get('review_summary')                // undefined: not global
+ctx.tools.get('review_summary', handle.agent)  // the reviewer-local tool
+
+await handle.dispose()
+ctx.tools.get('review_summary', handle.agent)  // undefined: scope is gone
+```
+
+Setup receives a full trusted Cordis context so it can compose ordinary plugins and services. Its contract is composition-only: driving or publishing the in-flight agent through casts or internal registry calls is unsupported.
+
+### The operation chooses the view
+
+Registration origin and operation subject are separate facts. Calling a service through `agent.ctx` selects where a new registration belongs; it does not bind later reads to that agent.
+
+Tool lookup and execution receive the agent they act for. Prompt assembly receives an assembly context for the agent whose request is being built. Event dispatch receives its domain subject. This keeps shared service instances reusable across agents while making each operation's view explicit.
+
+Only services that adopt the scope contract resolve an agent layer. `agent.ctx` does not automatically change arbitrary Cordis service calls.
+
+### Scoped events keep routing separate from event data
+
+An event about Agent A normally reaches unscoped listeners and A-scoped listeners, not B-scoped listeners. An event without an agent subject reaches only unscoped listeners.
+
+At the Cordis level, `Scoped<T>` is an opaque routing receiver. It carries the filter used to choose listeners but is not the domain object. Event signatures therefore keep the real `Agent`, tool execution, approval request, or other subject as an explicit argument that listeners can inspect.
+
+A listener registered with `{ global: true }` deliberately bypasses contextual audience filtering while its cleanup still follows the registering context. Registry-membership notifications remain unfiltered because they describe shared registry state rather than one agent's operation. The generated [event catalog](../../../../docs/cordis-catalog/events.md) is the exhaustive event reference.
+
+### Creation publishes last and disposal revokes last
+
+`ctx.agents.create()` and `resume()` build an unpublished session, scope, agent, and driver. They await `setup`, admit the final session and agent entries, announce them in order, start the loop, and only then return a handle.
+
+An optional creation signal cancels work only while create or resume is pending. After the promise resolves, the returned `AgentHandle` owns explicit disposal.
+
+If loading, setup, admission, or publication fails, the private transaction rolls back everything it prepared. Concurrent operations using the same caller-supplied live ID may both reach setup, but final registry entry admits only one; every loser rejects and cleans its private resources. Sequential reuse after awaited disposal remains valid.
+
+`AgentHandle.dispose()` reverses the boundary. It deactivates creation or driving, waits for synchronous publication to unwind, stops and drains the driver and final session flushes, detaches the agent and session, and finally disposes the scope. Repeated or racing disposal requests join one completion promise.
+
+The calling Cordis context and the concrete AgentLoop factory are structural co-owners. Unloading either disposes the transaction or live agent.
+
+```mermaid
+flowchart TB
+  request["Create or resume"] --> privateWorld["Build private session, scope, agent, and driver"]
+  privateWorld --> setup["Await composition through agent.ctx"]
+  setup --> admission["Admit final session and agent entries"]
+  admission --> publish["Announce lifecycle and start the driver"]
+  publish --> live["Return AgentHandle"]
+
+  privateWorld -->|"failure, cancellation, or owner loss"| rollback["Rollback private work"]
+  setup -->|"failure, cancellation, or owner loss"| rollback
+  admission -->|"duplicate or owner loss"| rollback
+  publish -->|"listener failure or owner loss"| rollback
+  live -->|"handle or owner disposal"| quiesce["Stop and drain work"]
+  rollback --> quiesce
+  quiesce --> detach["Detach agent, then session"]
+  detach --> revoke["Dispose the agent scope"]
+```
+
+## Security and authority are non-goals
+
+Agent scopes compose trusted same-process registrations. They do not sandbox plugins, define a parent-to-child authority lattice, freeze grants at creation, or guarantee that a child can do no more than its parent.
+
+A parent may own a child whose visible tools are wider than its own because lifetime ownership does not donate or cap registrations. A plugin holding a Cordis context also runs in the same process and can call available services directly.
+
+Deployments that need non-escalation require a separate authority representation, propagation rule, and execution check. Parent-subset grants, creation-time authorization snapshots, explicit future-grant APIs, and generic capability/output/termination tags are outside this decision.
+
+## Alternatives considered
+
+The rejected designs either separate visibility from cleanup, cover only one registration family, duplicate shared infrastructure, or conflate lifetime ownership with inheritance.
+
+### Pass an agent option to every registration
+
+An API such as `tools.register(definition, { agent })` repeats scope plumbing in every registry and permits visibility ownership to drift from cleanup ownership. Registering through `agent.ctx` makes both facts follow one Cordis effect owner.
+
+### Filter events while keeping registries global
+
+Listener filtering prevents the wrong hook from running but does not scope tool schemas, executable lookup, prompt sections, variables, or other registered data. Agent-local composition would still require temporary global mutation.
+
+### Create one service graph per agent
+
+The required view is shared deployment services plus one local registration layer. Per-agent graphs duplicate adapters and complicate shared persistence, provider registries, and application boot.
+
+### Inherit parent registration scopes
+
+Parentage describes lifetime and conversation lineage, not a universal merge policy. Hierarchical lookup makes unrelated services inherit accidentally and cannot define security without a separate authority model.
+
+## Consequences
+
+Contributors use one familiar pattern: register shared behavior through a plugin context, register local behavior through `agent.ctx`, select the real agent on operations, and dispose the returned handle. Setup is atomic from an observer's perspective, and teardown preserves local behavior until work stops.
+
+The cost is explicit subject selection, asynchronous programmatic creation, and service-specific scope adoption. Flat registration scope is intentionally not authority, and subagent composition controls remain a separate feature rather than hidden scope semantics.

.agents/notes/implemented/architecture/2026-07-08-tool-output-spill-files.md

@@ -0,0 +1,189 @@
+# Agent Note: Tool output spill policy
+
+Status: implemented
+
+## Problem
+
+Tool outputs need bounded model-facing previews, but some oversized results are still useful later. A fetched page body or a verbose tool response should not consume the next model request in full, but the model should be able to inspect the complete formatted result later with existing file-reading tools.
+
+Before this change the behavior was uneven. `dsh-bash-local` already writes complete stdout/stderr streams to private temp spill files when its in-memory tail overflows, but ordinary text tool results were returned inline unless the tool hand-rolled its own cap. The [tool result retention library](2026-07-06-tool-result-retention-library.md) owns preview mechanics, but it does not own storage or an execution-pipeline policy that applies those mechanics to final tool results.
+
+The shape matches the timeout policy design: a tool author normally returns the text result, and a policy plugin enforces the deployment's default context budget. Tool-specific early spill remains possible later for outputs that do not survive to the final `ToolExecutionResult`; the first cut proves the default final-result path.
+
+## Decision
+
+A thin spill storage seam plus a default spill policy plugin, in a new `packages/spill/` group:
+
+| Package | Role |
+|---|---|
+| `@deepseek-ai/dsh-spill` | Interface: `ctx.spillStore`, vocabulary types, no storage implementation. |
+| `@deepseek-ai/dsh-spill-local` | Local backend: private, session-scoped file storage on the host filesystem. |
+| `@deepseek-ai/dsh-spill-policy` | Tool-result policy plugin: wraps final text results after dispatch and replaces oversized results with a retained preview plus a spill locator. |
+
+There is no dedicated model-facing consumer package. The consumer is the existing `ctx.tools` execution pipeline: `dsh-spill-policy` consumes final tool results through the `tools/post-execute` waterfall, and the model follows the backend-supplied retrieval hint for the returned locator.
+
+### Spill seam
+
+The storage seam is minimal: save text and return a locator plus retrieval hint.
+
+```ts ignore-check
+interface SpillStore {
+  saveText(input: SaveTextSpill): Promise<SpillRef>
+}
+
+interface SpillSource {
+  toolName: string
+  callId: CallId
+  label: string
+}
+
+interface SaveTextSpill {
+  owner: { sessionId: SessionId }
+  source: SpillSource
+  suggestedName: string
+  content: string
+}
+
+type SpillLocator = Branded<'SpillLocator'>
+
+interface SpillRef {
+  locator: SpillLocator
+  bytes: number
+  retrievalHint: string
+}
+```
+
+`SpillLocator` is a [branded](../../../../packages/util/brand) model-facing handle returned by the backend. The local backend renders it as a filesystem path; a remote or database backend can render a URI, key, or command token. Consumers treat it as opaque and render it with `retrievalHint` instead of assuming `read` is always the right retrieval mechanism. `SpillOwner.sessionId` is the save-time storage namespace: forked sessions inherit existing spill locators from the seeded log without copying or re-owning them, and new spills after the fork use the child session id. A retention-period cleanup may expire old locators with other old session artifacts; the spill seam does not define a per-session cleanup policy.
+
+`dsh-spill-local` owns only storage details: session-scoped directory selection, safe names, path-traversal protection, the write, and returning `{ locator, bytes, retrievalHint }`. It does not own retention policy, tool-result replacement, search, or file inspection. Files land at `<root>/session-<hash>/<random>-<safeName>`, where `root` is a configured path or a lazily-created private (0700) per-process temp dir, the session subdir is a short `sha256(sessionId)` prefix, and the leaf is a random hex prefix plus the caller's `suggestedName` sanitized to one path segment (mirrors the JSONL backend's `encodeSegment`). The write is `open(path, 'wx', 0o600)` — exclusive and owner-only, so a planted symlink cannot redirect it. The locator is the path, and the retrieval hint tells the model it can use `read` or `grep` on that path.
+
+### Spill policy
+
+`dsh-spill-policy` is a `tools/post-execute` result transformer with one configuration knob:
+
+```ts ignore-check
+interface Config {
+  /** Omitted means no automatic spill policy. Present means apply to oversized plain text tool results. */
+  maxInlineBytes?: number
+}
+```
+
+When `maxInlineBytes` is omitted the plugin registers nothing (a true no-op). When set, it applies a default policy to final plain-text tool results:
+
+1. Let the tool run normally, delegating via `next()` so a downstream listener settles the result first.
+2. Flatten the accepted final `ContentBlock[]` only when it is entirely plain text; a result with any non-text block is left untouched.
+3. If its UTF-8 byte size is at or below `maxInlineBytes`, leave it unchanged.
+4. If it is larger, call `ctx.spillStore.saveText()` with the full final text.
+5. Replace the model-facing result with a retained head/tail preview plus the spill reference.
+
+The preview is an implementation default owned by the policy: a head/tail split of `maxInlineBytes` via the retention library's `TextRetainer`. Future config can expose preview sizing only after a second deployment needs it.
+
+The replacement text is intentionally generic because the policy only knows the final formatted tool result, not the tool's internal resource:
+
+```text
+<retained preview>
+
+(Omitted N bytes. Full formatted result stored at: /.../session-.../....txt. Use read with offset/limit, or grep this path to search within it.)
+```
+
+If `ctx.spillStore.saveText()` fails (permissions, ENOSPC, backend unavailable), or the call has no session owner, or no backend is loaded, the plugin logs the reason and returns the original result unchanged. Spill failure never turns a successful tool call into an `isError` result or hides the inline result.
+
+The policy skips `read` to avoid a circular `read -> spill file -> read again` loop. Additional opt-out configuration is deferred until a real second tool needs it.
+
+## Showcase: web_fetch
+
+`web_fetch` is the first showcase because it returns a naturally large text result and needs no tool-specific spill code. The tool is ordinary:
+
+```ts ignore-check
+ctx.tools.register(defineTool({
+  name: 'web_fetch',
+  async execute(args, exec) {
+    const result = await ctx.web.fetch({ url: args.url }, exec.signal ? { signal: exec.signal } : undefined)
+    return [{ type: 'text', text: formatFetchOutput(result) }]
+  },
+}))
+```
+
+With `dsh-spill-policy` configured, a large formatted fetch result is automatically retained and spilled. A deployment demonstrates the behavior by setting the provider resource cap higher than the policy cap:
+
+```yaml
+- id: web-fetch-local
+  name: '@deepseek-ai/dsh-web-fetch-local'
+  config:
+    maxBodyChars: 500000
+
+- id: spill-local
+  name: '@deepseek-ai/dsh-spill-local'
+
+- id: spill-policy
+  name: '@deepseek-ai/dsh-spill-policy'
+  config:
+    maxInlineBytes: 50000
+```
+
+This separation is important. `web-fetch-local` still owns resource caps (`maxResponseBytes`, `maxBodyChars`) to protect network, memory, and decoding work. `spill-policy` owns only the model-facing context cap after the result already exists. If the provider already returned `truncated: true`, the spill file contains the full formatted result the tool returned, not the full original webpage; the policy does not claim otherwise.
+
+## Relationship to retention and early spill
+
+Retention is separate from spill storage:
+
+- `@deepseek-ai/dsh-retention` owns preview mechanics (`TextRetainer`, `ItemRetainer`, and omitted metadata).
+- `@deepseek-ai/dsh-spill` owns saving final text and returning a locator plus retrieval hint.
+- `@deepseek-ai/dsh-spill-policy` applies the default final-result policy in the tool pipeline, composing the two.
+
+The final-result policy cannot replace tool-owned early spill. Some useful content is not present in final `ToolExecutionResult.content`:
+
+- `bash` final output is already a tail plus a temp spill path; the complete stdout/stderr streams live in executor files.
+- `subagent` final output is the child final answer, not the child rollout.
+- Future tools may produce runtime artifacts that are never represented by their final `ToolExecutionResult.content`.
+
+Those cases can consume `ctx.spillStore` directly in later work. They are not part of the first showcase.
+
+## Non-goals
+
+- No new model-facing `artifact_read` or `artifact_search` tool in v1.
+- No per-tool retention configuration in v1.
+- No model-facing timeout/truncation arguments.
+- No migration of `read` output into spill files.
+- No replacement for provider/resource caps such as `web-fetch-local.maxBodyChars`.
+- No bash temp-file normalization or subagent rollout capture in the first cut.
+
+## Deferred
+
+- `saveFile()` / `linkOrCopy` for existing executor spill files, needed for bash normalization.
+- Tool-owned spill for subagent rollouts (`await run.result`, read in-process child session before `run.dispose()`, save JSONL).
+- Per-tool opt-out or per-tool policy declarations if the built-in `read` skip is insufficient.
+- Remote or database storage backends for ACP or remote environments where a local path is not meaningful.
+- Cleanup and retention policy for old spill files, likely tied to session cleanup.
+
+## Testing
+
+- `dsh-spill` unit tests pin the seam contract: registration as `ctx.spillStore`, one-implementation-per-context, and disposal release.
+- `dsh-spill-local` unit tests cover `saveText`, `encodeSegment` sanitization (separators/tilde/whole-segment dots/empty), the session-hash directory, owner-only permissions, distinct paths per save, the configured/private root, and a storage-failure rejection.
+- `dsh-spill-policy` unit tests drive real tools through `ctx.tools.execute`: disabled-mode no-op, oversized-text replacement, small/non-text passthrough, `read` skip, best-effort fallback (save failure / no backend / no owner), and downstream-composition (bounding a replaced result, preserving `additionalContexts`).
+- `dsh-tool-web` integration drives `web_fetch` through `ctx.tools.execute` with the real `spill-local` backend + policy, proving the model-facing text changes only by the deliberate spill notice while the spill file holds the full formatted result.
+- The `repl-agent` example loads `spill-local` + `spill-policy`, so its keyless Loader smoke exercises the real load path (the namespace-plugin export shape + `inject`).
+
+## Consequences
+
+The default policy only sees final formatted text. It cannot preserve provider-internal content that was already capped or runtime artifacts that were never part of the result. This is acceptable for the first cut because the showcase is final-result spill, not early spill; tool-owned early spill remains deferred work.
+
+Returning real paths from the local backend keeps v1 simple and matches proven agent-tool behavior, while the seam itself only promises an opaque locator plus retrieval hint so remote backends can return non-file locators.
+
+The local-backend value proposition depends on the existing `read`/`grep` tools being able to inspect the returned local path, even when the spill directory is outside the session cwd. That holds today because the filesystem policy records observations and write guards but does not confine reads to the workspace. A future workspace-confinement policy must either allow local spill paths explicitly or use a non-file spill backend whose retrieval hint points at a supported reader.
+
+**Snapshot gap.** No ACP snapshot scenario covers the transcript-visible `web_fetch` spill notice yet. The ACP snapshot harness replays keyless and cannot hit the live web, and a `web_fetch` spill requires a real over-cap HTTP body; a deterministic scenario would need a seeded loopback fetch target the replay tree does not currently wire (the examples do not load `tool-web` at all). The behavior is covered instead by the `dsh-tool-web` integration test against a loopback server. Closing the gap is follow-up work: wire `tool-web` + a seeded fetch target into the ACP example, then record a `web-fetch-spill` scenario.
+
+The policy can become too large if it starts owning tool-specific semantics. It stays narrow: plain-text final results only. Tool-owned early spill remains future work.
+
+## Alternatives considered
+
+**Require each tool to opt in with a retention declaration.** Rejected for v1: the goal is a default behavior similar to Claude Code's generic tool-result persistence. A single `maxInlineBytes` deployment knob is enough to prove the shape.
+
+**Make `tool-results` a broad tool-result platform.** Rejected: a broad package name invites retention policy, result replacement, preview wording, search, and early spill into one seam. The shared storage part is smaller: save text and return a locator plus retrieval hint.
+
+**Use `ctx.fs.writeText` or the model-facing `write` tool.** Rejected: workspace filesystem writes carry project-file semantics, write/edit policy, observation state, and user-facing side effects. Spill files are runtime artifacts, not model-authored workspace edits. The existing `read` tool may inspect them later, but creation belongs to the runtime spill seam.
+
+**Let `web-fetch-local` fetch without caps and rely on spill-policy.** Rejected: spill-policy runs after the final tool result exists and cannot protect network, memory, or decoding resources. Provider resource caps stay mandatory.
+
+**Merge retention into spill.** Rejected: retention and spill have different responsibilities. `TextRetainer`/`ItemRetainer` decide what preview is kept and what was omitted; spill storage only saves the final text the policy asks it to save.

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md: f1a1868cd00007fb24efb21779dcc94c098b54e2
+2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md: a4993de2830301610bb2a9b0d28e8bbdf0ed9c46

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md

@@ -0,0 +1,61 @@
+# Agent Note: After-call compaction pressure and context-overflow recovery
+
+Status: implemented
+
+English | [中文](2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md)
+
+## Problem
+
+`agent/pre-step` runs before final request routing and before assistant output, tool results, buffered context, and steering exist. Even with the assembled prompt and session prefix, its pressure view is provisional because `agent/request` can still change routing or call configuration and tool schemas are not frozen with those inputs. Adding fields cannot make pre-call state describe a completed call and couples the generic seam to compaction.
+
+Successful calls are not the only pressure signal. A provider can reject a request for exceeding its context window before it returns usage, and some successful calls omit usage. The system therefore needs replayable post-call pressure plus a narrow failure-recovery path that preserves the provider error whenever compaction cannot prove useful progress.
+
+## Decision
+
+### Successful pressure moves to a durable post-step checkpoint
+
+`agent/pre-step` is narrowed to `(agent, turn, step, signal)`. It remains a generic serial checkpoint before `step/start`, but it carries no compaction-only prompt or prefix fields.
+
+The loop fires awaited serial `agent/post-step(agent, turn, step, signal)` after assistant output, every dispatched or synthetic tool result, post-tool context, and steering are durable, but before `step/end`. This placement gives pressure policy the complete successful-call state without splitting an assistant tool call from its result. A listener failure is an ordinary turn failure; it never enters model-request recovery.
+
+`dsh-compact-basic` reads the exact latest routed model from the durable request header only to establish that a completed route exists, then asks the singleton `ctx.tokenMeter` to measure the canonical logged envelope and current surface. It does not fall back to `AgentOptions.model` for automatic pressure. A headerless session has no completed routed request to assess and produces no work; any durable non-empty model name uses the same estimator. Operational measurement or summarization failures warn and continue with full history.
+
+### Request recovery is limited to the final model boundary
+
+`RequestError`, `RequestErrorDecision`, and the `agent/request-error` waterfall represent failures after the final adapter has been selected. Each returned stream handle owns a private failure set that preserves the original thrown error identity across dispatch, iterator construction, and iteration without leaking nested-call provenance into an outer call. Terminal in-band `error` or `aborted` finishes enter the same path. Prompt assembly, request middleware, request logging, result processing, tools, post-step listeners, and cleanup remain ordinary failures.
+
+The failed step closes before recovery runs. A retry opens the next numbered step and rebuilds the request from the durable log; consecutive recovery attempts reset only after a successful provider request. Both DeepSeek adapters normalize recognized provider context-limit failures to `CONTEXT_WINDOW_EXCEEDED`.
+
+If cancellation lands after assistant tool calls are durable but before all calls dispatch, the loop records a synthetic `tool/call` and aborted `tool/result` pair for every undispatched call before following the normal abort path. The surface therefore never retains orphaned durable tool calls merely because cancellation won the race.
+
+### CompactService exposes intent, not token accounting
+
+`CompactService.compactIfNeeded(agent, trigger, signal)` accepts `trigger: 'pressure' | 'context-overflow'`. The interface gains no estimation methods or token types; `ctx.tokenMeter` remains the reusable accounting owner.
+
+For `pressure`, compact-basic applies the service-wide threshold and retained-tail policy to one unified `ctx.tokenMeter.measure()` result. The same singleton meter owns range pricing, provenance, shadowed token counts, and non-shrinking-summary rejection. The common defaults remain threshold ratio `0.8`, retained history `floor(contextWindow × 0.16)`, summarization provider/model `''`, `maxTokens: 8192`, `compactionRetries: 1`, and `auto: true`.
+
+For canonical overflow, compact-basic bypasses scalar pressure and the normal retained-token budget. It chooses the maximal tool-balanced head range while leaving the newest indivisible unit, then attempts exactly one shrinking compaction under the same signal. The automatic listener snapshots `session.surface.replaceGeneration` and returns `{ action: 'retry' }` only when compaction succeeds and the generation increases. A backend returning a result without replacement cannot authorize retry.
+
+`maxOverflowRetries` is optional and defaults to `1`; `0` disables overflow recovery without disabling pressure. `auto: false` registers neither automatic listener. Noncanonical errors, exhausted attempts, an already-aborted signal, a missing routed model, no safe range, no generation change, and recovery throws all delegate to the next listener. With no later recovery, the loop reports the original provider error object and code. Cancellation or disposal remains authoritative even if recovery work completes concurrently.
+
+The default summarizer resolves explicit configuration, then the latest logged route, then agent options. Because direct `llm/stream` middleware may reroute that auxiliary call, `compact/summary.{provider, model}` records the final mutable `GenerateOptions` target observed after dispatch rather than the pre-waterfall candidate.
+
+## Testing
+
+Unit tests cover final-adapter failure provenance and identity, closed-step retry numbering and reset, cancellation and disposal, post-step ordering, routed-envelope pressure, balanced overflow reduction, generation proof, caps, delegation, and auxiliary-call routing. Real-loop tests cover thrown and in-band overflow through compaction to a reconstructed retry request.
+
+## Alternatives considered
+
+- **Keep provisional pre-step pressure and add more arguments** — rejected because later routing and request mutation remain outside any earlier snapshot, while generic lifecycle becomes coupled to one plugin.
+- **Retry the same numbered step** — rejected because recovery appends durable events after the failed boundary. A new step preserves balanced nesting and reconstructability.
+- **Retry whenever `compactIfNeeded` returns a result** — rejected because a custom backend can report success without changing model-visible state. `replaceGeneration` is the authoritative proof.
+- **Let compact-basic parse provider wording** — rejected because classification belongs at adapters and must cover both thrown and in-band delivery.
+- **Fall back to `AgentOptions.model` when no durable route exists** — rejected because automatic policy must describe a completed logged request. Headerless pressure and recovery delegate unchanged.
+
+## Consequences
+
+Post-step pressure describes the completed routed request, including durable tool results and request-only prefix fields. Canonical overflow supplies the backstop when no successful usage anchor exists. Recovery is bounded, cancellation-owned, and monotonic: it retries only after a visible surface generation change.
+
+The cost is one additional serial checkpoint on successful steps and adapter-maintained overflow classification. Provider wording and heuristic character density remain maintenance risks. Surface compaction still cannot repair an envelope that alone exceeds the window or split one indivisible oversized message/tool unit.
+
+This Agent Note supersedes only the pre-step automatic-trigger portion of the [compaction capability-seam Agent Note](../feature/2026-06-18-compaction-capability-seam.md). The service split, standalone token meter, balanced range contract, log-recorded lock, summary replacement, and sole `summarize()` subclass hook remain unchanged.

.agents/notes/implemented/architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.zh.md

@@ -0,0 +1,61 @@
+# Agent Note：调用后压缩压力与上下文溢出恢复
+
+Status: implemented
+
+[English](2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md) | 中文
+
+## 问题
+
+`agent/pre-step` 运行在最终请求路由之前，也早于 assistant 输出、工具结果、缓冲上下文与 steering 的产生。即使它接收已装配提示词与会话前缀，压力视图仍是临时的，因为 `agent/request` 还可以改变路由或调用配置，工具 schema 也没有与这些输入一同冻结。增加字段无法让调用前状态描述已完成调用，还会把通用 seam 与压缩耦合。
+
+成功调用也不是唯一的压力信号。提供方可能在返回 usage 之前就因上下文窗口超限拒绝请求，一些成功调用也不提供 usage。因此，系统需要可回放的调用后压力，以及一条狭窄的失败恢复路径；当压缩无法证明取得有效进展时，必须保留原始提供方错误。
+
+## 决策
+
+### 成功压力移动到持久 post-step 检查点
+
+`agent/pre-step` 收窄为 `(agent, turn, step, signal)`。它仍是 `step/start` 之前的通用串行检查点，但不再携带压缩专用的提示词或前缀字段。
+
+循环在 assistant 输出、所有已分发或合成的工具结果、工具后上下文与 steering 都持久化之后、`step/end` 之前，触发等待式串行 `agent/post-step(agent, turn, step, signal)`。该位置让压力策略看到完整的成功调用状态，同时不会拆开 assistant 工具调用与其结果。监听器失败属于普通 turn 失败，绝不会进入模型请求恢复。
+
+`dsh-compact-basic` 从持久请求头读取精确的最新实际路由模型，只用它确认已经存在完整路由，随后让单例 `ctx.tokenMeter` 计量规范日志信封与当前表层。自动压力不会回退到 `AgentOptions.model`。没有请求头的会话尚无已完成路由请求可供判断，因此不执行工作；任意持久记录的非空模型名都使用同一个估算器。操作性的计量或摘要失败会发出警告，并继续使用完整历史。
+
+### 请求恢复只覆盖最终模型边界
+
+`RequestError`、`RequestErrorDecision` 与 `agent/request-error` waterfall 表示最终适配器已经选定之后的失败。每个返回的流句柄都绑定一个私有失败集合；该集合在分发、异步迭代器构造与迭代过程中保留原始抛出错误的身份，同时防止把嵌套调用的错误来源误归到外层调用。终止性的带内 `error` 或 `aborted` finish 进入同一路径。提示词装配、请求中间件、请求日志、结果处理、工具、post-step 监听器与清理仍属于普通失败。
+
+恢复运行前，失败 step 已经关闭。重试会打开下一个编号 step，并从持久日志重建请求；连续恢复尝试计数只在提供方请求成功后重置。两个 DeepSeek 适配器都把识别出的提供方上下文限制错误规范化为 `CONTEXT_WINDOW_EXCEEDED`。
+
+如果取消发生在 assistant 工具调用已经持久化之后、所有调用完成分发之前，循环会为每个尚未分发的调用记录一对合成的 `tool/call` 与 aborted `tool/result`，随后进入正常中止路径。因此，表层不会仅因取消赢得竞态而留下孤立的持久工具调用。
+
+### CompactService 暴露意图，而不拥有 token 核算
+
+`CompactService.compactIfNeeded(agent, trigger, signal)` 接收 `trigger: 'pressure' | 'context-overflow'`。接口不增加估算方法或 token 类型；`ctx.tokenMeter` 继续作为可复用的核算所有者。
+
+对于 `pressure`，compact-basic 把服务级阈值与保留尾部策略应用到一次统一的 `ctx.tokenMeter.measure()` 结果。范围定价、来源、被遮蔽 token 数与非缩小摘要拒绝也由同一个单例 meter 完成。通用默认值保持为阈值比例 `0.8`、保留历史 `floor(contextWindow × 0.16)`、摘要提供方/模型 `''`、`maxTokens: 8192`、`compactionRetries: 1` 与 `auto: true`。
+
+对于规范化溢出，compact-basic 绕过标量压力与普通保留 token 预算。它在保留最新不可分割单元的同时，选择最大的工具配对平衡头部范围，并在同一 signal 下只尝试一次缩小压缩。自动监听器先记录 `session.surface.replaceGeneration`，只有压缩成功且 generation 增加时才返回 `{ action: 'retry' }`。后端若只返回结果但没有替换表层，不能授权重试。
+
+`maxOverflowRetries` 可选且默认为 `1`；`0` 只禁用溢出恢复，不会禁用压力检查。`auto: false` 不注册任何自动监听器。非规范化错误、尝试耗尽、已经中止的 signal、缺失路由模型、没有安全范围、generation 未变化，以及恢复抛错都会委托给下一个监听器。若没有后续恢复，循环报告原始提供方错误对象与代码。即使恢复工作并发完成，取消或销毁仍具有最终优先级。
+
+默认摘要器依次解析显式配置、最近记录的路由与 agent options。因为直接 `llm/stream` 中间件可以重新路由该辅助调用，`compact/summary.{provider, model}` 记录分发后最终可变的 `GenerateOptions` 目标，而不是 waterfall 之前的候选值。
+
+## 测试
+
+单元测试覆盖最终适配器失败的来源与身份、已关闭 step 的重试编号与重置、取消与销毁、post-step 顺序、已路由信封压力、平衡溢出缩减、generation 证明、上限、委托与辅助调用路由。真实循环测试覆盖抛出式和带内溢出，并验证压缩后的重试请求从替换表层重建。
+
+## 考虑过的替代方案
+
+- **保留临时 pre-step 压力并增加更多参数**——不予采纳，因为后续路由与请求变换仍在更早快照之外，同时通用生命周期会耦合到单个插件。
+- **重试相同编号的 step**——不予采纳，因为恢复会在失败边界之后追加持久事件。新 step 保持边界配对与可重建性。
+- **只要 `compactIfNeeded` 返回结果就重试**——不予采纳，因为自定义后端可能报告成功却没有改变模型可见状态。`replaceGeneration` 才是权威证明。
+- **让 compact-basic 解析提供方措辞**——不予采纳，因为分类属于适配器，而且必须同时覆盖抛出式与带内交付。
+- **没有持久路由时回退到 `AgentOptions.model`**——不予采纳，因为自动策略必须描述已完成且已记录的请求。没有请求头的压力检查与恢复会原样委托。
+
+## 后果
+
+Post-step 压力描述已完成的路由请求，包括持久工具结果与仅请求前缀字段。当成功 usage 锚点不存在时，规范化溢出提供兜底路径。恢复有明确上限、以取消为准，并保持单调：只有模型可见的表层 generation 变化后才重试。
+
+代价是成功 step 增加一个串行检查点，并需要适配器持续维护溢出分类。提供方措辞与启发式字符密度仍是维护风险。表层压缩依然无法修复仅信封本身就超出窗口的情况，也不能拆分单个不可分割的超大消息或工具单元。
+
+本 Agent Note 只取代[压缩能力接缝 Agent Note](../feature/2026-06-18-compaction-capability-seam.md) 中的 pre-step 自动触发部分。服务拆分、独立 token meter、平衡范围契约、日志记录锁、摘要替换与唯一 `summarize()` 子类 hook 均保持不变。

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-10-single-file-executable-sdk-runtime-distribution.md: 0d4686a5a233785ca4832ef068a118b484a872fe
+2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md: dcc9213c6b3a088b8b8bce2a442c5232ed5b7d0b

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.md

@@ -0,0 +1,85 @@
+# Agent Note: Single-file executable SDK runtime distribution (single-exe)
+
+Status: implemented
+
+English | [中文](2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md)
+
+## Problem
+
+DeepSeek Harness needs a dedicated SDK distribution form for the Python library — no Node installation, runs directly on the target platform: a single-file executable (hereafter "the exe") that exposes a stdio JSON-RPC serving surface (`HarnessSdkServer`, the Python SDK's peer), where the plugins and configuration actually booted are decided entirely by a `cordis.yml` supplied from outside the exe.
+
+- The JSONRPC protocol for talking to the Python SDK is already validated
+- A standardized way for cordis.yml to load every plugin (ESModule) is needed
+- The distribution must carry the Node runtime, and support a locally linked source debugging mode
+
+## Decision
+
+### Packaging route: @yao-pkg/pkg's `--sea` mode
+
+The exe is packaged with the **`--sea` (enhanced SEA) mode** of [@yao-pkg/pkg](https://github.com/yao-pkg/pkg) (the actively maintained fork after vercel/pkg was archived). Relative to Node's native SEA, pkg adds a `/snapshot` VFS and runtime module hooks on top, hands the ESM entry to Node's default ESM loader unchanged, and depends on no ESM→CJS transpilation.
+> Measured (macos-arm64, node24 target, pkg 6.21.0): bare-specifier ESM dynamic import inside the VFS (including top-level await), CJS interop, `node:sqlite`, fail-loud on package names outside the set, and on-disk ESM import outside the VFS all pass; `import.meta.url` comes through unchanged as `file:///snapshot/...`.
+
+`--sea` requires target ≥ node22; the exe uniformly targets node24. One pkg invocation packages exactly one target; multi-platform builds invoke it once per platform.
+
+Terminology reminder: pkg's `/snapshot` VFS has nothing to do with this repo's testing-system "snapshot" (ACP replay expected outputs, `$DSH_SNAPSHOT`); this document says "VFS" for the former.
+
+### The serving surface is a plugin: the two packages ui/jsonrpc + examples/jsonrpc-demo
+
+The deterministic protocol implementation (`server.ts` / `transport.ts`) lands as two packages on the existing `ui/acp` + `examples/acp-demo` pattern — the serving surface is itself a plugin:
+
+- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md) (`@deepseek-ai/dsh-jsonrpc`): the pure protocol plugin; on apply it mounts `HarnessSdkServer` plus a line-delimited JSON-RPC transport on the process stdio, with disposal through `ctx.effect()`. Whether to serve is decided by `cordis.yml`; a yml that does not mount it is a legitimate process that does not serve. Protocol-level exit belongs to the plugin (after answering the `shutdown` request it disposes its own fiber, then `exit(0)`; an HMR-style unload only stops the service without exiting the process).
+- [`packages/examples/jsonrpc-demo`](../../../../packages/examples/jsonrpc-demo/README.md) (`@deepseek-ai/dsh-jsonrpc-demo`): a thin app bin — `installFailLoud` + `loadEnv` + config discovery + `boot()` from [`dsh-app-boot`](../../../../packages/ui/app-boot/src/index.ts), done once boot completes; the server is brought up by the `dsh-jsonrpc` entry in the yml. Its only dependency is app-boot. Process-level exit belongs to the bin (stdin EOF/SIGTERM → dispose then 0, SIGINT → 130).
+
+Config discovery has two channels and fails loudly when both are missing: the `DSH_CORDIS_CONFIG` environment variable first (the SDK client convention), then an argv positional argument; no default path and no built-in fallback whatsoever — "the plugins actually booted are decided by an external cordis.yml" is a hard semantic.
+
+### Plugin resolution: the VFS holds a real package tree, the closure manifest IS the deploy root
+
+Inside the exe's VFS sits a **real package tree in build-artifact form** (each package's `lib/` plus a real `node_modules`); the Loader resolves plugin names through standard dynamic `import()`: bare specifiers resolve upward along `node_modules` from the Loader's position inside the VFS, and land inside the VFS naturally. The closed set needs no allowlist code — the set is whatever the VFS has installed, and importing a name outside the set fails.
+
+The deploy root is [`python/sdk-runtime/package.json`](../../../../python/sdk-runtime/package.json) (`dsh-jsonrpc-agent-pkg`, a pnpm workspace member and a zero-code pure dependency manifest) — the unified source of truth for "which plugins the exe ships" and "what the Python runtime distributes". Adding a plugin to the exe = adding one dependency line to the manifest and repackaging. [`scripts/verify-runtime-closure.ts`](../../../../scripts/verify-runtime-closure.ts) traverses every workspace package covered by that manifest and requires every non-optional workspace peer at the runtime root, reporting the complete referencing-package → missing-peer chain; CI static, pre-push, and the single-exe build run it before packaging. Deploy also packs by each package's `files`, so the shared chunks tsdown splits out must be covered by `files`.
+
+### Build pipeline and artifacts
+
+[`scripts/build-exe-for-python-sdk.ts`](../../../../scripts/build-exe-for-python-sdk.ts): runtime closure verification → `pnpm run build` → (after clearing) `pnpm --filter dsh-jsonrpc-agent-pkg deploy --legacy --prod --config.node-linker=hoisted --config.auto-install-peers=false --config.link-workspace-packages=true` **directly into** `python/sdk-runtime/src/deepseek_harness_runtime/runtime/node/` → inject the pkg configuration (`bin` points at `node_modules/@deepseek-ai/dsh-jsonrpc-demo/lib/bin.js` inside the closure, `assets` is a full glob — dynamic import is invisible to pkg's static analysis, so everything must be packed in explicitly) → one `pkg --sea` per target → the executables `dsh-jsonrpc-agent-pkg-<platform>-<arch>` land in `dist-exe/` and are copied back into the runtime directory. CI treats them as intermediate test inputs and retains their platform wheels. All four deploy flags are grounded in measurement: `--legacy` is the mandatory path with inject-workspace-packages off; hoisted yields a zero-symlink file tree (most stable for the pkg VFS, physically guaranteeing a single cordis instance); disabling automatic peer installation keeps unpublished package names from triggering registry resolution; link-workspace-packages points the closure at workspace/vendor sources.
+
+CI: [`.github/workflows/build-exe-for-python-sdk.yml`](../../../../.github/workflows/build-exe-for-python-sdk.yml), triggered explicitly only — `workflow_dispatch`, or the `build-exe` label on a pull request; native builds on the three platforms linux-x64 / linux-arm64 (`ubuntu-24.04-arm`) / macos-arm64, with `~/.pkg-cache` cached; macOS ad-hoc signing is handled by pkg. Each leg drives a mock SSE model through the SDK with the default config and a custom `cordis.yml`, drives the exe directly over NDJSON JSON-RPC, verifies the JSONL and final response, and installs release-shaped wheels into a clean venv without `runtime_bin`; Linux additionally inspects GLIBC requirements and runs in a manylinux 2.28 container. A full three-target run retains four artifacts, each containing one release file: the platform-independent SDK wheel and three native runtime wheels; a subset dispatch retains the SDK wheel and selected runtime wheels. Bare executables and source bundles remain intermediate test inputs. [`.gitlab-ci.yml`](../../../../.gitlab-ci.yml) accepts only `python-vX.Y.Z` tag pipelines whose version matches the root `package.json`, builds one SDK wheel and three native runtime wheels, then a single serialized job checks and publishes all four to the project PyPI registry. Windows is a non-goal.
+
+### Python SDK distribution: two carriers, exe for production, node for development
+
+The Python SDK lives at [`python/`](../../../../python/README.md): `python/sdk` (the client) + `python/sdk-runtime` (the runtime carrier package). The runtime package's data directory holds three kinds of content: the checked-in default `runtime/cordis.yml`, the build-injected platform exe, and the build-injected `runtime/node/` closure tree. `resolve_bundled_launch_args()` automatic resolution **finds the exe only**; the node carrier is enabled only by an explicit `DSH_RUNTIME_MODE=node` (running `runtime/node/node_modules/@deepseek-ai/dsh-jsonrpc-demo/lib/bin.js`, requiring a system node ≥22.19), positioned as the development-verification channel for members of this repo, and does not enter wheel distributions.
+
+[`scripts/build-python-release.py`](../../../../scripts/build-python-release.py) reads the authoritative stable `X.Y.Z` from the repository root `package.json` and stages both packages at that version, with the SDK depending exactly on `deepseek-harness-runtime-bin==X.Y.Z`. An optional `python-vX.Y.Z` release tag is a consistency assertion and is rejected when it differs from the repository version; the source `pyproject.toml` development sentinel never determines a release version. The SDK is a `py3-none-any` wheel; the wheel-only runtime package contains exactly one exe and uses one of `py3-none-manylinux_2_28_x86_64`, `py3-none-manylinux_2_28_aarch64`, or `py3-none-macosx_11_0_arm64`. Its Hatch hook rejects sdists, universal tags, mixed executable payloads, and unsupported platforms.
+
+The exe's "must be explicitly configured" hard semantic is unchanged; the zero-config experience is restored by the wrapper: when the caller gave no `cordis`, named no explicit runtime, and the environment has no `DSH_CORDIS_CONFIG`, the client explicitly injects the checked-in default `cordis.yml` (agent-core + preloaded llm-deepseek + JSONL persistence + bash-local + the `dsh-jsonrpc` serving entry, with `!!js` environment-variable fallbacks) via `DSH_CORDIS_CONFIG`.
+
+### Naming lineage
+
+`@deepseek-ai/dsh-jsonrpc-demo` (the package) → `dsh-jsonrpc-agent` (the bin) → `dsh-jsonrpc-agent-pkg` (the closure manifest; no scope prefix, deliberately sidestepping the constraints' package-shape rules for `@deepseek-ai/dsh-*`) → `dsh-jsonrpc-agent-pkg-<platform>-<arch>` (the exe artifacts). The wire `serverInfo.name` stays `deepseek-harness-sdk-runtime` (a protocol-stable value); the Python dist names are `deepseek-harness` / `deepseek-harness-runtime-bin`.
+
+## Disposition of worker-style plugins
+
+`dsh-workflow-workerthread` and `dsh-code-runtime-worker` are supported inside the exe. Their built hosts convert the sibling `lib/worker.cjs` URL with `fileURLToPath()` and pass the resulting filesystem string to `Worker`, which is the form pkg's Worker hook resolves inside the VFS. The worker entries are CommonJS because that hook compiles VFS worker files as CommonJS. The workflow engine keeps its data-URL bootstrap for unbuilt source execution; only its built sibling entry uses the filesystem string. The custom-config executable smoke loads both backends, invokes a real `run_code` call and a zero-agent `workflow` call, and requires each worker to return `42` from inside pkg's VFS.
+
+## Testing
+
+The verification surface has three tiers. Mechanism tier: the measured conclusions for the `--sea` chain are embedded in the Decision sections (ESM dynamic import inside the VFS, single cordis instance, fail-loud config chain, `node:sqlite`, macOS ad-hoc signing runs). SDK tier: the complete keyless pytest suite covers the client protocol against a fake runtime peer, subprocess cleanup, absolute cwd propagation, dual-carrier launch, and carrier resolution; root CI runs it on Python 3.10. End-to-end tier: every platform build completes a turn against a mock endpoint through the default SDK path, a custom config, and the direct binary protocol, with final text and JSONL checked. The custom config additionally drives `run_code` and a zero-agent `workflow` through their real worker files inside the packaged VFS. The same build leg runs a committed executable-specific snapshot through the Python SDK: a keyless scripted model mounts a Cordis plugin that registers a tool, invokes that tool from `run_code`, runs a direct spawn subagent and a workflow that starts a second spawn child, then unmounts the plugin. The comparison normalizes the SDK result and notification stream plus the parent and two child JSONL logs. This harness stays separate from ACP's `pnpm run test:snapshot` because the protocols and build artifacts differ. The platform wheel is then installed in a clean venv and run without `runtime_bin`.
+
+Manual-driving caveat: the bin treats stdin EOF as "the client is gone" and disposes immediately, so a short-lived pipe aborts an in-flight turn — pipe-driven runs must keep stdin open until the turn ends.
+
+## Alternatives considered
+
+**Bare Node native SEA.** The injected main script must be a single CJS file, and the blob carries no filesystem and no module resolution, so a dynamic import of a bare specifier has nothing to resolve against; the only option is compiling plugins statically into the main script and registering them by hand — bypassing standard module resolution and hardcoding the plugin set, contrary to "configuration decides everything". The final route is in fact "the official SEA foundation + pkg's VFS/module-hook layer"; what was rejected is the bare use, not SEA itself.
+
+**pkg standard mode.** Killed by the PoC, not a trade-off: it turns ESM into CJS + V8 bytecode via esbuild, the runtime vm compilation wires up no dynamic-import callback, every `import()` throws `ERR_VM_DYNAMIC_IMPORT_CALLBACK_MISSING`, and `--options experimental-require-module` has no effect; it also depends on community-patched Node binaries (no macos-arm64 prebuilt; compiling from source on the spot takes about 10 minutes). Zero viability for this repo's architecture.
+
+**Pre-bundling each package ESM→CJS into the VFS.** The compromise that keeps real resolution semantics and only downgrades the module format; `--sea` passed measurement outright, so this layer of build complexity never needed introducing.
+
+**jsonrpc-agent carrying the full closure dependencies.** The app bin would declare 53+ dependencies it never imports — a "packaging manifest" masquerading as real dependency relationships — and would force constraints to open two exceptions for it, cordis-in-dependencies and a files wildcard. With the closure manifest landing on the python-side manifest package, constraints needs no exception at all and the bin keeps the normal package shape isomorphic to acp-agent.
+
+**An open plugin set (loading user plugins from disk).** This round ships a closed set; the PoC incidentally confirmed that on-disk ESM import outside the VFS works (through the `ctx.baseUrl` relative-path channel). It is listed as a future evolution, which must separately solve sharing the cordis instance inside the exe with external plugins.
+
+## Consequences
+
+**Bought**: zero-dependency single-file distribution on target platforms; plugin semantics strictly identical to running from source (the same real package tree, no transpilation, no registry); the serving surface, the plugin set, and the configuration all converge on two sources of truth — `cordis.yml` plus one dependency manifest; the exe and node carriers share one tree and one semantics, so development verification never waits for packaging; official Node binaries remove the patched-binary supply-chain concern.
+
+**Paid**: artifacts on the order of 174MB with source entering the blob as-is (no bytecode obfuscation; a closed-source distribution requirement needs a separate evaluation); pkg's VFS/module-hook layer remains community-maintained (the build script pins `@yao-pkg/pkg@6.21.0`; upgrading is an explicit change); `--sea` is one invocation per target (matching CI's one leg per platform; local multi-platform builds are serial).

.agents/notes/implemented/architecture/2026-07-10-single-file-executable-sdk-runtime-distribution.zh.md

@@ -0,0 +1,85 @@
+# Agent Note: 单文件可执行的 SDK 运行时分发（single-exe）
+
+Status: implemented
+
+[English](2026-07-10-single-file-executable-sdk-runtime-distribution.md) | 中文
+
+## 问题
+
+DeepSeek Harness 需要为 Python 库专门提供一种无需安装 Node、可直接在目标平台运行的 SDK 分发形态：一个单文件可执行程序（下称 exe），通过 stdio 提供 JSON-RPC 对外服务接口（`HarnessSdkServer`，Python SDK 的对端），且实际启动的插件与配置完全由 exe 外部输入的 `cordis.yml` 决定。
+
+- 与 Python SDK 通信的 JSON-RPC 协议已经过验证
+- 需要提供通过标准化 `cordis.yml` 加载所有插件（ES 模块）的能力
+- 分发物要自带 Node 运行时，并支持本地源码链接的调试模式
+
+## 决策
+
+### 打包路线：@yao-pkg/pkg 的 `--sea` 模式
+
+exe 使用 [@yao-pkg/pkg](https://github.com/yao-pkg/pkg)（vercel/pkg 归档后的活跃维护 fork）的 **`--sea`（enhanced SEA）模式**打包。相比 Node 原生 SEA，pkg 在其上增加 `/snapshot` 虚拟文件系统（VFS）与运行时模块钩子，将 ESM 入口原样交给 Node 默认的 ESM loader，不依赖任何 ESM→CJS 转译。
+> 实测（macos-arm64、node24 构建目标、pkg 6.21.0）：VFS 内裸包名 ESM 动态 `import()`（含顶层 `await`）、CJS 互操作、`node:sqlite`、集合外包名明确报错、VFS 外磁盘 ESM `import()` 全部通过，`import.meta.url` 原样为 `file:///snapshot/...`。
+
+`--sea` 要求构建目标 ≥ node22，exe 统一以 node24 为构建目标；每次 pkg 调用只打包一个构建目标，多平台各调用一次。
+
+术语提醒：pkg 的 `/snapshot` VFS 与本仓库测试体系的“快照”（ACP 回放预期输出、`$DSH_SNAPSHOT`）无关，本文用“VFS”指前者。
+
+### 对外服务接口也是插件：ui/jsonrpc + examples/jsonrpc-demo 两包
+
+确定性协议实现（`server.ts` / `transport.ts`）按 `ui/acp` + `examples/acp-demo` 的既有模式落为两包——对外服务接口本身也是插件：
+
+- [`packages/ui/jsonrpc`](../../../../packages/ui/jsonrpc/README.md)（`@deepseek-ai/dsh-jsonrpc`）：纯协议插件；执行 `apply` 时，在进程 stdio 上挂载 `HarnessSdkServer` 与按行传输的 JSON-RPC 层，资源释放走 `ctx.effect()`。是否提供服务由 `cordis.yml` 决定；未挂载该插件的配置会启动一个不提供此服务的合法进程。协议级退出归插件所有（应答 `shutdown` 请求后 dispose 自身 fiber，再调用 `exit(0)`；HMR 式卸载只停止服务，不退出进程）。
+- [`packages/examples/jsonrpc-demo`](../../../../packages/examples/jsonrpc-demo/README.md)（`@deepseek-ai/dsh-jsonrpc-demo`）：轻量应用入口——`installFailLoud` + `loadEnv` + 配置发现 + [`dsh-app-boot`](../../../../packages/ui/app-boot/src/index.ts) 的 `boot()`；`boot()` 完成后入口即完成，服务器由 `cordis.yml` 中的 `dsh-jsonrpc` 条目启动。它只依赖 `app-boot`。进程级退出归 `bin` 所有（stdin EOF/SIGTERM → dispose 后返回 0，SIGINT → 130）。
+
+配置发现有两个通道，均缺失时立即报错：优先使用 `DSH_CORDIS_CONFIG` 环境变量（SDK 客户端约定），其次使用 argv 位置参数；没有默认路径或内置回退——“实际启动的插件由外部 `cordis.yml` 决定”是硬语义。
+
+### 插件解析：VFS 装载真实包树，闭包清单就是部署根目录
+
+exe 的 VFS 内是**构建产物形态的真实包树**（各包的 `lib/` + 真实 `node_modules`）。loader 通过标准动态 `import()` 解析插件名：裸包名从 VFS 内 loader 所在位置沿 `node_modules` 向上解析，自然落在 VFS 内。封闭集不需要白名单代码——VFS 中安装了什么，集合中就有什么；`import()` 集合外的名称会失败。
+
+部署根目录是 [`python/sdk-runtime/package.json`](../../../../python/sdk-runtime/package.json)（`dsh-jsonrpc-agent-pkg`，pnpm 工作区成员、零代码纯依赖清单），也是“exe 安装哪些插件”与“Python 运行时分发什么”的统一事实源。向 exe 添加插件，就是在清单中增加一行依赖后重新打包。[`scripts/verify-runtime-closure.ts`](../../../../scripts/verify-runtime-closure.ts) 遍历该清单覆盖的全部工作区包，要求每个非可选的工作区对等依赖（peer dependency）都显式列在运行时根目录，并报告“引用包 → 缺失对等依赖”的完整链路；CI 静态检查、pre-push 与 single-exe 构建都会在打包前运行该门禁。部署还会依据各包的 `files` 字段打包，因此 tsdown 拆出的共享分片必须被 `files` 覆盖。
+
+### 构建管线与产物
+
+[`scripts/build-exe-for-python-sdk.ts`](../../../../scripts/build-exe-for-python-sdk.ts)：运行时闭包校验 → `pnpm run build` →（清空后）`pnpm --filter dsh-jsonrpc-agent-pkg deploy --legacy --prod --config.node-linker=hoisted --config.auto-install-peers=false --config.link-workspace-packages=true` **直接写入** `python/sdk-runtime/src/deepseek_harness_runtime/runtime/node/` → 注入 pkg 配置（`bin` 指向闭包内的 `node_modules/@deepseek-ai/dsh-jsonrpc-demo/lib/bin.js`；`assets` 使用全量 glob，因为动态 `import()` 对 pkg 静态分析不可见，必须显式打入全部内容）→ 每个构建目标调用一次 `pkg --sea` → 可执行文件 `dsh-jsonrpc-agent-pkg-<platform>-<arch>` 写入 `dist-exe/`，并拷回运行时目录。CI 将这些文件作为测试中间输入，只保留对应平台的 wheel 包。四个部署标志都有实测依据：未启用 `inject-workspace-packages` 时必须使用 `--legacy`；`hoisted` 产出无符号链接的文件树（对 pkg VFS 最稳定，并从物理上保证只有一个 Cordis 实例）；关闭对等依赖自动安装可避免未发布包名触发注册表解析；`link-workspace-packages` 让闭包指向工作区/vendor 源码。
+
+CI 使用 [`.github/workflows/build-exe-for-python-sdk.yml`](../../../../.github/workflows/build-exe-for-python-sdk.yml)，且只允许显式触发：手动派发 `workflow_dispatch`，或给 PR 添加 `build-exe` 标签。linux-x64、linux-arm64（`ubuntu-24.04-arm`）和 macos-arm64 三个平台分别进行原生构建，并缓存 `~/.pkg-cache`；macOS 的 ad-hoc 签名由 pkg 处理。每个平台都使用模拟 SSE 模型，分别通过默认配置和自定义 `cordis.yml` 驱动 SDK，再通过 NDJSON JSON-RPC 直接驱动 exe，校验 JSONL 与最终响应；最后把发布形态的 wheel 包安装到干净的 venv 中，并在不传 `runtime_bin` 的情况下运行。Linux 还会检查 GLIBC 依赖，并在 manylinux 2.28 容器中运行。完整构建三个目标时保留 4 个产物，每个产物只含一个发布文件：平台无关的 SDK wheel 包与 3 个原生运行时 wheel 包；手动选择部分目标时保留 SDK wheel 与所选运行时 wheel。裸 exe 与源码包只作为测试中间输入。[`.gitlab-ci.yml`](../../../../.gitlab-ci.yml) 只接受版本与根目录 `package.json` 匹配的 `python-vX.Y.Z` 标签流水线，构建一个 SDK wheel 包和 3 个原生运行时 wheel 包，再由单个串行任务校验并将这 4 个文件发布到项目的 PyPI 注册表。Windows 不在目标范围内。
+
+### Python SDK 分发：双载体，exe 用于生产，`node` 用于开发
+
+Python SDK 位于 [`python/`](../../../../python/README.md)：`python/sdk` 是客户端，`python/sdk-runtime` 是运行时载体包。运行时包的数据目录包含三类内容：检入的默认 `runtime/cordis.yml`、构建注入的平台 exe，以及构建注入的 `runtime/node/` 闭包树。`resolve_bundled_launch_args()` 的自动解析**只查找 exe**；`node` 载体仅在显式设置 `DSH_RUNTIME_MODE=node` 时启用（运行 `runtime/node/node_modules/@deepseek-ai/dsh-jsonrpc-demo/lib/bin.js`，需要系统 Node ≥22.19），定位为本仓库成员的开发验证通道，不随 wheel 包分发。
+
+[`scripts/build-python-release.py`](../../../../scripts/build-python-release.py) 从仓库根目录的 `package.json` 读取权威的稳定版本 `X.Y.Z`，以该版本暂存两个包，并让 SDK 精确依赖 `deepseek-harness-runtime-bin==X.Y.Z`。可选的 `python-vX.Y.Z` 发布标签只是一项一致性断言，与仓库版本不同时会被拒绝；源码 `pyproject.toml` 中的开发占位版本从不决定发布版本。SDK 是 `py3-none-any` wheel 包；只提供 wheel 包的运行时包恰好包含一个 exe，标签为 `py3-none-manylinux_2_28_x86_64`、`py3-none-manylinux_2_28_aarch64` 或 `py3-none-macosx_11_0_arm64`。其 Hatch 钩子拒绝 sdist、通用标签、混合可执行载荷以及不支持的平台。
+
+exe“必须显式配置”的硬语义不变；零配置体验由包装层恢复：调用方没有提供 `cordis`、没有显式指定运行时，且环境中没有 `DSH_CORDIS_CONFIG` 时，客户端将检入的默认 `cordis.yml`（`agent-core` + 预载的 `llm-deepseek` + JSONL 持久化 + `bash-local` + `dsh-jsonrpc` 对外服务条目，并通过 `!!js` 使用环境变量兜底）显式注入 `DSH_CORDIS_CONFIG`。
+
+### 命名血统
+
+`@deepseek-ai/dsh-jsonrpc-demo`（包）→ `dsh-jsonrpc-agent`（`bin`）→ `dsh-jsonrpc-agent-pkg`（闭包清单；没有作用域前缀，刻意避开 `constraints` 对 `@deepseek-ai/dsh-*` 的包形状规则）→ `dsh-jsonrpc-agent-pkg-<platform>-<arch>`（exe 产物）。协议字段 `serverInfo.name` 保持为 `deepseek-harness-sdk-runtime`（协议稳定值）；Python 分发名为 `deepseek-harness` / `deepseek-harness-runtime-bin`。
+
+## 工作线程插件
+
+exe 内支持 `dsh-workflow-workerthread` 与 `dsh-code-runtime-worker`。两个后端构建后的宿主都通过 `fileURLToPath()` 转换相邻 `lib/worker.cjs` 的 URL，再将所得文件系统字符串传给 `Worker`；pkg 的 Worker 钩子可以用这种形式解析 VFS 内文件。该钩子会把 VFS 内的工作线程文件作为 CommonJS 编译，所以工作线程入口采用 CommonJS。工作流引擎在未构建的源码执行中仍保留 `data:` URL 引导程序，只有构建后的相邻入口使用文件系统字符串。自定义配置的可执行文件冒烟测试会加载两个后端，实际调用 `run_code` 与不启动 agent 的 `workflow`，并要求两个工作线程都从 pkg 的 VFS 内返回 `42`。
+
+## 测试
+
+验证面分三层。机制层：`--sea` 链路的实测结论内嵌在“决策”各节（VFS 内 ESM 动态 `import()`、单一 Cordis 实例、明确报错的配置链路、`node:sqlite`、macOS ad-hoc 签名可运行）。SDK 层：完整的无密钥 pytest 套件以假运行时对端覆盖客户端协议、子进程清理、绝对 `cwd` 传递、双载体启动与载体解析；根 CI 在 Python 3.10 上运行全部用例。端到端层：每个平台构建都通过默认 SDK 路径、自定义配置和直接二进制协议，对模拟端点完成一个轮次，并校验最终文本与 JSONL。自定义配置还会通过打包进 VFS 的真实工作线程文件执行 `run_code` 和不启动 agent 的 `workflow`。同一构建任务还会经 Python SDK 运行一组检入的 exe 专用快照：无密钥脚本化模型挂载一个会注册工具的 Cordis 插件，从 `run_code` 调用该工具，运行一个由 spawn 提供方直接启动的 subagent（子 agent）和一个会通过 spawn 启动第二个子 agent 的工作流，随后卸载该插件。比较时会规范化 SDK 结果与通知流，以及父会话和两个子会话的 JSONL 日志。该 harness 与 ACP 的 `pnpm run test:snapshot` 保持独立，因为二者的协议和构建产物不同。随后把平台 wheel 包安装进干净的 venv，并在不传 `runtime_bin` 的情况下运行。
+
+手工驱动注意：`bin` 将 stdin EOF 视为“客户端已离开”并立即 dispose，短命管道会中止进行中的轮次——管道驱动必须保持 stdin 打开，直到轮次结束。
+
+## 曾考虑的替代方案
+
+**裸用 Node 原生 SEA。** 注入的主脚本必须是 CJS 单文件，blob 内没有文件系统与模块解析，因此动态 `import()` 无法解析裸包名；只能把插件静态编译进主脚本并手工注册。这会绕过标准模块解析并硬编码插件集合，与“配置决定一切”相悖。最终路线实际是“官方 SEA 基础 + pkg 的 VFS/模块钩子层”；否决的是裸用方式，而不是 SEA 本身。
+
+**pkg 标准模式。** PoC 证明该模式不可行，而非权衡后放弃：它通过 esbuild 将 ESM 转为 CJS + V8 字节码，但运行时 VM 编译没有接入动态 `import()` 回调，任何 `import()` 都会抛出 `ERR_VM_DYNAMIC_IMPORT_CALLBACK_MISSING`，`--options experimental-require-module` 也无效；此外，它依赖社区补丁版 Node 二进制（macos-arm64 没有预编译版本，现场从源码编译约需 10 分钟）。该模式不适用于本仓库架构。
+
+**每包 ESM→CJS 预打包进 VFS。** 保持真实解析语义、只降级模块格式的折中；`--sea` 直接通过实测，这层构建复杂度无需引入。
+
+**让 jsonrpc-agent 承担完整闭包依赖。** 应用入口将声明 53 个以上自身并不 `import()` 的依赖，使“打包清单”伪装成真实依赖关系，还会迫使 `constraints` 为其增加 `cordis-in-dependencies` 与 `files` 通配符两个例外。将闭包清单放在 Python 侧的清单包后，`constraints` 不需要任何例外，`bin` 也能保持与 acp-agent 同构的正常包形状。
+
+**开放插件集（从磁盘加载用户插件）。** 本期采用封闭集；PoC 同时证实，可以通过 `ctx.baseUrl` 相对路径通道从 VFS 外的磁盘 `import()` ESM。该能力列为后续演进，届时还需解决外部插件与 exe 内 Cordis 实例的共享问题。
+
+## 后果
+
+**买到的**：目标平台零依赖的单文件分发；插件语义与源码运行严格一致（同一棵真实包树，无转译、无注册表）；对外服务接口、插件集与配置全部收敛到 `cordis.yml` 和一份依赖清单这两个事实源；exe 与 `node` 双载体使用同一棵树和相同语义，开发验证无需等待打包；官方 Node 二进制消除了补丁版二进制的供应链顾虑。
+
+**付出的**：产物约 174MB，且源码原样进入 blob（没有字节码混淆；闭源分发诉求需要另行评估）；pkg 的 VFS/模块钩子层仍由社区维护（构建脚本钉死 `@yao-pkg/pkg@6.21.0`，升级需要显式改动）；`--sea` 每个构建目标调用一次（与 CI 每个平台一个任务相匹配，本地多平台构建串行执行）。

.agents/notes/implemented/architecture/2026-07-12-agent-scope-runtime-design.md

@@ -0,0 +1,390 @@
+# Agent Note: Agent-scope runtime design and correctness
+
+Status: implemented
+
+## Problem
+
+The [agent-scope contract](2026-07-08-agent-scope-contexts.md) is simple for contributors: register through `agent.ctx`, resolve one global-plus-agent view, publish only after setup, and retain the scope until work stops. The runtime must preserve that contract across a cooperative plugin framework, asynchronous creation, reentrant listeners, durable session commits, and worker or process failure.
+
+The main design risk is adding a second mechanism for every race. Separate reservations, readiness sentinels, cancellation relays, snapshot layers, and protection registries can mirror the same fact until no reader can tell which one is authoritative. That machinery also encourages the runtime to treat trusted typed calls as hostile serialization boundaries.
+
+The implementation needs enough state to preserve real ownership and settlement boundaries, but no more. A correctness reviewer must be able to follow one fact from acceptance through publication and teardown without reconciling parallel representations.
+
+## Decision
+
+The runtime uses one mechanism per independent fact. Scope routing has an opaque carrier; each live registry object has one entry record; each create or resume operation has one transaction; typed same-process calls borrow readonly values; real data boundaries materialize once; the cooperative prompt-assembly result is authoritative; and worker/process code retains separate terminal and quiescence state only where different owners can genuinely race.
+
+The design can be skimmed as seven choices:
+
+| Problem | Authoritative mechanism |
+|---|---|
+| Select global plus one agent's registrations | Opaque scope key and routing carrier |
+| Own one live agent or session | One registry entry captured by its disposer |
+| Coordinate create/resume | One `AgentCreationTransaction` |
+| Protect durable, queued, model, or wire data | Materialize once at that boundary |
+| Pass typed values inside one process | Readonly borrowed contract |
+| Compose the model-visible prompt and tool surface | One shared tool view plus the authoritative assembly-waterfall result |
+| Coordinate subagent, worker, and process shutdown | One cancellation signal plus the independent terminal/quiescence facts of that boundary |
+
+The rest of this Agent Note expands those choices in dependency order: Cordis mechanics, scope routing, creation and session commit, tools and prompts, subagents and workflows, then executable checks.
+
+The [July 8 Agent Note](2026-07-08-agent-scope-contexts.md) remains the contributor contract. The separate [subagent composition-controls Agent Note](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns `persona`, `toolFilter`, and `maxDepth`; this document discusses only how their setup fits the lifecycle.
+
+## Cordis model: context, fiber, effect, receiver, and waterfall
+
+Five Cordis ideas are required to understand the implementation. A context selects services and registration ownership; a fiber is one live plugin or child lifecycle; an effect attaches cleanup to a fiber; an event receiver selects listeners; and a waterfall lets listeners transform or veto an operation in sequence.
+
+### A context is an ownership path through one service graph
+
+All agents share one Cordis service graph. A derived context does not clone `ToolRegistry`, `SystemPrompt`, persistence, or model adapters; it changes how registrations made through that context are tagged and which effects own their cleanup.
+
+`agent.ctx` is such a derived context. Service calls still reach the shared instances, while a registration can inspect its calling context and store a contribution under the nearest scope key. Ordinary plugin contexts carry no scope key and therefore register globally.
+
+### Fibers and effects make cleanup structural
+
+A Cordis fiber is the live instance created when a plugin or child context is activated. Its state records whether that lifecycle is active, unloading, failed, or disposed. `ctx.effect()` and `ctx.on()` return disposers and also attach those disposers to the registering fiber, so unloading a plugin or agent scope removes everything registered through that context without a separate inventory.
+
+The vendored Cordis fiber implementation establishes ownership before arbitrary setup or `internal/plugin` observers run. A reentrant unload can see the child fiber or effect that has started, reject effects added after unload begins, and join cleanup already started through a public single-shot disposer. Teardown observers are contained individually so one callback cannot prevent structural cleanup.
+
+These are framework lifecycle guarantees rather than agent-specific policy. Agent creation depends on them because setup can activate arbitrary plugins and synchronously reenter owner disposal.
+
+### Receivers route listeners; waterfalls compose decisions
+
+Cordis filters listeners using the dispatch receiver (`this`), while harness listeners need an explicit agent, execution, request, or other subject. `Scoped<T>` marks the receiver expected by a scoped event declaration, but the runtime carrier deliberately exposes no subject API.
+
+Product helpers therefore construct the carrier and pass the domain subject separately. This prevents listener routing from becoming an alternate object model and keeps event signatures understandable without knowledge of carrier internals.
+
+A Cordis waterfall is middleware-style dispatch. Each listener receives `next()`: calling it delegates to the remaining listeners and base operation, while returning without it vetoes or replaces the downstream result. Waterfalls power prompt assembly and tool policy; ordinary emit events notify synchronously, and parallel events await all listeners without a veto result.
+
+## Scope routing: one opaque key selects one layer
+
+The scope package implements the smallest object needed for Cordis routing. Its carrier holds only a composed service filter and scope predicate, while the package records the opaque key privately and exposes the scope fiber's quiescent disposer separately.
+
+### Scope identity uses object identity
+
+A `ScopeKey` is an opaque object compared by identity. The harness uses the live `Agent` as its own key, but the primitive is domain-neutral and supports other scoped owners.
+
+`createScope(parent, key)` returns a scope whose `ctx` shares the parent's services and whose effects are tagged with that key. `scopeOf(ctx)` reads the nearest registration key. `scopeTarget(base, key)` creates the event receiver whose filter preserves the base receiver's Cordis service filter, then admits unscoped listeners and listeners with that exact key.
+
+The receiver is a small carrier rather than a transparent proxy for the domain object. Code that needs the agent receives the explicit event argument; code that needs registration ownership receives `agent.ctx`.
+
+### Registry reads overlay one exact map
+
+Scope-aware registries store global contributions separately from identity-keyed local contributions. A read resolves the global layer and at most one local layer; it never traverses parentage.
+
+Each service retains its domain rule. Named prompt values and tools use local shadowing, tool restrictions filter globals before local tools are added, and events select listener audiences rather than registered data. Scope supplies identity and ownership, not a universal merge algorithm.
+
+### Fused dispatch helpers prevent subject drift
+
+`agentEvents(context, agent)` constructs the agent's carrier and injects the same agent as the event subject. Session, tool, approval, prompt, and subagent services likewise derive routing from the object they already own instead of accepting an unrelated key.
+
+The type marker rejects ordinary bare-receiver mistakes, and development invariants cover direct JavaScript or casted dispatch. The subject remains explicit because routing correctness and useful event data are different concerns.
+
+## Agent creation: one transaction owns the complete operation
+
+Create and resume are one asynchronous lifecycle with several phases, not several lifecycles. `AgentCreationTransaction` owns caller and factory liveness, optional cancellation, private resources, publication, rollback, and the memoized teardown observed by every owner.
+
+### Registry entries are the only live identity records
+
+AgentRegistry and SessionStore each keep one entry per live object. The entry holds the stable ID, object, scoped carrier, and the small amount of publication or append state that belongs to that object.
+
+A detach closure captures its exact entry. It deletes only when the map still points to that entry, so an old disposer cannot delete a later object that reuses the same ID. No registry rereads a mutable caller object to decide identity.
+
+There is no reservation API. Caller-supplied IDs are admitted at final entry. Concurrent same-ID operations may both complete private setup; exactly one final `enter()` succeeds, and every loser rolls its private resources back. Sequential reuse is valid after the earlier disposer reaches quiescence.
+
+### The transaction owns preparation before awaiting it
+
+The transaction is installed under both the calling Cordis context and the concrete AgentLoop factory before persistence load or setup can suspend. It also observes an optional create/resume signal until the public operation settles.
+
+Create prepares a new Session. Resume loads and validates the persisted Session before preparing the same live session identity. Both paths then build the scope, agent, and driver and invoke the same setup/publication algorithm.
+
+The factory stores concrete trace targets but invokes them through a caller-bound Cordis trace. This preserves dependency origin and caller ownership without stacking trace proxies.
+
+### Setup is trusted composition inside a private world
+
+Setup receives the full child context and may await plugin activation. It can register tools, prompt sections, restrictions, listeners, and other effects, but the public contract does not support driving or publishing the in-flight agent through casts or internal registry calls.
+
+The transaction races asynchronous load and setup against deactivation rather than waiting forever for a promise owned by external code. If cancellation or owner unload wins, public creation rejects after transaction-owned cleanup even when the external promise never settles.
+
+### Publication has one ordered commit path
+
+Publication admits and announces resources in the order required by observers:
+
+1. Enter the session.
+2. Enter the agent.
+3. Announce `session/created`.
+4. Announce `agent/created`.
+5. Enable public driving.
+6. Emit `agent/session-start`.
+7. Start the driver.
+
+The agent never drives before both registries and creation notifications agree. A synchronous listener may veto or dispose an owner; the transaction records publication in progress and waits for that callback stack to unwind before teardown continues. Every creation announcement that begins has a matching disposal announcement during rollback.
+
+The sequence diagram isolates the non-obvious race: a synchronous creation listener can request disposal while the publication call stack still owns both registry entries. Teardown must deactivate immediately but wait for that stack to unwind before stopping and detaching anything.
+
+```mermaid
+sequenceDiagram
+  participant Tx as AgentCreationTransaction
+  participant Registries
+  participant Listener as Synchronous listener
+  participant Driver
+
+  Tx->>Tx: mark publication in progress
+  Tx->>Registries: announce agent/created
+  Registries->>Listener: invoke inside the same call stack
+  Listener->>Tx: dispose reentrantly
+  Tx->>Tx: deactivate, teardown waits for publication
+  Tx-->>Listener: disposal request accepted
+  Listener-->>Registries: return
+  Registries-->>Tx: announcement unwound
+  Tx->>Tx: resolve publication settlement
+  Tx->>Driver: stop and drain
+  Tx->>Registries: detach agent, then session
+  Tx->>Tx: dispose scope and resolve teardown
+```
+
+### Teardown preserves work before revoking registrations
+
+Every teardown request joins one memoized path. The order is:
+
+1. Deactivate creation or driving and let synchronous publication finish.
+2. Stop and drain the driver, including idle injection flushes.
+3. Detach the agent.
+4. Detach the session.
+5. Dispose the agent scope.
+6. Retire transaction ownership tracking.
+
+This order lets final agent and session events use the matching scoped listeners and keeps persistence observers attached through the final flush. Scope disposal comes last because registration revocation is the externally visible lifetime boundary.
+
+## Session append: materialize, validate, commit, notify
+
+Session events cross a durable boundary, so append owns their data. The rest of the algorithm uses one attached entry and one commit point.
+
+### Durable data is materialized once
+
+Session headers, seeds, and appended events are lossless JSON data. The Session constructor or append path materializes and validates them before storage and exposes frozen snapshots, so later caller mutation cannot change persistence, replay, or model reconstruction.
+
+This is a real ownership boundary: the values leave the caller, may be persisted, and must reconstruct the same request later. It is intentionally stricter than a typed same-process callback or registry definition.
+
+### Pre-commit listeners can veto; post-commit observers cannot
+
+Append follows one sequence:
+
+1. Materialize the durable event and surface intent.
+2. Claim the SessionEntry and reject reentrant append on that entry.
+3. Resolve scoped callbacks and run internal invariant validation.
+4. Push exactly once; this is the commit point.
+5. Notify each observer independently, containing synchronous and asynchronous failures.
+6. Release append state and honor a detach requested during publication.
+
+No observer error makes a committed event look uncommitted, and one bad listener cannot starve later listeners. Session invariants stage their transition before commit and apply it only when the same event reaches the contained post-commit observer.
+
+`flush()` starts every persistence listener and awaits every result before reporting failure. This deliberate all-settled behavior prevents a synchronous failure from starving another backend or final flush.
+
+## Trust boundaries: copy only when ownership actually changes
+
+The runtime distinguishes typed in-process contracts from serialization and durability boundaries. This is the main simplification rule for values and callbacks.
+
+| Boundary | Ownership rule |
+|---|---|
+| Typed service/plugin call in the same process | Borrow readonly values and callbacks |
+| Parsed plugin configuration or external file | Validate semantic and structural input |
+| Queued inbox message | Materialize before asynchronous consumption |
+| Model/tool JSON input or output | Materialize at the model/tool boundary |
+| Durable session or persistence data | Materialize and validate before commit |
+| Worker, process, or wire message | Serialize, validate, and own the decoded value |
+
+Tests that fabricate hostile getters, replace typed callbacks after handoff, or cast fake service objects do not define a production contract by themselves. The runtime keeps checks where data crosses a parser, queue, model, durable, file, worker, process, or wire boundary and relies on readonly types plus plugin discipline inside the trusted process.
+
+Callback containment is separate from data ownership. Listeners are arbitrary extension code and can throw even when their arguments are trusted; publication and post-commit paths still contain failures according to their event contract.
+
+## Tools and prompts: one view, authoritative assembly, committed outcomes
+
+Tool presentation and execution share one private resolver. Prompt assembly remains trusted cooperative composition: registries supply the ordered input, and the assembly waterfall's returned value is exactly what the loop logs and sends. Execution uses separate one-way boundaries only where policy or outcome settlement must be monotonic.
+
+### One resolver defines the tool view
+
+The private resolver applies the current presentation mode, live global restrictions, exact local overlay, and local shadowing. Schemas, lookup, execution, Code Mode SDK generation, and restriction validation all use that resolver or its pre-restriction global-name view.
+
+The [subagent composition-controls Agent Note](../feature/2026-07-12-subagent-persona-tool-filter-and-depth.md#tool-filtering-is-one-live-global-view-rule) owns the user-visible allow/deny semantics. The implementation requirement is agreement: a filtered-away global cannot remain executable through a different lookup path, and a locally shadowed definition is the same definition presented and executed.
+
+`ToolRestriction` accepts readonly allow/deny names and compiles them into internal sets. Multiple restrictions intersect. Public `visible()` and `knownNames()` methods are unnecessary because only the registry needs the intermediate views.
+
+### Tool execution owns identity and boundary materialization
+
+The registry assigns every execution a fresh branded `Symbol` token. Nested Code Mode calls carry the outer token as `parent`, so structured output can correlate an inner capture with its enclosing `run_code` result by identity.
+
+A fresh registry-assigned Symbol provides collision-free execution identity without a WeakSet membership registry. Callers cannot supply the execution's own token through `ToolExecutionInput`; they only receive the pipeline-owned `ToolExecution` after the registry creates it. This is a trusted typed contract, not a runtime defense against arbitrary casts or JavaScript callers.
+
+Arguments are materialized once where model/tool JSON enters the pipeline. Pre-, around-, and post-execute listeners operate on the typed execution and decisions. Call ID correlation, approval, monotonic guards, and Code Mode nesting remain explicit relational checks.
+
+After the last post-execute listener, the registry materializes and freezes the accepted final result once. Every synchronous `tools/result` observer receives that exact committed object, and observer failures are contained individually. An outer pipeline failure is normalized into a committed error result, so observers can discard staged work against the same authoritative boundary.
+
+### The assembly waterfall owns the final model-visible composition
+
+SystemPrompt first resolves the global-plus-agent sections, variables, and tool providers into a deterministic registry contribution. The scope-filtered `system-prompt/assemble` waterfall may then reorder, replace, add, or remove any section, variable, or schema. Its returned assembly is authoritative; there is no later restoration pass and no finality metadata on ordinary prompt sections, tool definitions, or provider results.
+
+This is a trusted same-process extension seam, not an authority boundary. A listener that changes Code Mode's `run_code` schema or `tools:sdk` instructions, or a structured child's capture schema or instruction, owns preserving a coherent protocol in the assembly it returns. ToolRegistry still reserves `run_code` against ordinary tool registration and restriction because those are registry invariants, but assembly middleware remains free to transform the final model-visible surface.
+
+Scope solves the real isolation problem directly. Structured-output contributions register in the child's exact scope, while Code Mode derives its transport and SDK from the same resolved tool view. A second named-protection system would need another ownership and collision rule across arbitrary schema providers—including providers that intentionally contribute duplicate names—without creating a new trust boundary.
+
+### Structured output commits only authoritative outcomes
+
+Structured output combines child-scoped composition with a two-phase execution commit. The child registers its `structured_output` tool and instruction before publication; a trusted assembly listener may transform those ordinary contributions and is responsible for preserving the protocol if the child is expected to complete. The tool body validates a candidate and stages it by the current `ToolExecution`, but successful capture is decided only by immutable `tools/result` observations.
+
+For a native call, the observer deletes the stage and commits its value only when that exact execution's final result succeeds. A post-execute block or outer pipeline failure therefore cannot leave a captured value behind.
+
+For a Code Mode SDK call, the inner successful result records `{ parentToken, value }` rather than committing. The observer waits for the `run_code` execution whose token matches `parentToken` and commits only if that outer final result also succeeds. Program failure, runtime abort, or outer post-policy denial discards the pending value.
+
+Once a value is pending or committed, a scoped monotonic guard denies later tool calls. After commit, the ordinary serial `agent/turn-stop` listener returns a stop decision after continuation and steering have already folded. A schema-validation failure remains an ordinary `INVALID_ARGS` tool error and leaves the child able to retry within the same turn.
+
+Pure Code Mode's registry contribution omits `structured_output` from native wire schemas and exposes it through the generated SDK. The assembly waterfall may deliberately change that presentation; execution still validates against the child-scoped definition, and the listener owns the consistency of any alternate model-visible route it creates.
+
+### Three execution boundaries are deliberately one-way
+
+Prompt assembly is intentionally cooperative, but three execution facts need one-way settlement after their extensible stages:
+
+| Boundary | Final power | Why ordinary listener order is insufficient |
+|---|---|---|
+| Tool pre-policy | Deny monotonically | A later listener must not re-allow an already denied call |
+| Tool result | Observe the immutable committed outcome | Structured output must commit only the result that actually escaped the pipeline |
+| Turn continuation | Stop after ordinary continuation folding | A committed terminal output must end the turn |
+
+`ToolGuard` is the monotonic policy registry. Committed tool observation is the contained `tools/result` point described above. Terminal structured output listens on the ordinary serial `agent/turn-stop` fold after normal continuation and steering decisions; no public `strictSerial()` dispatcher is needed for the typed listener contract.
+
+### Skill and approval services trust typed callers
+
+Skill registry definitions and approval policies are readonly same-process contracts. Their services do not clone callback objects or defend against post-handoff callback replacement.
+
+Skill still validates external skill files and parsed provider output, routes catalogs through the calling agent's tool view, and disposes registrations exactly. Approval still resolves policy, observes cancellation, routes `approval/request` by `request.agent`, records the durable audit pair, and contains answerer and post-commit observer failures.
+
+## Subagents: readiness is the start promise
+
+Subagent startup has one ownership transfer. The provider owns partial resources until its start promise fulfills with a ready published run; the caller owns the returned run and must dispose it.
+
+### The service contract has one cancellation channel
+
+`SubagentProvider.start()` and `SubagentService.start()` return `Promise<SubagentRun>`. The promise fulfills only after the backend has established the child it promises, so callers and `subagent/start` observers never need a second `run.started` readiness promise.
+
+`SubagentStartRequest.signal` is required. Aborting it requests cancellation during startup and after readiness. `SubagentRun.dispose()` also requests cancellation and awaits quiescence. There is no separate public `run.cancel()` channel.
+
+Optional `sendMessage()` supports a live backend that can accept steering. Optional `resume()` returns `Promise<SubagentRun>` because the resumed child has the same asynchronous readiness boundary.
+
+The service validates provider capabilities and request semantics before calling the provider. A provider rejection cleans any partial resources before the rejection escapes and emits no `subagent/start`/`subagent/end` pair. After fulfillment, the service attaches result observation, emits scoped start, and returns the run. Provider removal prevents later starts but does not revoke a run already accepted by the provider.
+
+### In-process providers reuse the core transaction
+
+Spawn and fork share one in-process driver. It creates the child through `parent.ctx`, passes the required signal into the core creation transaction, and installs persona, tool restriction, and structured-output contributions during unpublished setup.
+
+The provider awaits creation and returns only the published run. At the handoff, core creation detaches its creation-only abort listener; the provider immediately rechecks the signal before installing the live-run listener, so an abort in that narrow interval disposes the new handle instead of escaping cancellation. Parent teardown follows the child because the operation belongs to `parent.ctx`; provider unload blocks new starts but does not become a second revocation owner for accepted runs. The run disposer cancels the child and awaits the AgentHandle's ordered teardown.
+
+Spawn uses an empty session seed. Fork uses a validated completed-turn prefix. Conversation seeding changes history only and does not import scope, tools, services, or authority.
+
+### ACP providers own the process until readiness or cleanup
+
+An ACP provider crosses a real process and wire boundary, so it retains validation, environment scrubbing, message serialization, abort/process races, and kill-to-exit quiescence.
+
+Start resolves only after `initialize` and `newSession` succeed. Abort, spawn failure, RPC failure, or invalid startup response reaps the process before rejection. After readiness, result maps the ACP prompt outcome and streamed output; dispose requests cancellation, closes the connection, and awaits process exit through one memoized path.
+
+## Workflows and ACP UI: retain only independent async facts
+
+Worker and editor bridges need more state than same-process registries because messages, process death, and rendering can settle independently. Their state is organized around those real facts rather than duplicate cancellation protocols.
+
+### Workflow children are pending starts or published records
+
+The workflow host keeps pending provider-start promises and published child records. A child moves from pending to published only when async `SubagentService.start()` fulfills; rejected starts clean their partial provider work and produce no child lifecycle pair.
+
+One host-owned AbortController supplies the required signal to pending and live children. Closing workflow admission aborts that signal, so there is no duplicate `ChildCancel` worker RPC or explicit host-side `run.cancel()` fanout. Quiescence waits for both pending starts and published child disposal.
+
+The worker boundary still serializes requests and outcomes. The host retains first-terminal-outcome arbitration, exact child accounting, worker-death handling, grace termination, late/duplicate message rejection, and bounded cleanup because result receipt, worker exit, and child quiescence are genuinely independent facts.
+
+### Terminal result and physical cleanup remain separate
+
+The workflow result records the first accepted terminal outcome according to the public precedence rules. Cleanup can continue after that result is chosen: live children still need disposal, a worker still needs termination, and a slow external backend may outlive the configured grace bound.
+
+Public disposal claims its memoized promise before invoking callbacks. Worker death closes admission before processing any queued late child request, synthesizes missing lifecycle ends, and starts child/process cleanup without rewriting an outcome already claimed.
+
+### ACP prompt settlement does not depend on rendering success
+
+The ACP UI correlates a prompt with its observed turn directly. It does not scan from a `logWatermark` or use session status as a second reconciliation oracle.
+
+Prompt handling settles correlation in a `finally` around transcript rendering. A rendering failure can fail presentation, but it cannot skip prompt settlement or leave the session permanently in flight. Concurrent loads of the same persisted caller-supplied session ID remain excluded because that is a real persistence identity race, not a UUID collision concern.
+
+## Correctness enforcement
+
+The design is enforced at types, runtime escape points, generated contracts, and behavioral tests. No one layer is asked to prove what it cannot observe.
+
+### Types make the ordinary path hard to misuse
+
+Readonly contracts describe borrowed same-process values. `Scoped<T>` marks event receivers, `agentEvents()` fuses carrier and subject, tool inputs omit registry-owned tokens, and subagent async return types expose readiness directly.
+
+TypeScript cannot govern JavaScript casts, direct Cordis dispatch, process messages, or durable files, so runtime enforcement remains at those escape points.
+
+### Runtime invariants cover cross-service facts
+
+The invariants plugin verifies that every declared scoped event uses a marked carrier and that event families exposing a subject use the matching key. Session trace validation stages before append commit and advances after the same event commits.
+
+The plugin does not police trusted setup by scanning registries or reject prompt assembly objects fabricated through casts. Those checks would turn composition contracts into speculative runtime machinery without protecting a real external boundary.
+
+### Generated artifacts keep public contracts aligned
+
+The event catalog, service catalog, producer/consumer matrix, configuration catalog, module graph, tool catalog, type-equivalence blocks, and scoped-event resolver map are generated or freshness-gated from source. The [TypeScript semantic-gates Agent Note](../process/2026-07-14-typescript-program-backed-semantic-gates.md) owns Program construction, semantic event discovery, and resolver-generation rules.
+
+Behavioral tests pin scoped routing and disposal, final-entry collision cleanup, publication rollback, ordered quiescence, durable pre/post-commit behavior, live tool filtering across presentation and execution, cooperative prompt assembly, structured-output commit in native and Code Mode, async subagent startup and signal cancellation, worker terminal arbitration, ACP settlement, and process teardown.
+
+## Alternatives considered
+
+The [July 8 Agent Note](2026-07-08-agent-scope-contexts.md#alternatives-considered) owns alternatives to the public flat-scope contract. The alternatives here concern implementation shape.
+
+### Use a transparent proxy as the scope carrier
+
+A proxy that impersonates the subject must preserve property, callable, constructable, private-field, descriptor, and proxy-invariant behavior that listener routing never needs. A small opaque carrier keeps the filter and key while the explicit event argument carries the subject.
+
+### Reserve agent and session IDs before setup
+
+Reservations prevent duplicate private setup work but require cross-service capabilities, release ordering, abandoned-reservation cleanup, and prepared-object binding. IDs are caller-supplied and concurrent reuse is caller error; final entry can choose the winner while the losing transaction rolls back cleanly.
+
+### Snapshot every typed same-process argument
+
+Universal copying defends against stateful getters and callers that violate readonly contracts, but it adds allocation, duplicated validators, and paths that can forget to copy. Materialization belongs at parser, queue, model, durable, worker, process, and wire boundaries where ownership actually changes.
+
+### Give readiness, cancellation, and disposal separate controllers
+
+Parallel sentinels can all mirror whether one operation is live. One transaction or start promise owns the operation; separate promises remain only where publication unwind, external work, terminal result, and physical quiescence can settle independently.
+
+### Keep synchronous subagent start plus `run.started`
+
+This splits provider acceptance from readiness and forces every consumer to register a partial run, attach result observation, await readiness, and clean up readiness failure. An async start promise makes provider-to-caller ownership transfer the readiness boundary itself.
+
+### Restore selected prompt or tool contributions after assembly
+
+A post-waterfall restoration pass would create a second composition rule after the documented cooperative seam. Correctly assigning canonical presence or absence would also require provider ownership and collision rules for arbitrary tool-schema providers, whose ordinary output may contain duplicate names. Scoped registration already supplies the required per-agent isolation, and trusted assembly listeners own the protocol consistency of what they return, so named restoration adds machinery without establishing an independent boundary.
+
+### Remove worker/process lifecycle guards with same-process hardening
+
+Worker messages, process death, and durable input do cross ownership and serialization boundaries. First-outcome arbitration, validation, environment scrubbing, and quiescent process cleanup remain necessary even though hostile same-process callback machinery does not.
+
+## Consequences
+
+The implementation is smaller and its proof follows the same shape as its ownership graph. One key selects a layer, one entry owns a live registry object, one transaction owns creation, one resolver owns a tool view, and one async promise transfers subagent ownership.
+
+### What the design guarantees
+
+- A scoped contribution is visible only in its exact agent view and is disposed with that scope.
+- Create and resume expose no partially configured handle; final-entry losers and publication failures clean every prepared resource.
+- Disposal retains scoped listeners and persistence through driver drain and final session work, then revokes the scope.
+- Durable, queued, model, worker, process, and wire values are owned at their real boundary; typed same-process values follow readonly contracts.
+- ToolRegistry's presentation, lookup, and execution resolve the same live view before expert assembly transforms, and committed results have one immutable observation point.
+- Registry contributions are deterministic inputs, while the trusted assembly waterfall owns the final model-visible composition.
+- Subagent start returns only a ready run, required signals cancel pending or live work, and disposal reaches the backend's quiescence contract.
+- Worker/process result precedence and cleanup remain correct under death, late messages, and bounded teardown.
+
+### Costs and limits
+
+Scope-aware services still maintain global and identity-keyed maps, and operations must carry their real agent explicitly. Async create/resume and subagent start require callers to await ownership transfer and dispose returned handles.
+
+A trusted `system-prompt/assemble` listener can remove or replace Code Mode and structured-output protocol pieces. This is deliberate: the listener owns final composition and must preserve any protocol the deployment expects to remain usable.
+
+The design trusts typed plugins in the same process. It does not defend against arbitrary casts, stateful getters, mutation that violates readonly contracts, or a plugin deliberately using ambient service access outside the supported composition API.
+
+The [security and authority non-goal](2026-07-08-agent-scope-contexts.md#security-and-authority-are-non-goals) remains fundamental. These mechanisms prove registration composition, publication, and lifetime ownership; they do not prove confinement or parent-to-child non-escalation.

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-14-provider-routed-llm-adapters.md: b7944bd31fdb5f63894e867d7c1224215d694f11
+2026-07-14-provider-routed-llm-adapters.zh.md: 7dcadf2521bab079e328b5f0d0a45185778b3b8d

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.md

@@ -0,0 +1,91 @@
+# Agent Note: Provider-routed LLM adapters and a generic pi-ai backend
+
+Status: implemented
+
+English | [中文](2026-07-14-provider-routed-llm-adapters.zh.md)
+
+## Problem
+
+`dsh-llm` registered adapters by exact model name. A plugin supplied a model list at Cordis startup, `LlmService` stored one adapter per listed string, and `GenerateOptions.model` selected the adapter and the provider model at once. This worked while both shipping adapters targeted the same two DeepSeek models, but it conflated two independent decisions: which upstream provider owns a request, and which model that provider should run.
+
+The conflation prevents a provider gateway from serving an open-ended model catalog. OpenRouter, for example, is one provider with many model ids, while a private OpenAI-compatible endpoint may add models without changing the Harness plugin tree. Every newly selected model currently needs to have been registered during plugin startup. The same model id can also exist at multiple providers, so model-only registration cannot state which provider the caller intended.
+
+`dsh-llm-pi-ai` exposed none of pi-ai's provider abstraction. It constructed an inline DeepSeek `openai-completions` model, applied DeepSeek-specific payload patches, and stamped every replayed assistant message as DeepSeek. pi-ai itself has a provider/model catalog, selects APIs such as `openai-responses`, `anthropic-messages`, and `google-generative-ai`, and preserves provider-specific response ids and reasoning/tool signatures for later turns. The Harness conversion dropped that provenance, so simply replacing the inline model with a catalog lookup would have made same-model replay and cross-provider handoff incomplete.
+
+The adapter configuration also assumes one DeepSeek API key and endpoint. A generic backend needs independent credentials and endpoint overrides per provider while leaving AWS, Google ADC, OAuth, and other ambient authentication mechanisms to pi-ai.
+
+## Decision
+
+### Provider is the adapter registration key
+
+`GenerateOptions` and `LlmCallConfig` carry `provider: string` beside `model: string`; `AgentOptions` carries the corresponding optional creation field. A loop request is valid only after both values are non-empty, and both values are part of the logged request header. `agent/request` may return a replacement pair on any step, so a session can switch providers and models without changing the Cordis plugin lifecycle.
+
+`LlmService` registers and resolves adapters by provider. `registerAdapter(providers, adapter)` checks the entire provider list before mutating the registry, rejects a duplicate with `DUPLICATE_ADAPTER`, and disposes the whole registration as one effect. Model ids are not registration keys; the selected adapter still validates or forwards them. The later [LLM catalog and ACP selection Agent Note](2026-07-15-llm-model-catalog-and-acp-selection.md) added advisory `listProviders()` / `listModels()` discovery without turning model membership into request validation.
+
+A provider has exactly one adapter owner in a Cordis context. `dsh-llm-deepseek` registers `deepseek`; `dsh-llm-pi-ai` may also register `deepseek`, but loading both owners is a configuration error rather than an ordering rule or fallback. A deployment that wants the hand-rolled DeepSeek implementation excludes `deepseek` from the pi-ai profiles. A deployment that wants pi-ai's DeepSeek implementation does not mount `dsh-llm-deepseek`.
+
+`dsh-llm-deepseek` removes its model registration list and accepts any model string routed through provider `deepseek`. Its request serialization, `/chat/completions` endpoint, thinking options, SSE parsing, and error behavior remain unchanged; `options.model` is still sent verbatim.
+
+### Explicit pi-ai provider profiles
+
+`dsh-llm-pi-ai` takes one non-empty list of provider profiles. Provider names must be unique within the list and present in pi-ai's `getProviders()` result. Each profile contains the provider name plus optional `apiKey`, `baseURL`, headers, reasoning level and budgets, cache retention, transport, timeouts, and retry settings. Credentials are never global: an explicit key applies only to its profile, while an absent key lets pi-ai resolve its standard environment variable, OAuth token, AWS credential chain, Google ADC, or other provider-native ambient authentication. An explicitly empty key is invalid configuration rather than an environment fallback.
+
+The plugin registers all configured provider names against one `PiAiAdapter` in one all-or-nothing call. A request uses its provider to select the matching profile and finds its model in `getModels(provider)` to obtain the catalog descriptor. An unknown provider fails at plugin load; an unknown model fails before network I/O with `UNKNOWN_MODEL`. The catalog object is never mutated. When a profile supplies `baseURL`, the adapter clones the selected descriptor and overrides only `baseUrl`, so a private endpoint can retain pi-ai's API, capabilities, compatibility flags, context limits, and reasoning map. The private endpoint must implement the selected provider's protocol, and the model id must still exist in the installed pi-ai catalog.
+
+The adapter calls pi-ai's `streamSimple()` so each catalog model chooses its registered API implementation, including OpenAI Responses instead of Chat Completions where the descriptor says `openai-responses`. Harness temperature, maximum tokens, signal, session id, and the profile's common stream options flow through directly. Profile headers merge with the mandatory Harness attribution headers, with Harness attribution winning its reserved names. The adapter no longer maintains DeepSeek-specific payload rewrites or a provider-protocol matrix.
+
+pi-ai's common stream options do not expose stop sequences. `dsh-llm-pi-ai` rejects a defined Harness `stop` option with `UNSUPPORTED_OPTION` rather than silently ignoring it or growing a second provider-specific payload implementation. `dsh-llm-deepseek` continues to support `stop` through its native request serializer.
+
+### Durable assistant provenance and replay state
+
+Assistant messages carry provider-neutral provenance containing the request's `provider` and `model`, plus an optional JSON-serializable adapter replay state. A successful `assistant/message` session event records this provenance and `deriveMessages()` returns it with the assistant message. User, system, context, and tool-result messages carry no assistant provenance. The provider/model fields are authoritative loop data; an adapter owns only its opaque replay-state payload.
+
+A terminal successful `finish` chunk may carry replay state, and `BlockAssembler` retains it alongside usage and finish reason. The loop attaches it to the assistant provenance only when the post-`agent/step-result` content is structurally equal to the assembled provider output. A listener that rewrites content keeps the provider/model provenance but loses the now-stale replay state. Error and aborted responses do not produce a normal assistant message and therefore do not enter future model history.
+
+The pi-ai replay state is a versioned, minimal projection of its successful `AssistantMessage`: source API/provider/model, response id/model, stop reason, and index-aligned text, thinking, and tool-call signatures. It does not duplicate text or tool arguments already carried by Harness content blocks, and it omits diagnostics, timestamps, usage, and errors. On a later request, `LlmService` gives replay state to the target adapter only when the historical provider and target provider are currently owned by the same adapter instance. That adapter combines the logged Harness content with replay state when it can restore the historical response, and owns any required cross-model or cross-provider conversion. An adapter receiving replay state with an unknown version or mismatched block shape fails explicitly; a different adapter receives only provider-neutral content and provenance.
+
+This state is model-visible replay input and therefore follows the existing [reconstructable-request rule](2026-07-05-reconstructable-requests.md): it is present in both the terminal `finish` chunk and the assembled `assistant/message` provenance that drives derivation. Resume and fork preserve it verbatim. Compaction that shadows the assistant message also removes its replay state from the active surface; the summary is ordinary provider-neutral content.
+
+### Propagate the target through every request producer
+
+Every model-selection surface carries provider and model together: declarative agents, ACP and stdio app config, the JSON-RPC initialize request, subagent overrides and inheritance, workflow child overrides, and direct compaction summarization. Subagents inherit both fields from their parent before applying request overrides. The system-prompt variable set gains `provider` beside `model`.
+
+Compaction configuration gains `summarizationProvider` beside `summarizationModel`. Both are empty to inherit, or both are non-empty to select an explicit target; a half-configured pair fails load. Inheritance uses the last logged request target when one exists and falls back to the agent's creation options. `compact/summary` records both fields with the existing model-call envelope.
+
+The JSON-RPC runtime receives provider and model explicitly. Its convenience fallback mounts `dsh-llm-deepseek` only for provider `deepseek` when that provider has no registered owner; other missing providers fail without guessing an adapter.
+
+The on-disk session format remains the pre-release pinned version `0`, with no compatibility promise. Seed/load validation rejects request headers lacking provider and assistant messages lacking required provenance instead of accepting an old shape that can no longer reconstruct the request.
+
+## Alternatives considered
+
+**Keep model names as registry keys and add wildcard adapters.** A wildcard introduces fallback ordering between exact registrations and catch-all plugins, makes duplicate ownership dependent on listener order, and still cannot distinguish the same model id at two providers without another convention.
+
+**Encode provider and model into one string.** Values such as OpenRouter's `openai/gpt-*` already contain provider-like prefixes and slashes. A delimiter convention would leak routing syntax into every model selector and require escaping rules; two explicit fields are unambiguous and independently loggable.
+
+**Add `backend + provider + model`.** A backend key would allow `dsh-llm-deepseek` and pi-ai's DeepSeek implementation to coexist and switch per request. The accepted deployment rule is instead one adapter owner per provider: implementations of the same upstream are alternatives selected by plugin composition. A third routing dimension would burden every request and configuration for a capability with no current consumer.
+
+**Let `dsh-llm-pi-ai` automatically register every pi-ai provider.** This would claim ambient credentials and provider names the deployment never intended to expose, and would conflict with native adapters such as `dsh-llm-deepseek`. Explicit profiles make capability and credential scope reviewable.
+
+**Mount one pi-ai plugin instance per provider.** Separate instances isolate config but repeat plugin declarations and cannot make profile registration atomic. One adapter already receives provider on every request, so a validated profile map is the smaller lifecycle surface.
+
+**Accept arbitrary inline pi-ai model descriptors.** This would support catalog-external private model ids, but it exposes pi-ai's model and compatibility schema as Harness configuration and makes the adapter responsible for validating protocol-specific combinations. The first version supports custom endpoints by overriding `baseURL` on catalog models; custom descriptors require a separate decision after a real catalog-external deployment is identified.
+
+## Consequences
+
+- Provider names are deployment-wide route ownership keys: two providers may use the same model string, but mounting two adapters for one provider fails at load instead of creating fallback order.
+- Model selection no longer changes the Cordis plugin graph. Catalog-backed adapters can accept any installed catalog model selected after startup, while the native DeepSeek adapter forwards arbitrary DeepSeek model ids.
+- A custom `baseURL` preserves the selected catalog model's protocol and capabilities; it does not make catalog-external model ids valid. Private endpoints must implement that catalog entry's protocol.
+- pi-ai credentials and transport knobs are scoped per provider profile. An omitted key delegates to pi-ai ambient authentication, while an explicitly empty key is invalid.
+- `dsh-llm-pi-ai` rejects stop sequences because pi-ai's common stream API cannot express them; the native DeepSeek adapter retains its stop support.
+- Replay state is portable only within the adapter instance that owns both the historical and target providers. Cross-provider and cross-model restoration is an adapter responsibility, and another adapter receives provider-neutral history without the opaque state.
+- Current pre-release session JSONL requires provider/model request headers and assistant provenance. Older shapes remain version `0` but are rejected rather than migrated.
+
+## Testing
+
+- Unit coverage exercises registry conflicts, request reconstruction, session validation, profile resolution, option forwarding, native API selection including OpenAI Responses, conversion, replay validation, error mapping, cancellation, content rewrites, and same-instance versus different-instance replay dispatch.
+- Keyless loop/session tests and ACP snapshots exercise durable provider/model metadata, resume and fork propagation, workflow/subagent overrides, and unchanged user-visible transcripts; the key-gated DeepSeek e2e retains real provider streaming and tool follow-up coverage.
+- Public JSDoc, package READMEs, architecture and core-data-structure docs, generated catalogs, examples, session fixtures, and Python SDK pairs use provider/model targets consistently and are checked by the repository documentation and type-equivalence gates.
+
+## Risks
+
+This is a repo-wide pre-release API break: model-only request construction, adapter registration, app protocols, fixtures, and persisted version-0 event shapes all change together, with no compatibility aliases. The provider exclusivity rule deliberately prevents two implementations of the same upstream from coexisting in one context. A pi-ai dependency update can change the accepted provider/model catalog, so the lockfile and adapter e2e matrix define the tested set. Custom `baseURL` endpoints inherit the chosen catalog model's protocol assumptions and cannot repair an incompatible proxy. Catalog-external model descriptors and multimodal content remain unsupported. pi-ai replay state may contain opaque encrypted reasoning signatures; it is persisted because the provider requires it for continuity, but it is never rendered or logged outside the existing session record.

.agents/notes/implemented/architecture/2026-07-14-provider-routed-llm-adapters.zh.md

@@ -0,0 +1,91 @@
+# Agent Note: 基于提供方路由的 LLM 适配器与通用 pi-ai 后端
+
+Status: implemented
+
+[English](2026-07-14-provider-routed-llm-adapters.md) | 中文
+
+## 问题
+
+`dsh-llm` 按精确模型名称注册适配器。插件在 Cordis 启动时提供模型列表，`LlmService` 为列表中的每个字符串保存一个适配器，`GenerateOptions.model` 同时选择适配器与提供方模型。两个正式适配器都只面向相同的两个 DeepSeek 模型时，这种方式可以工作，但它混淆了两个独立决策：由哪个上游提供方承接请求，以及该提供方应运行哪个模型。
+
+这种混淆使提供方网关无法提供开放的模型目录。例如，OpenRouter 是一个包含大量模型 ID 的提供方，私有 OpenAI 兼容端点也可能在不修改 Harness 插件树的情况下增加模型。目前，每个新选择的模型都必须在插件启动期间完成注册。同一个模型 ID 还可能存在于多个提供方中，因此仅按模型注册无法表达调用方预期使用的提供方。
+
+`dsh-llm-pi-ai` 没有暴露 pi-ai 的提供方抽象。它以内联方式构造 DeepSeek `openai-completions` 模型，应用 DeepSeek 专用的 payload 补丁，并将每条回放的助手消息标记为 DeepSeek。pi-ai 自身提供提供方/模型目录，能够选择 `openai-responses`、`anthropic-messages`、`google-generative-ai` 等 API，并保留提供方专用的响应 ID，以及后续轮次所需的推理和工具签名。Harness 转换丢弃了这些来源信息，因此仅将内联模型替换为目录查询，会导致同模型回放与跨提供方移交不完整。
+
+适配器配置同样假定只存在一个 DeepSeek API 密钥和端点。通用后端需要为各提供方分别配置凭据和端点覆盖，同时继续由 pi-ai 处理 AWS、Google ADC、OAuth 等环境认证机制。
+
+## 决策
+
+### 提供方作为适配器注册键
+
+`GenerateOptions` 与 `LlmCallConfig` 在 `model: string` 之外携带 `provider: string`，`AgentOptions` 则携带对应的可选创建字段。只有两个值都非空时，agent loop（智能体循环）请求才有效；两个值也都会写入请求头日志。`agent/request` 可以在任意步骤返回替换后的字段组合，因此会话可以切换提供方与模型，无需改变 Cordis 插件生命周期。
+
+`LlmService` 按提供方注册和解析适配器。`registerAdapter(providers, adapter)` 在修改注册表前检查整个提供方列表，遇到重复项时返回 `DUPLICATE_ADAPTER`，并将整组注册作为一个 effect 释放。模型 ID 不作为注册键；仍由选中的适配器负责验证或转发。后续的 [LLM 目录与 ACP 模型选择 Agent Note](2026-07-15-llm-model-catalog-and-acp-selection.md) 增加了建议性的 `listProviders()` / `listModels()` 发现接口，但不会把目录成员关系变成请求校验规则。
+
+在一个 Cordis 上下文中，一个提供方只能有一个适配器所有者。`dsh-llm-deepseek` 注册 `deepseek`；`dsh-llm-pi-ai` 也可以注册 `deepseek`，但同时加载两个所有者属于配置错误，不采用顺序规则或回退行为。若部署选择手写的 DeepSeek 实现，需从 pi-ai 配置中排除 `deepseek`；若部署选择 pi-ai 的 DeepSeek 实现，则不挂载 `dsh-llm-deepseek`。
+
+`dsh-llm-deepseek` 移除模型注册列表，接受通过 `deepseek` 提供方路由的任意模型字符串。其请求序列化、`/chat/completions` 端点、thinking 选项、SSE（Server-Sent Events）解析和错误行为保持不变；`options.model` 仍会原样发送。
+
+### 显式 pi-ai 提供方配置
+
+`dsh-llm-pi-ai` 接受一个非空的提供方配置列表。列表内的提供方名称必须唯一，并且存在于 pi-ai 的 `getProviders()` 结果中。每项配置包含提供方名称，以及可选的 `apiKey`、`baseURL`、headers、推理级别和预算、缓存保留设置、传输方式、超时和重试设置。凭据不设全局值：显式密钥仅对所属配置生效；未提供密钥时，pi-ai 使用标准环境变量、OAuth token、AWS 凭据链、Google ADC 或其他提供方原生环境认证。显式空密钥属于无效配置，不会回退到环境认证。
+
+插件通过一次全有或全无调用，将所有已配置的提供方名称注册到同一个 `PiAiAdapter`。请求按 provider 选择对应配置，并在 `getModels(provider)` 中查找模型以取得目录描述符。未知提供方会在插件加载时失败；未知模型会在网络 I/O 前以 `UNKNOWN_MODEL` 失败。适配器不会修改目录对象。当配置提供 `baseURL` 时，适配器复制选中的描述符，仅覆盖 `baseUrl`，使私有端点保留 pi-ai 的 API、能力、兼容标志、上下文限制与推理映射。私有端点必须实现所选提供方的协议，模型 ID 也仍须存在于已安装的 pi-ai 目录中。
+
+适配器调用 pi-ai 的 `streamSimple()`，因此每个目录模型会选择其注册的 API 实现；描述符为 `openai-responses` 时使用 OpenAI Responses，而非 Chat Completions。Harness 的 temperature、最大 token 数、signal、session ID，以及提供方配置中的通用流选项均直接传递。配置 headers 与 Harness 强制归因 headers 合并；发生保留名称冲突时，以 Harness 归因为准。适配器不再维护 DeepSeek 专用 payload 重写或提供方协议矩阵。
+
+pi-ai 的通用流选项不支持停止序列。若 Harness `stop` 选项已定义，`dsh-llm-pi-ai` 会以 `UNSUPPORTED_OPTION` 拒绝请求，不会静默忽略，也不会增加第二套提供方专用 payload 实现。`dsh-llm-deepseek` 继续通过原生请求序列化器支持 `stop`。
+
+### 持久化助手来源信息与回放状态
+
+助手消息携带提供方无关的来源信息，其中包含请求的 `provider` 和 `model`，以及可选的 JSON 可序列化适配器回放状态。成功的 `assistant/message` 会话事件记录这些来源信息，`deriveMessages()` 返回助手消息时也会包含这些信息。用户、system、context 与工具结果消息不携带助手来源信息。provider/model 字段是 agent loop 的权威数据；适配器仅拥有其不透明回放状态 payload。
+
+成功的终止 `finish` 分片可以携带回放状态，`BlockAssembler` 会将其与 token 用量和结束原因一起保留。只有当 `agent/step-result` 处理后的内容与提供方组装输出在结构上相等时，agent loop 才会把回放状态附加到助手来源信息。监听器重写内容后，provider/model 来源信息仍会保留，但已经陈旧的回放状态会被移除。错误或中止响应不会生成正常助手消息，因此不会进入后续模型历史。
+
+pi-ai 回放状态是其成功 `AssistantMessage` 的带版本最小投影，包含源 API/provider/model、响应 ID/model、停止原因，以及按索引对齐的文本、thinking 和工具调用签名。它不会重复 Harness 内容块中已有的文本或工具参数，也不包含诊断信息、时间戳、用量或错误。后续请求中，只有历史提供方和目标提供方当前归同一个适配器实例所有时，`LlmService` 才会把回放状态交给目标适配器。适配器在能够恢复历史响应时，将 Harness 记录的内容与回放状态组合，并负责所需的跨模型或跨提供方转换。适配器收到未知版本或块形状不匹配的回放状态时会显式失败；其他适配器只能收到提供方无关的内容与来源信息。
+
+该状态属于模型可见的回放输入，因此遵循现有的[请求可重建规则](2026-07-05-reconstructable-requests.md)：它同时存在于终止 `finish` 分片和驱动派生的已组装 `assistant/message` 来源信息中。恢复和 fork 会原样保留该状态。压缩（compaction）遮蔽助手消息时，也会从活动 surface 中移除其回放状态；摘要属于普通的提供方无关内容。
+
+### 在所有请求生产方中传播目标
+
+每个模型选择接口都同时携带 provider 与 model：声明式 agent、ACP（Agent Client Protocol）和 stdio 应用配置、JSON-RPC initialize 请求、subagent 覆盖与继承、工作流子 agent 覆盖，以及直接压缩摘要。subagent 先从父 agent 继承两个字段，再应用请求覆盖。系统提示词变量集合在 `model` 之外增加 `provider`。
+
+压缩配置在 `summarizationModel` 之外增加 `summarizationProvider`。两个值均为空时继承，均非空时选择显式目标；只配置其中一个会导致加载失败。继承优先使用最近一次记录的请求目标，没有时回退到 agent 创建选项。`compact/summary` 使用现有模型调用 envelope 记录两个字段。
+
+JSON-RPC 运行时显式接收 provider 与 model。仅当 `deepseek` 提供方没有注册所有者时，其便利回退才会挂载 `dsh-llm-deepseek`；其他缺失的提供方会直接失败，不会猜测适配器。
+
+磁盘会话格式仍使用预发布阶段固定的版本 `0`，且不承诺兼容性。seed/load 验证会拒绝缺少 provider 的请求头，以及缺少必需来源信息的助手消息，不会接受已无法重建请求的旧格式。
+
+## 考虑过的替代方案
+
+**继续以模型名称作为注册表键，并增加通配适配器。** 通配机制会在精确注册与兜底插件之间引入回退顺序，使重复所有权取决于监听器顺序；若不再增加其他约定，仍无法区分不同提供方中相同的模型 ID。
+
+**将提供方与模型编码到一个字符串中。** OpenRouter 的 `openai/gpt-*` 等值已经包含类似提供方的前缀和斜杠。分隔符约定会把路由语法泄漏到每个模型选择接口，并需要转义规则；两个显式字段更清晰，也可以分别记录日志。
+
+**增加 `backend + provider + model`。** backend 键可以让 `dsh-llm-deepseek` 与 pi-ai 的 DeepSeek 实现共存，并按请求切换。最终采用的部署规则是一个提供方对应一个适配器所有者：同一上游的不同实现属于由插件组合选定的替代项。第三个路由维度会增加每个请求与配置的负担，却没有当前消费方。
+
+**让 `dsh-llm-pi-ai` 自动注册所有 pi-ai 提供方。** 这种方式会占用部署无意暴露的环境凭据和提供方名称，并与 `dsh-llm-deepseek` 等原生适配器冲突。显式配置可以审查能力和凭据范围。
+
+**每个提供方挂载一个 pi-ai 插件实例。** 独立实例可以隔离配置，但会重复插件声明，也无法实现配置注册的原子性。每个请求本就向同一个适配器提供 provider，因此经过验证的配置映射具有更小的生命周期接口。
+
+**接受任意内联 pi-ai 模型描述符。** 这种方式可支持目录外的私有模型 ID，但会将 pi-ai 的模型与兼容性 schema 暴露为 Harness 配置，并要求适配器验证协议专用组合。当前版本通过覆盖目录模型的 `baseURL` 支持自定义端点；只有实际出现目录外部署需求后，才会另行决策是否支持自定义描述符。
+
+## 影响
+
+- 提供方名称是部署范围内的路由所有权键：两个提供方可以使用相同的模型字符串，但为同一个提供方挂载两个适配器会在加载时失败，不会形成回退顺序。
+- 模型选择不再改变 Cordis 插件图。目录型适配器可以接受启动后选择的任意已安装目录模型，原生 DeepSeek 适配器则会转发任意 DeepSeek 模型 ID。
+- 自定义 `baseURL` 会保留所选目录模型的协议与能力，但不会让目录外模型 ID 变为有效。私有端点必须实现该目录项对应的协议。
+- pi-ai 凭据与传输选项按提供方配置隔离。省略密钥时委托 pi-ai 使用环境认证；显式空密钥无效。
+- pi-ai 的通用流 API 无法表达停止序列，因此 `dsh-llm-pi-ai` 会拒绝停止序列；原生 DeepSeek 适配器仍支持停止序列。
+- 仅当历史提供方与目标提供方归同一个适配器实例所有时，回放状态才可移植。适配器负责跨提供方和跨模型恢复；其他适配器只接收不含不透明状态的提供方无关历史。
+- 当前预发布会话 JSONL 要求请求头包含 provider/model，助手消息包含来源信息。旧格式仍使用版本 `0`，但会被拒绝，不执行迁移。
+
+## 测试
+
+- 单元测试覆盖注册表冲突、请求重建、会话验证、配置解析、选项转发、包括 OpenAI Responses 在内的原生 API 选择、转换、回放验证、错误映射、取消、内容重写，以及同一实例与不同实例间的回放分发。
+- 无密钥的 agent loop/会话测试和 ACP 快照覆盖持久化 provider/model 元数据、恢复与 fork 传播、工作流/subagent 覆盖，以及不变的用户可见 transcript（文本记录）；密钥门控的 DeepSeek e2e 测试保留真实提供方的流式输出与工具后续调用覆盖率。
+- 公共 JSDoc、package README、架构与核心数据结构文档、生成目录、示例、会话 fixture（测试前置数据）和 Python SDK 配对文档统一使用 provider/model 目标，并由仓库文档与类型等价门禁校验。
+
+## 风险
+
+这是一次覆盖全仓库的预发布 API 破坏性变更：仅模型的请求构造、适配器注册、应用协议、fixture，以及持久化版本 0 事件格式会同时变化，不提供兼容别名。提供方排他规则有意禁止同一上游的两个实现共存于同一上下文。pi-ai 依赖升级可能改变可接受的提供方/模型目录，因此锁文件与适配器 e2e 矩阵定义已验证集合。自定义 `baseURL` 端点会继承所选目录模型的协议假设，无法修复不兼容的代理。目录外模型描述符与多模态内容仍不受支持。pi-ai 回放状态可能包含不透明的加密推理签名；提供方需要该信息维持连续性，因此系统会持久化该状态，但不会在现有会话记录之外渲染或记录它。

.agents/notes/implemented/architecture/2026-07-15-agent-initiator-scope.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-15-agent-initiator-scope.md: 69648100e76cfc212469854188d664357fec22f1
+2026-07-15-agent-initiator-scope.zh.md: 835d7a5b2ab6d2d6fce7971de4fd9d6c69e50d77

.agents/notes/implemented/architecture/2026-07-15-agent-initiator-scope.md

@@ -0,0 +1,65 @@
+# Agent Note: Initiating Agent scope over AsyncLocalStorage
+
+Status: implemented
+
+English | [中文](2026-07-15-agent-initiator-scope.zh.md)
+
+## Problem
+
+The harness has two useful but different notions of context. A Cordis `Context` selects services, registration ownership, and lifetime; `agent.ctx` is the flat registration scope owned by one live Agent. Agent and Session identity instead describe the subject of an asynchronous operation. Changing a root `ctx.agent` to mean “whichever Agent is running” would conflate those meanings and fail when one process drives Agents concurrently.
+
+Deep process-local infrastructure sometimes needs a trusted initiating Agent below explicit loop, tool, and request parameters—for example, a host-aware transport, tracing helper, logger, or gateway client. Requiring every private helper to forward `agent` adds repetition, while a process-global mutable slot is incorrect across `await`. Model-visible arguments are unsuitable because a model must not choose a trusted Session or routing header. The carrier belongs to the Agent service rather than optional model-visible context.
+
+## Decision
+
+The mandatory `ctx.agents` service uses Node `AsyncLocalStorage` to carry the initiating Agent. It stores the exact `Agent` directly rather than introducing a one-field frame; a separate private run token records nested boundary lineage only for teardown bookkeeping and carries no identity. The [core-data catalog](../../../../docs/core-data-structures/core.md#initiating-agent) identifies the carried type.
+
+`currentInitiator()` reads optionally, `requireInitiator()` throws `no initiating agent is active`, and `withInitiator(agent, operation)` preserves the operation's exact synchronous value or Promise. `withoutInitiator(operation)` establishes a clearing boundary for work that must not inherit an Agent. Session remains derived as `agent.session`; turn, step, tool call, `signal`, model, `cwd`, sandbox, and authorization stay with their existing owners.
+
+`AgentLoop` already injects `ctx.agents` and wraps each concrete driver's complete `runLoop` lifetime in `agents.withInitiator(agent, ...)`. Its package-private loop, turn, step, and tool-call orchestration entries recover the exact Agent from `ctx.agents`, derive `agent.session` once, and let operation-local helpers capture it instead of forwarding the concrete driver or `Session` through shallow interfaces. A leaf helper keeps a narrow `Session` parameter when that is its actual interface rather than accepting a broader `Context` only for an ambient lookup.
+
+Concurrent drivers receive independent stores. A child driver's continuations carry the child, while the caller resumes in its prior store as soon as `withInitiator()` returns; active-run tracking keeps the returned Promise in the teardown drain until it settles. Creation, persistence load, and unpublished `setup(agentCtx)` remain outside the child's driver boundary: creation initiated by a parent runs under the parent identity, while `agentCtx.agent` explicitly identifies the child.
+
+Ambient identity does not replace explicit contracts. `ToolExecution.agent`, `AssembleContext.agent`, `GenerateOptions.sessionId`, task ownership, parent/child requests, `ctx.agent`, `agentCtx.agent`, approval and hook subjects, `cwd` selection, cancellation, worker/process messages, persistence records, and wire identity remain explicit. A remote boundary materializes the identity it needs into its typed request because ALS is process-local.
+
+`AgentRegistry` owns an ordered initiator lifecycle. Teardown first rejects new boundaries; removing `ctx.agents` then drains injected dependents such as AgentLoop, and the registry waits for active returned-Promise boundaries before calling `AsyncLocalStorage.disable()`. If a boundary's inherited async chain starts an owning Cordis fiber's unload, the private run-token lineage releases that nested boundary chain from the drain, which prevents teardown from waiting on itself while unrelated boundaries still drain. `currentInitiator()` and `requireInitiator()` remain usable through a retained in-flight service reference while the ordinary drain runs; after disposal, initiator methods throw `agent initiator scope is disposed`. Root Context disposal may start sibling fiber teardown concurrently, so active-boundary counting remains necessary in addition to Cordis dependency ordering.
+
+Initiator scope does not own detached work: registry drain tracks only the Promise returned by `withInitiator()` or `withoutInitiator()`. Asynchronous resources created inside a boundary inherit its store until they settle or ALS is disabled, so their owning seam must stop unreturned work explicitly. Agent-owned foreground work returns its lifetime and keeps its cancellation contract. Unrelated timers, queues, and deployment infrastructure start under `withoutInitiator(operation)`; queue, worker, process, and wire boundaries serialize identity rather than expecting ALS propagation.
+
+A host-aware transport may derive a deployment-owned header such as `X-Harness-Session-Id` from `ctx.agents.requireInitiator().session.id`; the header is absent from model-visible schema and arguments. No production MCP or Web transport adopts such a header in this decision. A test-double transport proves the trusted boundary without assigning host routing policy to an existing provider-neutral seam.
+
+This decision extends the [Agent registration-scope contract](2026-07-08-agent-scope-contexts.md) and its [runtime design](2026-07-12-agent-scope-runtime-design.md); it does not change their static `agent.ctx` meaning.
+
+## Verification
+
+Agent service tests pin optional and required reads, exact synchronous and cross-realm Promise identity, intrinsic Promise settlement observation, overlapping, nested, and cleared boundaries, restoration after throws or rejection, ordinary and reentrant drain ordering, and retained-reference errors. AgentLoop integration pins concurrent and nested drivers, agentless calls, AgentRegistry restart, root teardown, and package-private loop and tool scheduling through the ambient lookup. Composition, module-graph, build, and runtime-closure checks keep `ctx.agents` wired through the default bundle, SDK spine, Python runtime closure, and direct AgentLoop harnesses without another provider.
+
+A test-double host-aware transport derives `X-Harness-Session-Id` internally and verifies that tool schema and logged arguments contain no identity field. The service deliberately does not drain async work omitted from the Promise returned by the boundary operation; that work remains subject to its owner's explicit stop contract.
+
+## Alternatives considered
+
+**Pass Agent through every function.** Public, worker, process, persistence, and wire boundaries continue to do this, but requiring every process-local private helper to carry Agent adds repetitive forwarding without improving trust. ALS is confined to the asynchronous chain inside those explicit boundaries.
+
+**Make `ctx.agent` dynamic.** `ctx.agent` already means the static Agent associated with an Agent-scoped Cordis context. Changing the root meaning would mix registration and execution scopes and make concurrent behavior surprising.
+
+**Add a separate `ctx.agentExecution` service.** The carrier has no independent backend, configuration, or identity type: it stores the same `Agent` that `ctx.agents` already owns, and AgentLoop already depends on that service. A second mandatory provider would add package, composition, lifecycle, generated-catalog, and test-harness wiring without separating a real capability.
+
+**Store a named or complete runtime frame.** A one-field `{ agent }` frame only wraps the value, while Agent, Session, inbox, cancellation, turn, step, tool execution, and persistence already have authoritative owners. Adding more fields would create stale snapshots and another lifecycle; carrying `Agent` directly keeps the boundary named by its methods without duplicating state.
+
+**Include a step `AbortSignal`, `cwd`, sandbox, or authorization.** Their lifetimes and authority do not match the driver boundary, and their existing seams already pass them explicitly. Adding a control capability requires a separate decision and nested lifecycle contract.
+
+**Use a process-global `currentAgent`.** Concurrent Agents and subagents overwrite one another across awaited continuations, so a mutable global is correct only under a serialization guarantee the harness does not make.
+
+**Derive identity from model-visible arguments.** Model or user input cannot be trusted to select Session, tenant, or sandbox routing.
+
+**Add routing identity to every capability seam.** That spreads hosting concerns through provider-neutral APIs. A host-aware implementation owns its transport header while public boundaries remain explicit.
+
+## Consequences
+
+Deep infrastructure gains one trusted process-local initiating Agent without widening existing tool and capability requests. Concurrent and nested drivers isolate automatically, AgentLoop gains no additional mandatory service, and HMR/root disposal reaches quiescence before ALS is disabled.
+
+The dependency is implicit in function signatures and carries a capability-bearing Agent object. Consumers must restrict it to cross-cutting infrastructure, treat ambient presence as neither liveness nor authorization, and retain explicit cancellation and ownership checks. ALS also has an always-on propagation cost and does not cross worker, process, HTTP, or durable queue boundaries.
+
+The teardown design deliberately accepts Node's [Stability 1 (Experimental)](https://nodejs.org/api/async_context.html#asynclocalstoragedisable) `AsyncLocalStorage.disable()` dependency. Node requires `disable()` before an ALS instance can be garbage-collected, which matters when HMR replaces AgentRegistry-owned instances; the service state guard prevents a later boundary from re-entering the instance after disposal.
+
+The scope deliberately carries only the Agent, omitting turn, step, `signal`, `cwd`, sandbox, and authorization. A real consumer that cannot use existing explicit fields must justify any refinement separately; a stale copied field may at most mislabel telemetry, never grant control.

.agents/notes/implemented/architecture/2026-07-15-agent-initiator-scope.zh.md

@@ -0,0 +1,65 @@
+# Agent Note: 基于 AsyncLocalStorage 的发起 Agent 作用域
+
+Status: implemented
+
+[English](2026-07-15-agent-initiator-scope.md) | 中文
+
+## 问题
+
+Harness 中存在两种有用但不同的上下文概念。Cordis `Context` 负责选择服务、注册归属和生命周期；`agent.ctx` 是一个存活 Agent 所拥有的扁平注册作用域。Agent 与会话身份描述的则是异步操作主体。若把根 `ctx.agent` 改成「当前正在运行的 Agent」，就会混淆这两种含义，并在单进程并发驱动多个 Agent 时失效。
+
+进程内深层基础设施有时需要在显式传递的循环、工具及请求参数之下获取可信的发起 Agent，例如宿主感知传输层、追踪辅助函数、日志器或网关客户端。要求每个私有辅助函数都转发 `agent` 会造成重复，而进程级可变槽会在跨 `await` 时发生并发错误。模型可见参数也不适用，因为模型不得选择可信的会话或路由请求头。该载体归 Agent 服务所有，而非模型可见的可选上下文。
+
+## 决策
+
+必需的 `ctx.agents` 服务使用 Node `AsyncLocalStorage` 携带发起 Agent。它直接存储同一个 `Agent`，不引入只有一个字段的帧；另一个私有运行标记只记录嵌套边界的谱系，供 teardown 记账使用，不携带身份。[核心数据目录](../../../../docs/core-data-structures/core.md#initiating-agent)标明了所携带的类型。
+
+`currentInitiator()` 用于可选读取，`requireInitiator()` 抛出 `no initiating agent is active`，`withInitiator(agent, operation)` 保留操作返回的同步值或 Promise 本身。`withoutInitiator(operation)` 会建立清空边界，供不得继承 Agent 的工作使用。会话仍通过 `agent.session` 推导；轮次、步骤、工具调用、`signal`、模型、`cwd`、沙箱和授权继续由现有归属方管理。
+
+`AgentLoop` 已经注入 `ctx.agents`，并用 `agents.withInitiator(agent, ...)` 包裹每个具体驱动的完整 `runLoop` 生命周期。循环、轮次、步骤和工具调用的包内私有入口从 `ctx.agents` 恢复同一个 Agent，一次推导 `agent.session`，再由操作内辅助函数捕获该值，避免在浅层接口中转发具体驱动或 `Session`。若 `Session` 本身就是底层辅助函数的实际接口，该函数会保留狭窄的 `Session` 参数，而不会只为隐式查找而接收更宽泛的 `Context`。
+
+因此，并发驱动使用彼此独立的存储。子驱动的异步延续携带子 Agent；`withInitiator()` 返回后，调用方立即恢复之前的存储，而活动运行计数仍持续跟踪返回的 Promise，直到其结束。创建、持久化加载和尚未发布的 `setup(agentCtx)` 位于子驱动边界之外：由父 Agent 发起的创建使用父身份，而 `agentCtx.agent` 显式标识子 Agent。
+
+隐式身份不会取代显式契约。`ToolExecution.agent`、`AssembleContext.agent`、`GenerateOptions.sessionId`、任务归属、父子请求、`ctx.agent`、`agentCtx.agent`、审批与 hook 主体、`cwd` 选择、取消、worker 和进程消息、持久化记录及协议身份都保持显式传递。远程边界会把所需身份写入类型化请求，因为 ALS 只在进程内有效。
+
+`AgentRegistry` 管理一个有序的发起方生命周期。teardown 会先拒绝新边界；移除 `ctx.agents` 后，AgentLoop 等注入方开始排空，注册表随后等待活动的返回 Promise 边界，最后调用 `AsyncLocalStorage.disable()`。如果某个边界继承的异步调用链启动所属 Cordis fiber 的卸载，私有运行标记谱系会从排空范围中释放该嵌套边界链，从而避免 teardown 等待自身完成，同时继续排空无关边界。在普通排空期间，进行中代码可通过保留的服务引用继续调用 `currentInitiator()` 和 `requireInitiator()`；dispose 后，发起方方法会抛出 `agent initiator scope is disposed`。根 Context dispose 可能并发启动同级 fiber 的 teardown，因此除 Cordis 依赖顺序外仍必须统计活动边界。
+
+发起方作用域不负责管理脱离返回链的工作：注册表排空只跟踪 `withInitiator()` 或 `withoutInitiator()` 返回的 Promise。边界内创建的异步资源会继承其存储，直到自身结束或 ALS 被禁用；所属 seam 必须显式停止未纳入返回 Promise 的工作。Agent 所有前台工作会把完整生命周期纳入返回值，并保留显式取消契约。无关的定时器、队列和部署基础设施在 `withoutInitiator(operation)` 下启动；队列、worker、进程和协议边界必须序列化身份，不能期待 ALS 传播。
+
+宿主感知的传输层可以从 `ctx.agents.requireInitiator().session.id` 推导由部署方拥有的 `X-Harness-Session-Id` 等请求头；模型可见 schema 和参数中不包含该请求头。本决策不让现有生产 MCP 或 Web 传输层采用此请求头。测试替身传输层用于证明可信边界，而不会把宿主路由策略分配给现有的提供方无关 seam。
+
+本决策扩展 [Agent 注册作用域契约](2026-07-08-agent-scope-contexts.md)及其[运行时设计](2026-07-12-agent-scope-runtime-design.md)，不会改变其中 `agent.ctx` 的静态含义。
+
+## 验证
+
+Agent 服务测试锁定可选与必需读取、同步值和跨 realm Promise 的引用身份、内建 Promise 结束状态观察、并发、嵌套及清空边界、同步抛错或 Promise 拒绝后的恢复、普通与重入排空顺序及保留引用的错误。AgentLoop 集成测试锁定并发与嵌套驱动、无 Agent 调用、AgentRegistry 重启、根 Context 销毁，以及包内私有的循环和工具调度通过隐式查找完成。组合、模块图、构建及运行时闭包检查确保默认组合包、SDK 主干、Python 运行时闭包及直接 AgentLoop harness 通过 `ctx.agents` 完成接线，无需其他提供方。
+
+测试替身形式的宿主感知传输层在内部推导 `X-Harness-Session-Id`，并验证工具 schema 与记录参数都不包含身份字段。服务有意不排空边界操作所返回 Promise 之外的异步工作；这类工作仍由所属方的显式停止契约管理。
+
+## 考虑过的替代方案
+
+**在每个函数中传递 Agent。** 公开、worker、进程、持久化和协议边界继续显式传递，但要求每个进程内私有辅助函数都携带 Agent 只会造成重复转发，不会提高可信度。ALS 仅限于这些显式边界内部的异步调用链。
+
+**让 `ctx.agent` 变成动态值。** `ctx.agent` 已经表示与 Agent 作用域 Cordis 上下文静态关联的 Agent。改变根上下文的含义会混合注册作用域与执行作用域，并让并发行为变得意外。
+
+**新增独立的 `ctx.agentExecution` 服务。** 该载体没有独立后端、配置或身份类型：它存储的是 `ctx.agents` 已经管理的同一个 `Agent`，而 AgentLoop 本就依赖该服务。第二个必需提供方会增加包、组合、生命周期、生成目录及测试 harness 接线，却没有拆出真实能力。
+
+**保存命名帧或完整运行时帧。** 只有一个字段的 `{ agent }` 帧只是包装该值，而 Agent、会话、inbox、取消、轮次、步骤、工具执行和持久化已经有各自的真源。增加更多字段会产生陈旧快照和另一套生命周期；直接携带 `Agent`，由方法名标识边界，无需重复保存状态。
+
+**包含步骤级 `AbortSignal`、`cwd`、沙箱或授权。** 它们的生命周期及权限范围与驱动边界不一致，而且现有 seam 已经显式传递这些值。新增控制能力需要独立决策和嵌套生命周期契约。
+
+**使用进程级 `currentAgent`。** 并发 Agent 和 subagent 会在异步延续执行之间相互覆盖，因此可变全局值只在 Harness 不具备的串行保证下才正确。
+
+**从模型可见参数推导身份。** 不能信任模型或用户输入来选择会话、租户或沙箱路由。
+
+**给每个能力 seam 增加路由身份。** 这会把宿主关注点扩散到提供方无关 API。宿主感知实现拥有其传输请求头，而公开边界继续显式传递身份。
+
+## 后果
+
+深层基础设施可以获得一个可信的进程内发起 Agent，而无需加宽现有工具和能力请求。并发及嵌套驱动会自动隔离，AgentLoop 不增加新的必需服务，HMR 或根 Context dispose 会在禁用 ALS 前完成排空。
+
+该依赖不会出现在函数签名中，并且携带一个具有控制能力的 Agent 对象。消费方必须将其限制在横切基础设施中，把隐式存在视为既不证明存活、也不授予权限，并保留显式取消和归属检查。ALS 还有常驻传播成本，也无法跨越 worker、进程、HTTP 或持久化队列边界。
+
+该销毁设计有意依赖 Node 的 [Stability 1（实验性）](https://nodejs.org/api/async_context.html#asynclocalstoragedisable) API `AsyncLocalStorage.disable()`。Node 要求在 ALS 实例可被垃圾回收前调用 `disable()`，这对 HMR 替换 AgentRegistry 所拥有的实例尤为重要；服务状态守卫会阻止 dispose 后通过后续边界重新进入该实例。
+
+该作用域有意只携带 Agent，省略轮次、步骤、`signal`、`cwd`、沙箱和授权。若真实消费方无法使用现有显式字段，必须另行论证扩展；陈旧字段最多只能误标遥测数据，绝不能授予控制权。

.agents/notes/implemented/architecture/2026-07-15-llm-model-catalog-and-acp-selection.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-15-llm-model-catalog-and-acp-selection.md: 6cc8afc6c7431fbf3eb29fc358b432db4f72b529
+2026-07-15-llm-model-catalog-and-acp-selection.zh.md: 1cce7a58d0ec83dc01feaf72ccb61d294a78ddd5

.agents/notes/implemented/architecture/2026-07-15-llm-model-catalog-and-acp-selection.md

@@ -0,0 +1,66 @@
+# Agent Note: Advisory LLM catalogs and per-session ACP model selection
+
+Status: implemented
+
+English | [中文](2026-07-15-llm-model-catalog-and-acp-selection.zh.md)
+
+## Problem
+
+Provider-routed adapters let every request choose `provider + model`, but `LlmService` exposed only routing and streaming. A UI could not discover which providers were registered or which models an adapter was prepared to recommend. ACP clients therefore received no `model` session config option, so Zed, JetBrains, and VS Code integrations had no model list even though the request seam already supported runtime switching.
+
+Model discovery cannot become request validation. The hand-written DeepSeek adapter deliberately forwards arbitrary model ids to a public or private endpoint, while pi-ai has a finite installed catalog that is authoritative for its own request resolution. Treating one shared catalog as a whitelist would remove the private-endpoint behavior that provider routing was designed to preserve.
+
+ACP selection must also preserve the provider dimension. The same model id may appear under multiple routes, and switching a global adapter or agent template would leak one editor session's choice into every other session. Prompt variables and request routing must change together; a selection that lands during asynchronous prompt assembly cannot make `{{model}}` name one model while the request reaches another.
+
+## Decision
+
+### Provider-neutral advisory discovery
+
+`LlmAdapter` gains `providerInfo(provider)` and asynchronous `listModels(provider)` methods. Their provider-neutral results are `LlmProviderInfo { id, name }` and `LlmModelInfo { provider, id, name, description? }`. The defaults preserve existing adapter behavior by naming a provider after its route and advertising no models.
+
+`LlmService.listProviders()` returns detached metadata in registration order. `LlmService.listModels(provider)` delegates to the route owner, validates non-empty ids and names, rejects a mismatched provider or duplicate model id with `INVALID_CATALOG`, and returns detached values. Unknown providers still fail with `NO_ADAPTER`. Provider metadata is validated atomically during `registerAdapter()` so a malformed display record cannot leave a partial registration.
+
+Catalog membership is advisory. It drives selectors and diagnostics but never changes `stream()` routing and never rejects an otherwise valid request. Provider ownership remains exclusive and lifecycle-bound; model ids remain request-time adapter input.
+
+`dsh-llm-pi-ai` maps the configured provider's installed `getModels(provider)` entries into the neutral catalog. Its existing request-time catalog lookup remains authoritative and still rejects unknown models with `UNKNOWN_MODEL`. `dsh-llm-deepseek` accepts an optional `models` config containing display entries, defaulting to `deepseek-v4-flash` and `deepseek-v4-pro`. An explicit list replaces those defaults and an empty list disables discovery. The entries improve selector UX for known public or private models, while every unlisted model id continues to pass through unchanged.
+
+### ACP session config option
+
+The ACP bridge advertises one select with `id: model` and `category: model` in `session/new` and `session/load` when the session has a complete target whose provider is registered. Each opaque option value encodes the full provider/model pair. Models are grouped by provider when multiple non-empty provider groups exist; a single group is flattened for clients that render simple selects better.
+
+The session's current target is added to the displayed options when its adapter omits it. This preserves custom DeepSeek and private-endpoint models while keeping the adapter catalog advisory. A target with an unregistered provider is not advertised, and a model-less agent remains available to another `agent/request` supplier.
+
+`session/set_config_option` accepts only values from the current catalog snapshot and updates a target reference owned by that ACP session. No global `LlmService` or `AgentOptions` state changes, so concurrent sessions may select different providers and models. The existing permission select remains independent, and every response returns the complete refreshed option state.
+
+### Prompt/request consistency and durability
+
+Agent setup installs scoped `system-prompt/assemble` and `agent/request` listeners. Prompt assembly snapshots the selected pair once per step, overwrites the assembled `provider` and `model` variables after downstream prompt listeners, and the request listener applies that same snapshot after downstream request listeners. A selection during asynchronous assembly therefore starts on the next step rather than splitting prompt text from routing. Other call-config fields remain untouched.
+
+The request header remains the durable source of truth. When a selected target is actually used, the existing full `request/header` snapshot records it. `session/load` initializes the ACP selection from the folded last request header before falling back to bridge config. A selection that is never used by a request is intentionally in-memory only because it never became model-visible state.
+
+ACP's experimental `providers/*` capability is not used. That draft surface configures provider base URLs, protocols, and headers, including secrets; it does not enumerate models and would give the UI authority to rewrite deployment-owned adapter configuration.
+
+## Alternatives considered
+
+**Return model strings only.** A model-only value loses the provider route and becomes ambiguous as soon as two providers expose the same id.
+
+**Make catalogs mandatory whitelists.** This conflicts with the hand-written adapter's arbitrary model pass-through and private deployments. The selected adapter already owns authoritative request validation.
+
+**Store selection in `AgentOptions` or `LlmService`.** Those are creation-wide or deployment-wide objects. Mutating them would couple concurrent ACP sessions and bypass the logged `agent/request` replacement path.
+
+**Persist a new model-selection session event immediately.** An unused UI selection has not affected a model request. Recording the existing request header when the target is consumed preserves the model-visible-if-and-only-if-logged rule without adding a second source of truth.
+
+**Use ACP `providers/*`.** That unstable API changes endpoint and authentication configuration rather than selecting a model for one session, and its lifecycle and secret-handling semantics do not match this feature.
+
+## Consequences
+
+- Any adapter can expose a dynamic model list without leaking provider-library types into the core seam.
+- Catalog consumers must treat absence as “not advertised,” never “invalid request.”
+- pi-ai-backed ACP deployments automatically inherit the installed pi-ai provider catalogs; hand-written DeepSeek deployments list known choices explicitly and retain arbitrary model support.
+- ACP clients receive a standard stable model config option, with provider-aware values and per-session isolation.
+- Request headers remain compatible with the provider-routed session shape; no new JSONL event or format version is required.
+- A catalog read can be asynchronous. ACP reads a detached snapshot before creating or resuming an agent, so discovery failure cannot leave a partially published session.
+
+## Testing
+
+Unit coverage validates catalog detachment and malformed metadata, pi-ai and DeepSeek catalog projection, ACP provider grouping, custom-current insertion, invalid values, provider/model request routing, prompt-variable alignment, concurrent-session isolation, model-less fallback, and load restoration from the request header. The existing ACP transport suites verify that the additional config option does not change prompt, cancellation, replay, approval, or tool-rendering behavior.

.agents/notes/implemented/architecture/2026-07-15-llm-model-catalog-and-acp-selection.zh.md

@@ -0,0 +1,66 @@
+# Agent Note: 建议性 LLM 目录与 ACP 会话级模型选择
+
+Status: implemented
+
+[English](2026-07-15-llm-model-catalog-and-acp-selection.md) | 中文
+
+## 问题
+
+基于提供方路由的适配器允许每次请求选择 `provider + model`，但 `LlmService` 只暴露路由和流式调用。UI 无法发现已注册的提供方，也无法知道适配器愿意推荐哪些模型。因此，ACP 客户端收不到 `model` 会话配置项；即使请求接缝已经支持运行时切换，Zed、JetBrains 和 VS Code 集成仍没有模型列表。
+
+模型发现不能变成请求校验。手写 DeepSeek 适配器会把任意模型 ID 原样转发给公开或私有端点，而 pi-ai 的有限安装目录则是其自身请求解析的权威依据。将共享目录视为白名单，会破坏提供方路由需要保留的私有端点能力。
+
+ACP 选择还必须保留提供方维度。同一个模型 ID 可能存在于多个路由下；切换全局适配器或 agent 模板会让一个编辑器会话的选择泄漏到其他会话。Prompt 变量与请求路由必须同时变化；如果选择发生在异步 prompt 组装期间，不能让 `{{model}}` 表示一个模型、实际请求却到达另一个模型。
+
+## 决策
+
+### 提供方中立的建议性发现
+
+`LlmAdapter` 增加 `providerInfo(provider)` 与异步 `listModels(provider)` 方法。其提供方中立结果分别为 `LlmProviderInfo { id, name }` 和 `LlmModelInfo { provider, id, name, description? }`。默认实现以路由名称作为提供方名称，并且不展示模型，从而保持现有适配器行为。
+
+`LlmService.listProviders()` 按注册顺序返回分离后的元数据。`LlmService.listModels(provider)` 委托给路由所有者，校验非空 ID 和名称，并在提供方不匹配或模型 ID 重复时以 `INVALID_CATALOG` 失败，最后返回分离后的值。未知提供方仍以 `NO_ADAPTER` 失败。提供方元数据在 `registerAdapter()` 期间进行原子校验，错误展示记录不会留下部分注册。
+
+目录成员关系仅提供建议。它驱动选择器与诊断，但不会改变 `stream()` 路由，也不会拒绝原本有效的请求。提供方所有权仍然具有排他性并绑定生命周期；模型 ID 仍是请求时传给适配器的输入。
+
+`dsh-llm-pi-ai` 将已配置提供方的安装目录 `getModels(provider)` 映射为中立目录。其现有请求时目录查询仍是权威依据，未知模型仍以 `UNKNOWN_MODEL` 失败。`dsh-llm-deepseek` 接受可选的 `models` 配置作为展示条目，默认包含 `deepseek-v4-flash` 和 `deepseek-v4-pro`。显式列表会替换这些默认值，空列表则关闭发现。这些条目改善已知公开或私有模型的选择体验，而所有未列出的模型 ID 仍会原样透传。
+
+### ACP 会话配置项
+
+当会话具有完整目标且目标提供方已注册时，ACP bridge 会在 `session/new` 与 `session/load` 中展示一个 `id: model`、`category: model` 的选择项。每个不透明选项值都编码完整的提供方/模型字段组合。存在多个非空提供方分组时按提供方分组；只有一个分组时将其展开，以便对简单选择器支持更好的客户端展示。
+
+如果适配器目录未包含会话当前目标，该目标仍会加入展示选项。这能保留自定义 DeepSeek 与私有端点模型，同时维持目录的建议性。提供方未注册的目标不会展示；缺少模型的 agent 仍可由其他 `agent/request` 提供者补齐。
+
+`session/set_config_option` 只接受当前目录快照中的值，并更新该 ACP 会话独占的目标引用。它不会修改全局 `LlmService` 或 `AgentOptions` 状态，因此并发会话可以选择不同的提供方和模型。现有权限选择项保持独立，每次响应都返回完整的刷新后配置项状态。
+
+### Prompt/请求一致性与持久化
+
+Agent setup 会安装作用域内的 `system-prompt/assemble` 与 `agent/request` 监听器。Prompt 组装为每个 step 只快照一次选中的字段组合，在下游 prompt 监听器完成后覆盖组装结果中的 `provider` 与 `model` 变量；请求监听器则在下游请求监听器完成后应用同一个快照。因此，异步组装期间发生的选择会从下一个 step 生效，不会导致 prompt 文本与路由分裂。其他调用配置字段保持不变。
+
+请求头仍是持久化事实来源。当选中目标被实际使用时，现有的完整 `request/header` 快照会记录它。`session/load` 先从折叠后的最后请求头初始化 ACP 选择，再回退到 bridge 配置。一个从未被请求使用的选择只保留在内存中，因为它从未成为模型可见状态。
+
+本功能不使用 ACP 的实验性 `providers/*` 能力。该草案接口配置提供方 base URL、协议和 headers，其中可能包含密钥；它不枚举模型，并且会赋予 UI 改写部署所有的适配器配置的权力。
+
+## 考虑过的替代方案
+
+**只返回模型字符串。** 仅模型值会丢失提供方路由；两个提供方暴露相同 ID 时立刻产生歧义。
+
+**将目录设为强制白名单。** 这与手写适配器的任意模型透传和私有部署冲突。请求的权威校验本就属于被选中的适配器。
+
+**将选择存入 `AgentOptions` 或 `LlmService`。** 这些对象分别面向创建过程或整个部署。修改它们会耦合并发 ACP 会话，并绕开带日志归因的 `agent/request` 替换路径。
+
+**立即写入新的模型选择会话事件。** 尚未使用的 UI 选择没有影响模型请求。目标被消费时记录现有请求头，既满足“模型可见当且仅当已记录”的规则，也不会引入第二个事实来源。
+
+**使用 ACP `providers/*`。** 该不稳定 API 用于修改端点与认证配置，而不是为单个会话选择模型；其生命周期和密钥处理语义都不适合本功能。
+
+## 结果
+
+- 任意适配器都能暴露动态模型列表，无需把提供方库类型泄漏到核心接缝。
+- 目录消费者必须把缺失理解为“未展示”，而不是“请求无效”。
+- 基于 pi-ai 的 ACP 部署会自动继承已安装的 pi-ai 提供方目录；手写 DeepSeek 部署显式列出已知选项，同时保留任意模型能力。
+- ACP 客户端会收到稳定标准的模型配置项，其中的值保留提供方信息，并按会话隔离。
+- 请求头继续使用基于提供方路由的会话结构；不需要增加 JSONL 事件或格式版本。
+- 目录读取可以是异步的。ACP 在创建或恢复 agent 前读取分离后的快照，因此发现失败不会留下部分发布的会话。
+
+## 测试
+
+单元测试覆盖目录分离与错误元数据、pi-ai 和 DeepSeek 目录投影、ACP 提供方分组、自定义当前模型补入、无效值、提供方/模型请求路由、prompt 变量一致性、并发会话隔离、无模型回退，以及从请求头恢复选择。现有 ACP 传输测试验证新增配置项不会改变 prompt、取消、回放、审批或工具展示行为。

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-15-replay-token-meter-service.md: 9bbc177f456e006179c466f8c245e4599db3dd5a
+2026-07-15-replay-token-meter-service.zh.md: 4437626c8651a80537d45197a93733271a592173

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.md

@@ -0,0 +1,60 @@
+# Agent Note: Replay token meter service
+
+Status: implemented
+
+English | [中文](2026-07-15-replay-token-meter-service.zh.md)
+
+## Problem
+
+Context pressure is useful outside compaction. A compaction backend, an overflow guard, or a future request-policy plugin can all need the same answer: how much of the configured context window does the durable request consume? Keeping that fold inside `dsh-compact-basic` duplicates replay logic, makes measurement unavailable without compaction, and encourages callers to reuse stale accounting.
+
+Provider usage is not a complete answer. It describes one successful call under one exact request envelope, while the current surface can grow, shrink, or be replaced afterward. Sessions also switch providers and models, old logs can lack chunk provenance, and usage fields separate input, cache-read, cache-write, output, and reasoning counts. A useful service therefore combines the latest exact anchor with conservative heuristic repricing and exposes the log revision consumed by each result.
+
+## Decision
+
+### One concrete LLM-family service
+
+`@deepseek-ai/dsh-token-meter` is one concrete package under `packages/llm/` and registers `ctx.tokenMeter`. It is not split into an interface and backend before a second implementation exists. `TokenMeterService` itself exposes `contextWindow`, `measure(session, requestHeader?)`, and `estimateMessage(message)`; consumers call the singleton service directly.
+
+The service has one `contextWindow`, defaulting to 128,000 tokens and configurable as a positive integer. Estimation uses a fixed four-characters-per-token heuristic plus structural overhead. There are no model profiles, density settings, tokenizer backends, or language-specific strategies.
+
+### Per-session replay folds
+
+Each session owns one isolated incremental fold. Active folds advance from `session/event`; every read catches up through the durable tail, so listener ordering, seeded sessions, and service reload do not change the answer. The fold tracks canonical full request-header snapshots, step boundaries, surface appends and replacements, assistant usage, and assistant-chunk provenance. A malformed next event fails transactionally and remains unread rather than partially mutating state.
+
+`measure(session, requestHeader?)` synchronizes the fold once and returns scalar pressure together with positional per-node prices. `totalTokens` remains request-and-response pressure; `surfaceTokens` is the surface-only heuristic total and equals the sum of `nodes[].tokens`. A `requestHeader` override changes pressure pricing only, while the surface fields always describe the current session. `estimateMessage(message)` applies the fixed heuristic without session state. Each result is one detached, deeply immutable snapshot carrying one `logRevision`. Every measurement clones the current nodes and is therefore O(surface).
+
+Provider usage is reused only when the measured canonical request envelope equals the latest successful-call anchor. Any provider, model, system, prefix, tool, or call-config change causes complete heuristic repricing. Surface changes remain a signed delta from a matching anchor, including negative values after a shrinking replacement. A later successful request replaces the earlier anchor, including across provider or model switches.
+
+Usage sums the disjoint input, cache-read, cache-write, and output buckets. Reasoning is not added a second time. Every successful model call records an `assistant/message`, including content-less and max-token calls, with its exact earlier chunk seqs. An explicit empty provenance list means a known empty provider stream; absent legacy provenance conservatively treats the durable assistant output as provider output.
+
+### Compact-basic consumes, but does not own, measurement
+
+`dsh-compact-basic` requires `ctx.tokenMeter`; `CompactService` gains no token methods or types. Configuration, the region transaction, and summarization stay in separate modules; the service registers automatic listeners itself, while `summarize()` remains its sole subclass hook. The singleton meter consistently prices pressure, retention, shadowed content, provenance, and non-shrinking-summary rejection.
+
+Automatic compaction uses one unified measurement for each threshold-and-retention decision. The region transaction measures after appending its durable `compact/start` lock and again after asynchronous summarization; any intervening durable append changes `logRevision` and prevents replacement.
+
+Compact policy has service-wide defaults: threshold ratio `0.8`, retained tail `floor(contextWindow × 0.16)`, `summarizationProvider: ''`, `summarizationModel: ''`, `maxTokens: 8192`, `compactionRetries: 1`, `maxOverflowRetries: 1`, and `auto: true`. Top-level `thresholdRatio` and `retainTokens` override the pressure policy; retention must remain below the resulting threshold. The summarization provider and model must both be set or both be empty; an empty pair resolves the latest logged request target, then the `AgentOptions` pair.
+
+Automatic pressure runs at `agent/post-step` and measures the canonical durable envelope produced under the provider/model actually selected by `agent/request`. A headerless session has no completed routed request to assess and produces no work; any routed target can use the singleton estimator. Canonical overflow recovery uses the same measurement for forced range selection and retries only after a proven surface replacement.
+
+## Testing
+
+Unit tests cover fixed estimation, envelope invalidation and anchor replacement, replay boundaries, immutable snapshots, routed pressure, convergence, overflow generation proof, and rollback. A real Loader/Include fixture verifies the zero-config token-meter and compact-basic load path in dependency order.
+
+## Alternatives considered
+
+- **Keep estimation inside `CompactService`** — rejected because measurement has consumers and replay semantics independent of compaction; it would also force every compactor to expose the same unrelated API.
+- **Split a token-meter interface from a heuristic backend immediately** — rejected because only one implementation exists. One concrete service preserves the future seam without speculative packages or configuration.
+- **Keep model-keyed windows and density profiles** — rejected because the deployment currently has one context policy and one estimator. Model registries, unknown-model failures, and configurable density add branches without a second behavior to select.
+- **Keep separate scalar and surface measurements** — rejected because callers would need two reads and revision matching for one decision. A scalar-only read could avoid cloning nodes below threshold, but the split API introduces a caller-side race window; the unified snapshot accepts O(surface) cloning in exchange for coherence.
+- **Treat provider usage as portable between envelopes** — rejected because model, tools, prefixes, and call config are request facts. Mismatch reprices the whole current request.
+
+## Consequences
+
+- Token pressure has one replay-aware owner that compaction and future plugins can share.
+- The default makes the bundled composition usable with two zero-config plugin entries; deployments override one context capacity when needed.
+- Fixed heuristic pricing remains an estimate of provider behavior and is not an exact tokenizer or request serializer.
+- Every measurement clones the current positional surface and therefore costs O(surface), including pressure checks that finish below threshold.
+- Measurements fail loudly on malformed durable boundaries. This turns corrupted replay into a named integration failure instead of silently drifting pressure.
+- Post-step pressure reads the exact logged routing/tools/prefix boundary; provider overflow classification remains the adapter-maintained backstop for requests rejected before a successful usage anchor.

.agents/notes/implemented/architecture/2026-07-15-replay-token-meter-service.zh.md

@@ -0,0 +1,60 @@
+# Agent Note: 回放式 token 计量服务
+
+Status: implemented
+
+[English](2026-07-15-replay-token-meter-service.md) | 中文
+
+## 问题
+
+上下文压力并不只对压缩有用。压缩后端、溢出保护或未来的请求策略插件都可能需要回答同一个问题：持久请求占用了已配置上下文窗口的多少容量？如果把该折叠逻辑留在 `dsh-compact-basic` 内部，就会重复实现回放逻辑，使未加载压缩的调用方无法使用计量，并诱使调用方复用陈旧的核算结果。
+
+提供方 usage 也不是完整答案。它只描述某个精确请求信封下的一次成功调用，而当前表层之后还可能增长、缩小或被替换。会话也可能切换提供方与模型，旧日志可能缺少分片来源，usage 字段还会分别报告输入、缓存读取、缓存写入、输出与推理计数。因此，可用的服务必须把最新精确锚点与保守的启发式重新定价结合起来，并公开每个结果已经消费的日志修订号。
+
+## 决策
+
+### 一个具体的 LLM 家族服务
+
+`@deepseek-ai/dsh-token-meter` 是 `packages/llm/` 下的单个具体包，并注册 `ctx.tokenMeter`。在第二种实现出现之前，它不会被拆成接口与后端。`TokenMeterService` 本身公开 `contextWindow`、`measure(session, requestHeader?)` 与 `estimateMessage(message)`；消费方直接调用这个单例服务。
+
+服务只有一个 `contextWindow`，默认值为 128,000 token，并允许配置为正整数。估算采用固定的每 token 四个字符启发式规则，并加上结构开销。服务不提供模型 profile、密度设置、分词器后端或语言专用策略。
+
+### 逐会话回放折叠
+
+每个会话都有一个隔离的增量折叠。活跃折叠通过 `session/event` 前进；每次读取都会追到持久日志尾部，因此监听器顺序、种子会话与服务重载不会改变答案。折叠跟踪规范的完整请求头快照、步骤边界、表层追加与替换、assistant usage，以及 assistant 分片来源。下一个畸形事件会以事务方式失败并保持未读，不会让状态只修改一半。
+
+`measure(session, requestHeader?)` 只同步一次折叠，并在返回标量压力的同时给出逐位置节点价格。`totalTokens` 仍表示请求与响应压力；`surfaceTokens` 是仅针对表层的启发式总量，并等于 `nodes[].tokens` 之和。`requestHeader` 覆盖只改变压力定价，表层字段始终描述当前会话。`estimateMessage(message)` 不依赖会话状态，直接应用固定启发式规则。每个结果都是一个分离且深度不可变的快照，只携带一个 `logRevision`。每次计量都会复制当前节点，因此成本为 O(surface)。
+
+只有当待计量的规范请求信封等于最近一次成功调用的锚点时，服务才复用提供方 usage。提供方、模型、系统提示词、前缀、工具或调用配置任一变化都会触发完整的启发式重新定价。表层变化相对匹配锚点保留有符号增量，包括缩小替换后的负值。后续成功请求会替换先前锚点，提供方或模型切换时也一样。
+
+Usage 会对互不重叠的输入、缓存读取、缓存写入与输出 bucket 求和，不会再次加入推理计数。每次成功模型调用都会记录 `assistant/message`，包括无内容调用与达到 token 上限的调用，并带上精确的更早分片 seq。显式空来源列表表示已知为空的提供方流；旧日志中缺失的来源则保守地把持久 assistant 输出视为提供方输出。
+
+### compact-basic 消费计量，但不拥有计量
+
+`dsh-compact-basic` 要求 `ctx.tokenMeter`；`CompactService` 不增加 token 方法或类型。配置、区域事务与摘要器分别保留在独立模块中，服务自身注册自动监听器，而 `summarize()` 仍是唯一的子类 hook。单例计量器一致用于压力、保留、被遮蔽内容、来源以及非缩小摘要拒绝的定价。
+
+自动压缩的每次阈值与保留联合决策只使用一次统一计量。区域事务先追加持久 `compact/start` 锁，再执行一次计量，并在异步摘要完成后再次计量；期间任何持久追加都会改变 `logRevision`，从而阻止替换。
+
+压缩策略采用服务级默认值：阈值比例 `0.8`、保留尾部 `floor(contextWindow × 0.16)`、`summarizationProvider: ''`、`summarizationModel: ''`、`maxTokens: 8192`、`compactionRetries: 1`、`maxOverflowRetries: 1` 与 `auto: true`。顶层 `thresholdRatio` 与 `retainTokens` 覆盖压力策略；保留值必须小于最终阈值。摘要提供方与模型必须同时设置或同时为空；空组合先解析最近记录的请求目标，再使用 `AgentOptions` 中的组合。
+
+自动压力检查运行在 `agent/post-step`，并计量 `agent/request` 实际所选提供方/模型产生的规范持久信封。没有请求头的会话尚无已完成的路由请求可供判断，因此不执行工作；任意路由目标都可使用这个单例估算器。规范化溢出恢复使用同一计量结果强制选择范围，并且只有在表层替换得到证明后才重试。
+
+## 测试
+
+单元测试覆盖固定估算、信封失效与锚点替换、回放边界、不可变快照、已路由压力、收敛、溢出 generation 证明与回滚。真实 Loader/Include fixture 验证零配置 token-meter 与 compact-basic 按依赖顺序加载的路径。
+
+## 考虑过的替代方案
+
+- **把估算保留在 `CompactService` 内**——不予采纳，因为计量拥有独立于压缩的消费方与回放语义；它还会强迫每个压缩器暴露同一套无关 API。
+- **立即把 token meter 拆成接口与启发式后端**——不予采纳，因为目前只有一种实现。单个具体服务保留未来接缝，同时避免推测性的包与配置。
+- **保留模型键控的窗口与密度 profile**——不予采纳，因为当前部署只有一种上下文策略与一个估算器。模型注册表、未知模型错误和可配置密度只增加分支，却没有第二种行为可供选择。
+- **保留独立的标量与表层计量**——不予采纳，因为消费方必须为一次决策执行两次读取并匹配修订号。仅读取标量可以避免在低于阈值时复制节点，但拆分 API 会在消费方引入竞态窗口；统一快照接受 O(surface) 复制成本，以换取结果一致性。
+- **在不同信封之间移用提供方 usage**——不予采纳，因为模型、工具、前缀与调用配置都是请求事实。不匹配时会重新定价完整当前请求。
+
+## 后果
+
+- Token 压力拥有一个可供压缩与未来插件共享的回放感知所有者。
+- 默认值让内置组合只需两个零配置插件条目即可使用；部署需要时只覆盖一个上下文容量。
+- 固定启发式定价仍然只是提供方行为的估计，并不是精确分词器或请求序列化器。
+- 每次计量都会复制当前的位置表层，因此成本为 O(surface)，低于阈值即可结束的压力检查也不例外。
+- 遇到畸形持久边界时，计量会明确失败。这会把损坏的回放转化为具名集成错误，而不是让压力静默漂移。
+- post-step 压力检查读取精确记录的路由、工具与前缀边界；对于在成功 usage 锚点出现前就被拒绝的请求，提供方溢出分类仍是由适配器维护的兜底路径。

.agents/notes/implemented/feature/2026-06-14-acp-agent-client-protocol.md

@@ -0,0 +1,57 @@
+# Agent Note: Agent Client Protocol (ACP) support — drive the coding agent from external editors
+
+Status: implemented
+
+## Problem
+
+The harness originally exposed agents only through a readline loop. That surface could carry text, but it gave an editor no structured way to create or resume sessions, correlate prompt completion, stream reasoning and tool activity, render tool-specific UI, ask for permission, or cancel one conversation without disturbing another. ACP defines those interactions as JSON-RPC over stdio, and Zed is the target client used to make concrete compatibility decisions.
+
+The bridge must preserve the harness's existing ownership boundaries. It cannot depend on the concrete agent loop, bypass the tool registry, execute shell commands in the editor, or invent a second source of session truth. stdout is also the protocol transport, so any accidental log output corrupts the connection.
+
+## Decision
+
+`@deepseek-ai/dsh-acp` is a UI/client-driver plugin under `packages/ui/acp`. It uses `@agentclientprotocol/sdk`'s `AgentSideConnection` over stdin/stdout and programs only interface services: the agent create/resume factory, session persistence, tool registry, user interaction, and optional approval/bash capabilities. It does not change the agent loop and is not a capability-seam implementation.
+
+The bridge implements the following stable session path:
+
+- `initialize` negotiates the protocol version, advertises text plus `resource_link` prompts, and advertises `loadSession`.
+- `session/new` validates an absolute `cwd`, stores it in `SessionHeader`, creates an agent through `ctx.agents`, and returns any composition-backed config options.
+- `session/load` validates the requested cwd against persisted metadata before constructing an agent, reserves the id across the asynchronous resume, replays user/assistant/tool events as ACP updates, and reports the resumed config-option fold.
+- `session/prompt` accepts text and resource links, rejects unsupported or empty content, allows one in-flight prompt per session, and settles against that prompt's owning `turn/end`. An error turn rejects the RPC; other closed turn reasons map through a total ACP stop-reason codec.
+- `session/cancel` calls the queue-aware agent cancel path and settles only the addressed session's prompt.
+
+Tool-call presentation remains tool-owned. A tool's `presentCall` and `presentResult` return the `generic`, `terminal`, or `diff` render-intent variants; the bridge switches on that union and maps it to ACP. Presenter-less tools receive a generic fallback. Bash terminal cards use Zed's capability-gated `_meta.terminal_info`, `_meta.terminal_output`, and `_meta.terminal_exit` convention; the harness still executes the command through `ctx.bash`, preserving sandbox, environment scrub, ownership, and cwd. Clients without that extension receive ordinary text content. Filesystem tools provide diff cards and file locations without hard-coded tool-name branches in the bridge.
+
+Permission handling is an answerer on the [user-approval seam](2026-07-06-approval-seam.md), not an ask-every-tool policy in ACP. An `approval/request` for a bridge-owned agent with a call id becomes `session/request_permission` on that agent's editor session, with one-shot allow/reject choices. Foreign or call-less requests delegate; a missing or failed answerer remains fail-closed. The plugin that asks—such as a pre-execute policy or bash escalation—owns the decision to ask.
+
+When `ctx.permission` is composed, the bridge exposes one `permission` select from the deployment's preset table. The shipped `workspace-write` and `danger-full-access` presets each bundle a sandbox mode with an approval policy; unmatched effective knobs produce the switch-away-only `custom` state. `session/set_config_option` validates through `PermissionService.set()` and writes both owning knob events. A switch during an open turn appends immediately; an idle switch is overlaid in responses and anchored at the next `agent/prompt-submit`, before request assembly. Until then it is memory-only, so a crash restores the durable fold. ACP session modes are not modeled because config options are the forward protocol surface; `AcpConfig.model` remains connection-wide.
+
+The bridge also provides the ACP-backed `UserInteractionProvider`: `ask_user_question` requests become form elicitations on the owning session. Select, multi-select, option descriptions, and custom-answer override semantics are preserved.
+
+Lifecycle ownership is explicit. The bridge holds an `AgentHandle` per live session. Disconnect and Cordis disposal cancel pending prompts, dispose every handle in parallel, await loop quiescence and persistence flush, and then remove the records. Stream notification failures are contained so a vanished client cannot corrupt an agent turn. The ACP app composition loads no stdout logger; a test guards stdout as framed JSON-RPC only.
+
+The precise supported and deferred protocol rows live in [`packages/ui/acp/acp-feature-support.md`](../../../../packages/ui/acp/acp-feature-support.md); the package README is the operational contract.
+
+## Alternatives considered
+
+**A prepended `tools/execute` listener that asks on every ACP-owned call** — rejected. It would hard-code permission policy into the UI bridge, ask even when no policy requires it, and could not serve approval requests that arise after execution begins. The shared user-approval seam keeps mechanism, asking policy, and UI answerer separate.
+
+**Inject the concrete `agentLoop`** — rejected. Agent creation, resume, idle observation, and disposal are interface-level ownership operations on `dsh-agent`; a UI plugin does not need a dependency-rule exception.
+
+**Execute bash through ACP `terminal/*`** — rejected. That would move execution outside the harness and bypass its sandbox, credential scrub, task ownership, cwd resolution, and session log. Terminal metadata is presentation only.
+
+**Represent permission presets as ACP session modes** — rejected. The deployment-defined preset is already one config-option select, while session modes are the legacy surface slated for removal in ACP v2.
+
+**Hijack stdout defensively** — rejected. Process-wide monkey-patching is outside Cordis effect ownership and races the protocol transport. The app composition owns stdout purity.
+
+## Consequences
+
+Editors can create, load, prompt, cancel, render, ask, and reconfigure multiple harness sessions over one ACP connection without a loop-specific dependency. The session event log remains the durable source for replay, prompt settlement, cwd, and per-session configuration. Tool presentation and human-answer channels remain extensible plugin contracts instead of ACP-specific behavior.
+
+The bridge deliberately does not implement session list/delete/resume/close capabilities, MCP passthrough, additional directories, image/audio/embedded-resource prompts, plans, slash commands, usage updates, editor filesystem delegation, or the ACP terminal execution sub-protocol. Runtime model selection was added later through standard session config options by the [LLM catalog and ACP selection Agent Note](../architecture/2026-07-15-llm-model-catalog-and-acp-selection.md).
+
+An idle config selection is truthful in the live response but not durable until the next `agent/prompt-submit` anchors it inside the open turn. Crashing before that boundary loses the pending selection; this is the cost of keeping session events turn-enclosed and replay-safe.
+
+## Verification
+
+The ACP suites cover the in-memory protocol codec, create/load replay, exact prompt settlement, cancellation races, unsupported content, tool presentation, terminal capability fallback, permission outcome mapping, config-option validation and persistence, multi-session isolation, disconnect/disposal quiescence, and HMR cleanup. Snapshot and built-bin tests exercise the app composition, while the real-API e2e self-skips without a key.

.agents/notes/implemented/feature/2026-06-14-acp-multi-session.md

@@ -0,0 +1,37 @@
+# Agent Note: Multiplex concurrent ACP sessions over one connection
+
+Status: implemented
+
+## Problem
+
+An ACP editor can keep several conversations alive over one agent subprocess. A single-active-session bridge would force extra processes and would not match Zed's client model, which tracks multiple session ids and concurrent loads. Multiplexing introduces isolation risks: events, prompt completion, cancellation, permission prompts, config selections, and predictable background-task ids must never cross session boundaries.
+
+## Decision
+
+The ACP bridge stores live sessions in `Map<SessionId, SessionRecord>`. Agent-scoped callbacks use `ownedRecord`: look up `agent.session.id` in that forward map and accept the record only when it owns the exact agent object, so a foreign same-id object cannot claim the session. A record owns its agent handle, in-flight prompt, live tool-call presentation state, pending idle config switches, session cwd, and client capability snapshot. A separate loading-id set reserves each id before asynchronous resume so two pipelined loads cannot construct duplicate agents; distinct ids may load concurrently.
+
+Every `session/event` and `agent/status` callback resolves the owning record before sending or settling anything. Each session permits one in-flight prompt independently. The prompt records a log watermark, captures its own `turn/start`, and settles only on the matching `turn/end`; a late end from a cancelled prior turn cannot resolve a newer prompt. `session/cancel` addresses one record and calls only that agent's queue-aware cancel path.
+
+Permission ownership uses the same exact-agent check against the forward map. The ACP `approval/request` answerer prompts only the editor session that owns the requesting agent and delegates foreign requests. User-interaction elicitations likewise route by agent ownership. Per-session sandbox and approval config values fold only that session's events, with pending idle switches stored on that record until the next turn anchors them.
+
+Background bash tasks carry an opaque owner token equal to the owning session id. `bash_output` and `bash_kill` compare the caller's token with the executor's task ownership before reading or killing; a predictable task id alone grants no access. Ownership is stored with the executor task, so a tool plugin reload does not erase it.
+
+Connection teardown clears the live map, settles each pending prompt as cancelled, and disposes all `AgentHandle`s in parallel. Each handle stops and awaits its loop, flushes the session while attached, unregisters the agent, and removes the session. Teardown is memoized and shared by client disconnect and plugin disposal.
+
+## Alternatives considered
+
+**One live session per connection** — rejected. It adds process overhead and contradicts the target client's multi-session shape without removing multiplexing needs from the editor.
+
+**A per-session `ctx.extend()`** — rejected. A child context does not by itself create a child plugin fiber, so listeners would still belong to the bridge fiber. The implemented bridge instead uses global listeners with explicit O(1) demultiplexing and per-session owned records; agent lifecycle is owned by `AgentHandle`.
+
+**Agent object identity as bash-task ownership** — rejected. A resumed or replaced agent object may legitimately represent the same durable session. The opaque session token is the cross-boundary identity that should survive plugin reloads.
+
+## Consequences
+
+N sessions can stream, prompt, request permission, switch config, and run background tasks concurrently without interleaving or cross-settling. A cancel or dispose in one session does not affect its neighbors. The bridge pays for explicit maps and isolation tests, but it does not add one listener set per session and therefore avoids listener fan-out during long-lived connections.
+
+The bridge still exposes no protocol method to close one live session independently. Today records leave together on connection teardown; session close/resume lifecycle capabilities remain deferred in the ACP feature checklist.
+
+## Verification
+
+The multi-session suite drives concurrent sessions through interleaved updates, independent in-flight prompts, targeted cancellation, same-id and distinct-id load races, permission routing, config isolation, and teardown. Tool-bash tests prove one session cannot read or kill another session's background task.

.agents/notes/implemented/feature/2026-06-15-code-mode.md

@@ -0,0 +1,130 @@
+# Agent Note: Code Mode — the model writes TypeScript against the tool registry
+
+Status: implemented
+
+## Problem
+
+In the registry's native presentation, the agent loop advertises every visible capability as a JSON-schema function definition. `ToolRegistry` contributes its schemas to the system-prompt assembly, the assembly's `tools` land on the wire (and in the logged request header), the model invokes one `tool-call` block per step, and the loop dispatches each call through `ctx.tools.execute()` **sequentially** (parallel tool execution is an explicit open TODO in `dsh-tools` and [docs/architecture.md](../../../../docs/architecture.md)), with **every** intermediate `tool-result` re-entering the model's context on the next request.
+
+For multi-step tool work this is token-heavy and serial. The model cannot compose tools — loop over a result set, branch on an intermediate value, fan out, post-process — without a full model round-trip per call, and each round-trip drags the entire intermediate result back into context whether the model needs it or not.
+
+Cloudflare's [Code Mode](https://blog.cloudflare.com/code-mode/) proposes an alternative grounded in a simple observation: LLMs are better at writing code than at emitting tool calls, because they have seen millions of lines of real code and comparatively few contrived tool-calling traces. Instead of one tool call per step, the model writes a TypeScript program against a generated API over the tools, the program executes in a sandboxed runtime, and the model curates what comes back — only what it prints or returns — instead of every intermediate result.
+
+Tool presentation belongs to the registry that owns tool visibility: implementing a second presentation as an after-the-fact waterfall transform would make correctness depend on listener order and fight [reconstructable requests](../architecture/2026-07-05-reconstructable-requests.md). The execution substrate is also part of the foundation rather than a placeholder: Node `worker_threads` provides a separate isolate, an empty environment, heap caps, and termination of a hot synchronous loop, while fitting the harness's existing trust model (§Trust posture).
+
+## Decision
+
+Three decisions, each elaborated in its own section below:
+
+1. **Code Mode is a first-class presentation mode of `ToolRegistry`** (`dsh-tools`), selected by a validated `mode` config: `'native'` (the default, contributing the visible capability schemas), `'code'` (the registry contributes only its reserved `run_code` transport plus a generated SDK `.d.ts` in the system prompt), or `'both'` (native schemas and the transport + SDK). The registry shapes its canonical contribution at the source; the cooperative prompt-assembly result remains authoritative, and the logged request header records exactly that returned presentation.
+2. **Code execution is a capability seam** — `packages/code-runtime/` contains the interface package `@deepseek-ai/dsh-code-runtime`, which owns `ctx.codeRuntime` ([capability seams](../architecture/2026-06-13-capability-seams.md); consumer = `dsh-tools`, with core-consumes-a-seam precedent in `agent-loop` → `dsh-llm`). The runtime knows nothing about tools: it is handed a program and named async bindings, runs the program, and reports `{ value, logs, error? }`. Language and substrate are backend properties, so a future Python or container backend is another implementation package, not a redesign.
+3. **The shipped implementation is `@deepseek-ai/dsh-code-runtime-worker`**: one fresh Node worker thread per run, executing the model's TypeScript after type-strip, with bindings bridged over the message port, an empty environment, configurable heap/output/time caps, and hard termination. Its trust posture is bash-equivalent by design — no unsafe-acknowledgement flags — because the harness already ships `dsh-bash-local`, which executes arbitrary model-written shell commands with strictly *more* ambient authority.
+
+### The registry owns the mode
+
+`ToolRegistry` gains a schemastery-validated config (`static Config`), its first: `mode: 'native' | 'code' | 'both'`, default `'native'`. A deployment flips it from `cordis.yml` (`tools: { mode: code }`) — no code edit, per the no-hardcoded-tunables convention.
+
+**Wire tool list.** The registry contributes visible capabilities in `'native'`, only `run_code` in `'code'`, and both in `'both'`. The final `PromptAssembly.tools` list is logged in the request header. `run_code` is a reserved presentation transport outside registration and restriction layers; direct prompt providers and the assembly waterfall remain responsible for their own contributions.
+
+**Interaction with `toolOrder`, stated up front:** a configured `systemPrompt.toolOrder` naming native capabilities rejects every assembly under `mode: 'code'`, because those names are outside that mode's wire-validation universe. This is correct behavior, not a bug: a deployment using Code Mode updates its order config or drops it.
+
+**SDK prompt section.** In `'code'` and `'both'`, the lazy `tools:sdk` section in the tool-guidance order band renders TypeScript declarations plus fixed usage instructions for the scope's visible capabilities. It shares lookup and execution visibility, excludes `run_code`, and sorts tools lexicographically for byte-stable output.
+
+**Assembly ownership.** `run_code` and `tools:sdk` enter the trusted `system-prompt/assemble` waterfall as normal assembly inputs. A scoped `tools:sdk` section may shadow the global default before dispatch, and a listener may remove or replace either contribution. The waterfall's returned assembly is final, so whoever changes these inputs owns preserving a viable Code Mode protocol when the deployment expects Code Mode to remain usable; no restoration pass overrides deliberate composition.
+
+**Codegen.** `jsonSchemaToTs()` maps the `defineTool` JSON-Schema subset to TypeScript, carries schema descriptions into JSDoc, and degrades unsupported constructs to `unknown`. The SDK exposes tools as quoted object keys, supporting arbitrary names without aliases or collisions. Typing is advisory because the runtime strips types before execution.
+
+### The run_code tool and the dispatch bridge
+
+Under `'code'` and `'both'` the registry owns `run_code` as a reserved presentation transport with one required parameter, `{ code: string }`. It is represented by a normal `ToolDefinition` for dispatch but stays outside the filterable capability layers, so restrictions cannot accidentally remove Code Mode's only entry point. Calls traverse the complete tool pipeline — `tools/pre-execute` → monotonic guards → `tools/execute` around dispatch → `tools/post-execute` → immutable `tools/result` notification — exactly like native calls; a permission plugin can inspect the program text before it runs, and final-result observers see the normalized outer outcome. Its `execute(args, exec)`:
+
+1. **Build bindings.** One run-scoped signal follows outer cancellation and is aborted whenever the run settles. Each visible tool binding JSON-normalizes its arguments—rejecting lossy values before dispatch—waits on the serialization queue, executes with a deterministic call id and the outer token as `parent`, defers returned contexts through the outer execution, and logs `tool/code-dispatch`. Successful text becomes a string and non-text blocks become placeholders; tool errors reject the binding promise. Every sub-call retains its own immutable execution identity and traverses the full tool pipeline.
+2. **Runs the program**: `ctx.codeRuntime.run({ program: args.code, bindings: [{ global: 'tools', functions }], signal: runController.signal })`. The runtime receives the run-scoped signal, not only the caller's outer signal, so any way the outer run settles also aborts work inside the runtime.
+3. **Settle after quiescence.** When the runtime settles, the bridge aborts outstanding work and drains the dispatch queue before returning. Success returns captured output and presentation metadata. A runtime failure becomes `CodeRunFailedError`; backend rejection uses the registry's normal error boundary. Both produce structured error results, and no sub-call can append after `run_code` settles.
+
+**Sub-call contexts are deferred through the parent.** Injecting inside `run_code` would break parent call/result adjacency, so `ToolRunContext.deferContext()` collects every sub-result `additionalContexts` entry in dispatch order. The registry carries that array even when the program later throws, and the loop appends each entry only after the outer result and every sibling result in the step. An outer post-execute block discards tool-deferred entries and exposes only contexts explicitly attached by the blocking decision.
+
+**Concurrency is serialized.** Each run owns a dispatch queue, so even `Promise.all` executes tool calls in submission order. Settlement abandons queued calls that have not started. Parallelism requires per-tool concurrency-safety metadata.
+
+**Presentation.** `run_code`'s render intent is decided here per the [render-intent Agent Note](../architecture/2026-07-02-tool-render-intent-union.md): `presentCall` → a `generic` card, `kind: 'execute'`, title = the program text, `rawInput` = the same program text; `presentResult` → a `generic` card whose content is the captured output (from `meta`). The program is the title because ACP execute cards reliably render that field while some clients omit body and raw-input content. This is not a `terminal` card: that card's semantics are "a shell command in a working directory", which a program is not.
+
+### Observability: `tool/code-dispatch`
+
+Each sub-dispatch appends a log-only `tool/code-dispatch` event containing parent and child call ids, tool identity, normalized arguments, and result summary. It remains outside model history but available to persistence and UIs. Appends occur inside the open `run_code` turn. Direct executions without an agent still run but cannot log the event.
+
+### The code-runtime seam
+
+`packages/code-runtime/code-runtime/` — `@deepseek-ai/dsh-code-runtime`, depending only on `cordis`. An abstract `CodeRuntime extends Service` (`super(ctx, 'codeRuntime')`) plus the vocabulary:
+
+- `CodeRunRequest = { program: string; bindings: CodeBindingNamespace[]; signal?: AbortSignal }`
+- `CodeBindingNamespace = { global: string; functions: Record<string, (args: unknown) => Promise<unknown>> }` — the runtime exposes each namespace as a global object of async functions inside the program; binding arguments and resolutions must be structured-cloneable (a runtime may cross a serialization boundary; ours does).
+- `CodeRunResult = { value?: unknown; logs: CodeLogEntry[]; error?: CodeRunFailure }` — program execution outcomes, including exception, timeout, abort, and worker exit, resolve as the `error` field. `run()` may reject only for caller/seam misuse (for example a duplicate binding namespace); consumers still contain a non-conforming backend rejection at their own error boundary.
+- `CodeLogEntry = { source: 'console' | 'stdout' | 'stderr'; level?: 'log' | 'info' | 'warn' | 'error' | 'debug'; text: string }`
+- `CodeRunFailure = { kind: 'exception' | 'timeout' | 'abort' | 'worker-exit'; message: string }` — orthogonal outcomes reported independently per [defensive patterns](../../../../docs/defensive-patterns.md); a timed-out run is not an exception, an abort is not a timeout.
+- Two readonly backend descriptors, informational not gating: `language` (what the program must be written in — `'typescript'` for the shipped backend; a Python backend would say so, and pair with its own SDK generator on the presentation side) and `isolation` (`'worker-thread'` for the shipped backend; `'process'`, `'container'`, … for future ones). `dsh-tools` requires `language === 'typescript'` in the MVP — its codegen emits TS — and fails the assembly loudly otherwise, the same misconfiguration idiom as `toolOrder` violations (as when `mode` is non-native with no `ctx.codeRuntime` loaded at all).
+
+Requests contain every runtime input; implementations own validated timeout and cap defaults. The registry looks up the optional runtime only when Code Mode is assembled, so native mode does not depend on one. Missing or language-incompatible runtimes fail loudly. Alternate substrates or languages can replace the implementation behind the same seam, paired with the appropriate SDK generator.
+
+### The worker-thread runtime
+
+`@deepseek-ai/dsh-code-runtime-worker`, the second package of the `packages/code-runtime/` group. Per `run()`:
+
+1. **Type-strip host-side** with Node's built-in `stripTypeScriptTypes` (`node:module`; present across the repo's whole engines range, `^22.19.0 || >=24.0.0`, and position-preserving, so runtime error line numbers match the model's source). Strip-only mode rejects non-erasable syntax (`enum`, namespaces) — that rejection returns as `error.kind: 'exception'` with Node's message, the SDK instructions say "erasable TypeScript only", and the model self-corrects like any other program error. A syntax-level failure never spawns a worker.
+2. **Spawn one fresh `Worker` per run** from the package's own bootstrap module: `env: {}` (truly empty — stronger than the scrubbed-env rule for spawned commands), `resourceLimits` from config, `stdout`/`stderr` captured into `logs` rather than inherited. No pooling and no cross-run state: a program's world dies with its worker, which keeps runs reconstructable from the log alone and makes state bleed unrepresentable.
+3. **Execute** in the bootstrap: the stripped program becomes the body of an `AsyncFunction` whose parameters are the binding globals and a capturing `console` shim, so top-level `await` and `return` work and the program's completion value is the run's `value` (structured-cloneable values cross as-is; anything else is replaced by its `util.inspect` rendering, documented).
+4. **Bridge bindings over the message port**: each binding function in the worker posts `{ id, global, name, args }` and awaits the reply; the host validates the name against the request's bindings, invokes, and replies `{ id, ok, value }` or `{ id, ok: false, message }` (a host-side binding rejection becomes a program-side rejection). The worker-side namespace objects are built null-prototype via `defineProperty`, so a binding named `__proto__`, `constructor`, or `toString` is an ordinary own property, not a prototype collision. Unknown names, duplicate ids, and post-settlement messages are rejected or ignored — the port protocol assumes a hostile peer, because the peer runs model code.
+5. **Enforce independent budgets.** `computeMs` meters worker busy time, allowing slow awaited tools without excusing a hot loop. `maxWallMs` bounds total elapsed time, including unresolved waits. Expiry, cancellation, and completion terminate the worker. Heap exits and truncation are reported explicitly; compute, wall, heap, log, and return-value caps are validated configuration.
+6. **Dispose to quiescence**: the service's own disposal terminates in-flight workers and *awaits* their exits before resolving, per [defensive patterns](../../../../docs/defensive-patterns.md).
+
+### Trust posture
+
+The worker runtime provides containment, not a security boundary: model code can reach Node APIs and has authority comparable to the bash tool. `worker.terminate()` stops the thread but not OS processes it spawned. Code Mode uses the same `tools/pre-execute` policy gate as bash and adds an empty environment, heap limits, a separate isolate, and hard termination of the program itself. Deployments that need a hard multi-tenant boundary need a container-class backend for both code and bash; the runtime's isolation descriptor lets them distinguish that backend.
+
+### What the model sees
+
+The SDK instructs the model to write an async erasable-TypeScript body, call tools through `await tools.name(args)`, catch rejected tool calls when needed, and return or log only the output that should re-enter context. Calls remain sequential even under `Promise.all`. The declaration prefix can be as large as native schemas, especially in `'both'`, but remains stable for provider caching.
+
+## Consequences
+
+Deployments switching to `'code'` must update any native-only `toolOrder`. Assembly listeners own the integrity of any rewritten protocol surface. Sub-dispatch remains serialized, while per-call contexts retain their source, envelope, and metadata through the outer result.
+
+## Testing
+
+- **Worker runtime:** Real-worker tests cover output and value capture, failure kinds, compute and wall budgets, hostile binding traffic, empty environment, structured-clone fallback, output caps, and disposal to quiescence. A built-package test runs the worker entry under plain Node.
+- **Registry integration:** Tests cover code generation, all presentation modes, reserved-name and restriction rules, scoped visibility, authoritative assembly rewrites, `toolOrder`, runtime compatibility failures, full-pipeline sub-dispatch, parent-token correlation, serialization, cancellation and queue drain, JSON normalization, error propagation, log events, ordered context deferral across successful and failed programs, outer-block suppression, and HMR cleanup.
+- **With-key e2e:** A real model composes two bash calls in one program; another discovers nested workspace instructions through a Code Mode fs dispatch. The tests verify collapsed request headers, correlated dispatch events, resulting files, deferred context, and model behavior.
+- **Snapshot:** The `code-mode-turn`, `both-mode-turn`, and `code-mode-workspace-context` fixtures pin SDK text, header tool lists, dispatch events, deferred context, and result cards.
+
+## Alternatives considered
+
+**An add-on consumer plugin with zero core changes.** Rejected because `agent/request` is call-config-only under [reconstructable requests](../architecture/2026-07-05-reconstructable-requests.md), while transforming an assembled tool list would have to undo `toolOrder` canonicalization without owning its config and would depend on listener order. Which tools the model is offered, and in which representation, is the registry's single concern: native schemas and the SDK are two projections of one visible store.
+
+**`node:vm` as the reference runtime, with hardening deferred.** Rejected: `node:vm` is not isolation (prototype-chain escapes reach the host realm) and cannot interrupt a hot loop. A worker thread provides a separate isolate, empty environment, `resourceLimits`, and reliable `terminate()` at bash-equivalent trust, so the reference and production implementation are one package without an unsafe-acknowledgement ceremony.
+
+**Result elision / summarization over native tool-calling.** Addresses only the context-bloat half of the problem: trimming old `tool-result`s is cheap to add as a logged surface replacement under reconstructable requests, but still pays one model round-trip per call and cannot express loops, branches, or joins. Complementary, not competing; it can layer under Code Mode for residual native calls.
+
+**Parallel native dispatch in the loop.** The other answer to round-trip cost; still valid future work (the open TODO), still blocked on concurrency-safety metadata, and still no composition — it parallelizes calls the model already decided on in one step. Code Mode's serialized-queue decision keeps the two compatible: when the metadata lands, both native parallel dispatch and per-tool binding parallelism unlock together.
+
+**Always-exclusive (Cloudflare-faithful, no mode).** Rejected for this SDK's primary consumer: a coding agent's bread-and-butter single calls (`bash`, `read`, `edit`) are already ideal as native calls, and forcing every edit through a program taxes the common case. The mode config keeps the faithful form (`'code'`) one line away without imposing it.
+
+**Per-tool visibility tiers (this tool native, that tool code-only).** Deferred: it needs per-tool metadata and a presentation split that `'native' | 'code' | 'both'` does not, and its design depends on evidence about how models split usage under `'both'`.
+
+**Sanitized identifier aliases in the SDK** (`my-tool` → `my_tool`, Cloudflare's approach). Rejected: quoted keys on a `declare const` make every name reachable with zero alias-collision logic; models handle `tools["my-tool"](…)` fine.
+
+**A REPL-style persistent kernel** (state survives across `run_code` calls). Rejected for the MVP: cross-call state would be invisible to the session log, breaking the reconstructability guarantee that every request is a pure function of the log; fresh-per-run keeps it. A kernel-style backend remains expressible behind the seam later, with its own logging story.
+
+## Risks
+
+**The worker is not a hard security boundary.** Deliberate and documented (§Trust posture): posture equals the existing bash tool, containment exceeds it, gating uses the same seams. Deployments needing more need a future `isolation: 'container'` backend — tracked as the seam's designed extension, not a TODO on this design.
+
+**`stripTypeScriptTypes` is marked experimental.** It is the same engine (amaro/swc) behind Node's own native `.ts` execution, exposed as an API across this repo's whole engines range. Mitigations: the runtime's unit suite pins the behaviors relied on (position preservation, erasable-only rejection message shape loosely), the call sits behind one private function, and `amaro`/`sucrase` are drop-in replacements if the API shifts. The erasable-only subset is a model-facing contract line, and the error path is a working feedback loop, not a dead end.
+
+**Prompt cost of the SDK, especially under `'both'`.** The `.d.ts` can rival the native schemas it complements; `'both'` carries two representations. Prefix stability + provider caching amortize per-session cost; the mode is per-deployment; the Agent Note makes no unconditional-savings claim. Measured guidance (when to prefer which mode) is explicitly post-ship learning.
+
+**Registry scope growth.** `dsh-tools` absorbs codegen, a tool, a bridge, and an event. Contained by module boundaries inside the package (`ts-types.ts`, `code-mode.ts` beside `schema.ts`/`json-schema.ts`/`presentation.ts`) and by the seam: everything substrate-shaped lives behind `ctx.codeRuntime`.
+
+**Structured-clone values can exceed JSON.** Tool bindings therefore JSON-normalize arguments before dispatch, ensuring every executed call can be logged. The lower-level runtime keeps its wider port contract, while stricter consumers validate at their boundary. Non-text sub-results become placeholders.
+
+**Serialized-only sub-dispatch.** `Promise.all` gains no wall-clock parallelism yet, only fewer round-trips; models may over-expect. The instructions state it; lifting it is tied to the same concurrency-safety metadata the native parallel-dispatch TODO needs.
+
+**Budget metering reads the event loop, not a flag.** Busy-time polling (`eventLoopUtilization()`) is coarser than an exact CPU meter — a budget expires up to one poll interval late — and its correctness claim ("a pending dispatch cannot pause it") is load-bearing against a hostile program. Both sides are unit-tested (hot loop with a pending decoy dispatch dies at `computeMs`; idle-on-slow-binding survives to `maxWallMs`), and the poll interval is an internal constant, not config — nothing a deployment could mis-tune into a bypass.

.agents/notes/implemented/feature/2026-06-17-filesystem-tool-schemas.md

@@ -0,0 +1,110 @@
+# Agent Note: Filesystem tool schemas — model-facing read/write/edit shapes
+
+Status: implemented
+
+## Problem
+
+[The filesystem capability-seam Agent Note](../architecture/2026-06-17-filesystem-capability-seam.md) defines the filesystem capability seam (`ctx.fs`), the package split (`dsh-fs`, `dsh-fs-local`, `dsh-tool-fs`, plus the `dsh-fs-policy` policy plugin), and the observed-file/stale-version policy for read-before-write/edit checks — which the [split-fs-seam](../simplification/2026-06-26-fsspec-style-fs-seam.md) and [event-gate](../architecture/2026-06-26-file-context-as-event-gate.md) Agent Notes moved off `ctx.fs` into the `dsh-fs-policy` plugin on the `fs/*` event gate. The remaining decision for the first filesystem tool delivery is the model-facing schema surface: what arguments the model sees for `read`, `write`, and `edit`.
+
+The schema should be small enough to implement in the first `dsh-tool-fs` pass, but stable enough that future local/remote/sandboxed filesystem backends do not require model-facing churn. It should also avoid importing every option from reference systems. Claude Code and OpenCode expose similar core file tools but differ in naming style and extra flags; this Agent Note chooses the minimal shared surface for the prototype.
+
+## Decision
+
+`@deepseek-ai/dsh-tool-fs` exposes these three model-facing tools in the first filesystem suite:
+
+| Tool | Our schema | Claude Code | OpenCode | Notes | Part of prototype |
+|---|---|---|---|---|---|
+| `read` | `read(file_path, offset?, limit?)` | `Read(file_path, offset?, limit?, pages?)` | `read(filePath, offset?, limit?)` | Files only; 1-indexed `offset`; no image/PDF/multimodal support in the first pass. | YES |
+| `write` | `write(file_path, content)` | `Write(file_path, content)` | `write(content, filePath)` | Creates or overwrites UTF-8 text. Under the default fs-policy, updates to existing files require a prior observation; new-file creates do not. | YES |
+| `edit` | `edit(file_path, old_string, new_string, replace_all?)` | `Edit(file_path, old_string, new_string, replace_all?)` | `edit(filePath, oldString, newString, replaceAll?)` | Literal string replacement; unique match required by default; under the default fs-policy requires a prior observation (any windowed read counts). | YES |
+
+The schema uses snake_case field names (`file_path`, `old_string`, `new_string`, `replace_all`) to align with Claude Code and with existing DeepSeek Harness tool-schema examples. The consumer package translates these model-facing names into `ctx.fs` calls and `fs/*` event dispatches.
+
+## Tool schemas
+
+### `read`
+
+`read` inspects a UTF-8 text file and returns line-numbered content.
+
+Arguments:
+
+- `file_path: string` — required. Path to read, resolved by `ctx.fs`.
+- `offset?: number` — optional. 1-based first line to return. Defaults to the first line.
+- `limit?: number` — optional. Maximum number of lines to return. Defaults and caps are implementation details of `dsh-tool-fs` / `ctx.fs`.
+
+Non-goals for the first pass:
+
+- No PDF `pages` argument.
+- No image or multimodal file reads.
+- No directory listing through `read`; if needed, listing becomes a separate future tool.
+
+### `write`
+
+`write` creates or fully replaces a UTF-8 text file.
+
+Arguments:
+
+- `file_path: string` — required. Path to write, resolved by `ctx.fs`.
+- `content: string` — required. Full UTF-8 text content to write.
+
+Under the default fs-policy, updating an existing file with `write` requires a prior observation (a read/write/edit) of that file by the same execution context; the `dsh-fs-policy` plugin supplies the observed version as the stale guard on `fs/write-intent`. Creating a new file does not require a prior observation. With the policy plugin absent, `write` is an unconditional bare-provider create-or-overwrite.
+
+The schema does not expose `expected_hash`, `expected_version`, or `create_only` as model-facing parameters. Stale-version checks are driven by backend-produced versions and the policy plugin's observed state, not by asking the model to copy version tokens through the schema.
+
+### `edit`
+
+`edit` updates an existing UTF-8 text file by replacing literal text.
+
+Arguments:
+
+- `file_path: string` — required. Path to edit, resolved by `ctx.fs`.
+- `old_string: string` — required. Literal text to replace. Empty strings are invalid in the first pass.
+- `new_string: string` — required. Literal replacement text; an empty string deletes the match.
+- `replace_all?: boolean` — optional. Defaults to false. When false, `old_string` must identify exactly one match.
+
+`edit` requires a prior observation of the file in the same execution context (any windowed read counts — authorization is version freshness, not a full-view requirement), or a prior write/edit by that context. The `dsh-fs-policy` policy plugin derives the owner and supplies the recorded version as the stale guard; the provider's mutation lock enforces it.
+
+The first pass rejects Codex-style patch grammars and multi-mode edit APIs. It uses one strict literal replacement mode so the model-facing contract stays simple and the backend can own exact-match, duplicate-match, line-ending, and stale-version semantics.
+
+## Result shape
+
+The first implementation returns `ContentBlock[]` through the existing `ToolDefinition.execute()` contract. `ctx.fs` returns structured filesystem results and owns file-state recording/refreshing; `tool-fs` formats those results into the model projection.
+
+Default native projections:
+
+| Tool | Structured `ctx.fs` outcome consumed by `tool-fs` | Default model projection |
+|---|---|---|
+| `read` | returned lines, returned line count, total line count, target display path, file version, partial-view flag | line-numbered text plus pagination footer |
+| `write` | create/update operation, target display path, new file version | concise create/update success text |
+| `edit` | replacement count, replace-all flag, target display path, new file version | concise edit success text |
+
+The structured outcome does not restate model arguments such as `file_path`, `old_string`, or `content` unless the backend has resolved them into new information such as `displayPath`, `targetKey`, or a new version. Token-conscious truncation is part of the model projection, not the backend's canonical result.
+
+## Deferred
+
+The following are deliberately out of scope for the first filesystem schema pass:
+
+- Model-facing `expected_hash`, `expected_version`, or `create_only` parameters.
+- Directory listing, glob, grep, and search tools.
+- Binary-safe read/write operations.
+- PDF/image/multimodal `read`.
+- Code Mode projection values for filesystem tools.
+- A canonical edit diff format.
+
+## Testing
+
+Schema tests pin the required/optional argument set per tool, empty-`old_string` rejection, the `replace_all` default, the snake_case field names, description prose that states the observation policy, and root-plugin suite registration; integration tests execute all three tools through `ctx.tools.execute()` against the real `dsh-fs-local` provider and verify the model arguments translate into the expected `ctx.fs` calls and `fs/*` dispatches.
+
+## Alternatives considered
+
+- **A Codex-style patch grammar or multi-mode edit API** — rejected: one strict literal replacement mode keeps the model-facing contract simple and lets the backend own exact-match, duplicate-match, line-ending, and stale-version semantics.
+- **camelCase argument names (OpenCode's style)** — snake_case aligns with Claude Code and the existing harness tool-schema examples, and naming is public surface once shipped.
+- **Model-facing `expected_hash` / `expected_version` / `create_only` parameters** — rejected: stale checks are driven by backend-minted versions and the policy plugin's observed state, never by fragile model-copied tokens.
+
+## Consequences
+
+**The first schema is intentionally smaller than Claude Code's.** Dropping PDF pages, multimodal read, rich grep/list flags, and expected hash fields keeps the implementation focused, but users may ask for those quickly. They arrive as separate Agent Notes or focused follow-ups rather than overloads of the initial schema.
+
+**No explicit model-facing stale guard in v1.** The schema does not ask the model to provide an expected hash/version. That is intentional: stale checks come from backend-produced versions and the `dsh-fs-policy` plugin's observed state, not from fragile model-copied tokens. Filesystem safety failures surface through structured `FsError` codes owned by `dsh-fs`, not through model-supplied version fields.
+
+**Naming becomes public surface.** Once shipped, changing `file_path` to `filePath` or `old_string` to `oldString` would churn prompts, examples, and downstream clients. This Agent Note chooses snake_case up front and treats it as the stable model-facing contract.

.agents/notes/implemented/feature/2026-06-18-acp-terminal-and-tool-rendering.md

@@ -1,12 +1,12 @@
-# RFC: Rich ACP bash rendering — the terminal card via the `_meta` convention
+# Agent Note: Rich ACP bash rendering — the terminal card via the `_meta` convention
 
 Status: implemented
 
 ## Problem
 
-The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](../../proposed/feature/2026-06-14-acp-agent-client-protocol.md) and `packages/core/tools`). For `bash` we surface the exact command as the `tool_call` title, the model's `description` as a content text block, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
+The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](2026-06-14-acp-agent-client-protocol.md) and `packages/core/tools`). For `bash` we surface the exact command as the `tool_call` title, the model's `description` as a content text block, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
 
-That is a correct, capability-free baseline, but not how the reference editors render a *terminal* tool at its best. An editor like Zed has a dedicated terminal tool-call card — a header showing the working directory, the command as the label, the command output rendered as a terminal, and an exit-status pill — but it only builds that card when the `tool_call` carries terminal metadata (below). With a plain text block the output appears as static markdown and there is no cwd header. (Zed also HIDES `rawInput` for `kind: 'execute'`, which is why the command IS the title — both reference adapters do the same. The human-readable description rides as a separate content block above the card; note this is a DELIBERATE divergence — claude-agent-acp DROPS the description in terminal mode and renders only the card — we keep the summary visible alongside.)
+Reference editors render terminal metadata as a dedicated card with cwd, command, live-style output, and exit status; plain text loses that structure. The command is the title because execute cards hide raw input, while the human-readable description remains a separate block above the card.
 
 ## Key finding: agent-executed terminals use a `_meta` convention, NOT `terminal/create`
 
@@ -28,7 +28,12 @@ Keep `dsh-bash` agent-side execution; render the terminal card via the `_meta` c
 3. **Bridge mapping.** When the client advertised the capability, the bridge maps that presentation to: on `tool_call`, `content:[…, {type:'terminal', terminalId}]` (any tool `content`, e.g. the description, rendered BEFORE the terminal block) + `_meta.terminal_info.{terminal_id,cwd}`; on `tool_call_update`, `_meta.terminal_output.{terminal_id,data}` (the captured output) + `_meta.terminal_exit.{terminal_id, exit_code|signal}` (the parsed exit), with the update's text `content` OMITTED (an ACP `tool_call_update.content` REPLACES the call's content collection, so re-sending the fenced block would clobber the terminal content block). `terminalId` is derived from the harness `callId` (stable, unique per call). When the capability is absent, the bridge sends the description content block on the call and the existing ` ```console ` text content on the update — unchanged.
 4. **The exit pill is parsed from the rendered output; no new execution path, no live streaming.** Output is attached at completion (from the agent's own `tool/result`), not streamed token-by-token. The exit-status pill (`_meta.terminal_exit.{exit_code,signal}`) IS emitted: the pure `presentResult(args, result)` seam sees only content blocks, so `dsh-tool-bash` recovers the structured exit by parsing the status markers (`[exit code: N]` / `[killed by signal: …]`) that `renderResult` appended — the parse is the exact inverse of the marker emission, the two co-evolve in one file, and a round-trip test guards the pair. Disposal is unaffected: nothing new to tear down, since the bridge never creates a client-side terminal.
 
-## Risks / trade-offs
+## Alternatives considered
+
+- **The ACP client-side terminal sub-protocol (`terminal/create`)** — explicitly rejected: the editor would execute the process, bypassing `dsh-bash`'s env scrub, background-task ownership, and per-session cwd, and forking execution into two backends. Both reference agents reject it the same way (the key finding above); agent-side execution plus the `_meta` convention is the only shape that yields the terminal card while keeping the harness's execution policy.
+- **Threading a structured exit through the event schema** — rejected in favor of the marker round-trip: the pure `presentResult(args, result)` seam sees only content blocks, and the parse is the exact inverse of the marker emission, co-evolving in one file under a round-trip test.
+
+## Consequences
 
 - **Zed-convention `_meta` keys.** The terminal card rides on Zed-specific keys (`terminal_info`/`terminal_output`/`terminal_exit`) inside ACP's spec-blessed `_meta` extensibility point, NOT on the ACP terminal sub-protocol. A client that doesn't recognize the keys still gets the text fallback (the capability gate ensures we only emit them when the client opted in via `_meta.terminal_output`), so a non-Zed client is never worse off. If ACP later standardizes agent-executed terminals, migrate to that and drop the convention keys.
 - **Capability honesty.** Emit terminal metadata ONLY when the client advertised `_meta.terminal_output`; the text fallback is the contract for everyone else and must never regress. Covered by a no-capability test asserting the ` ```console ` path.
@@ -38,4 +43,4 @@ Keep `dsh-bash` agent-side execution; render the terminal card via the `_meta` c
 
 ## Out of scope / non-goals
 
-The text-block baseline stays the no-capability default. Client-side `terminal/create` execution is explicitly rejected (it bypasses `dsh-bash`). Two follow-ups are deliberately NOT built here and would each warrant their own RFC when someone takes them on: **live incremental streaming** (`_meta.terminal_output_delta` as chunks arrive, which needs an incremental-output seam on `dsh-bash`), and **command classification** (parsing a `cat`/`sed` as a `read` card with a file location, a `grep` as a `search`, etc., falling back to the terminal card — display-only, must never change what executes).
+The text-block baseline stays the no-capability default. Two follow-ups are deliberately NOT built here and would each warrant their own Agent Note when someone takes them on: **live incremental streaming** (`_meta.terminal_output_delta` as chunks arrive, which needs an incremental-output seam on `dsh-bash`), and **command classification** (parsing a `cat`/`sed` as a `read` card with a file location, a `grep` as a `search`, etc., falling back to the terminal card — display-only, must never change what executes).

.agents/notes/implemented/feature/2026-06-18-compaction-capability-seam.md

@@ -0,0 +1,125 @@
+# Agent Note: Compaction as a capability seam (abstract contract + basic backend)
+
+Status: implemented
+
+## Problem
+
+A long-running agent conversation grows without bound. As the event log accumulates turns, the derived message history eventually approaches the model's context window — the model then truncates mid-response (`max-tokens`) or degrades. **Compaction** is the mitigation: replace a run of older history with a concise summary, keeping recent context intact.
+
+The [session surface](../architecture/2026-06-18-session-surface.md) was built as the foundation for exactly this — an ordered projection over the event log with a `surfaceOp: { op: 'replace', start, end }` operation purpose-built to shadow a range of entries and insert a replacement, with `sourceEventSeqs` recording provenance so the decision replays deterministically. What remained was the plugin that *decides what to compact and produces the summary*.
+
+Two forces shape the design. First, compaction policy and reusable token measurement vary independently: measurement belongs to the LLM-family [`ctx.tokenMeter` service](../architecture/2026-07-15-replay-token-meter-service.md), while summarization can be a model call, a template, or a remote service. Second, `SurfaceEventType` is closed to five event types (`user/message`, `assistant/message`, `tool/result`, `context/message`, `steering/message`); only those may carry `surfaceOp`. A bespoke `compaction/*` event therefore **cannot** itself appear on the surface — the compiler rejects `surfaceOp` on it and the invariants plugin rejects it at runtime.
+
+## Decision
+
+### Compaction is a capability seam, split interface / implementation
+
+Per the [capability-seams Agent Note](../architecture/2026-06-13-capability-seams.md), compaction ships as separate packages so the contract, the algorithm, and (later) the consumer surface evolve independently:
+
+1. **Interface** — `@deepseek-ai/dsh-compact`: an abstract `CompactService` owning the `ctx.compact` key, the `CompactionResult` vocabulary, and the `compact/*` session events. It declares `compactIfNeeded()` and `compactRegion()` as **abstract** — the contract states *what* compaction does, not *how*.
+2. **Implementation** — `@deepseek-ai/dsh-compact-basic`: a concrete `BasicCompactService` that consumes `ctx.tokenMeter` and owns the tail→head retention walk, summarization via `ctx.llm.stream()`, the surface replacement, the lock, post-step pressure, and canonical context-overflow recovery. `summarize()` is its sole subclass hook; pricing and replay stay with the meter.
+3. **Consumer** — deferred. A `/compact` tool and slash command will `inject: ['compact']` and call the contract; they are intentionally out of scope here so the seam settles first.
+
+### The contract depends on `dsh-session` and `dsh-llm` — a deliberate deviation
+
+The capability-seams Agent Note states the interface package "depends only on cordis" (true of `dsh-bash`, whose vocabulary is self-contained). Compaction **cannot** honor that: its verbs act on an agent-owned `Session` (`compactRegion(start, end, agent)`) and its output uses the content vocabulary (`CompactionResult.summary: ContentBlock[]`). There is no way to express the contract without naming `Session`/`SessionEvent` (from `dsh-session`) and `ContentBlock` (from `dsh-llm`).
+
+This is not a coupling smell — it is the contract's domain. The "only cordis" guidance was always shorthand for "the interface depends only on what the contract genuinely names, and never on an implementation." `dsh-session` and `dsh-llm` are themselves interface/vocabulary packages, not implementations; `dsh-compact` still imports no backend. The seam's real invariant — *consumers and implementations evolve independently behind an abstract service* — holds intact.
+
+### Abstract `compactIfNeeded` / `compactRegion`, algorithm in the backend
+
+An earlier draft put the full algorithm (the retention walk, token-summing, text extraction) as concrete methods on the interface. That recouples the contract to one strategy: a backend that wants a different retention policy or event sequence would have to fight inherited concrete code. Making both core methods abstract puts every *how* decision in the backend and keeps the interface a statement of *what*. Token measurement is not a compaction hook at all; the singleton service lets multiple consumers share one per-session replay fold.
+
+`compactIfNeeded(agent, trigger, signal)` takes an explicit `'pressure' | 'context-overflow'` trigger and cancellation. It reads only the latest durable routed request; no header means no work, while any routed provider/model target uses the singleton estimator. `compactRegion(start, end, agent, signal?)` uses `agent.session` as its single session identity and keeps an optional signal for manual callers. The default summarizer resolves its target from explicit config, the latest logged routed target, then agent options, and records the provider/model pair after any `llm/stream` routing.
+
+### Automatic pressure runs after successful durable step work
+
+Successful-call pressure cannot run at pre-step because final `agent/request` routing, provider output, tool results, buffered context, and steering do not exist there. Serial `agent/post-step(agent, turn, step, signal)` fires after those facts are durable and before `step/end`. `dsh-compact-basic` measures the canonical logged request through `ctx.tokenMeter`, so the next request sees any replacement without a speculative envelope override.
+
+Canonical provider context overflow takes a separate path. The failed step closes, `agent/request-error` receives the original request error and consecutive retry count, and compact-basic forces one useful balanced reduction. It returns retry only if `session.surface.replaceGeneration` increases; the loop then opens a new numbered step and reconstructs its request from the durable log. No range, no replacement, recovery failure, cancellation, an exhausted cap, or an unrelated error preserves the original provider failure. The complete lifecycle decision is in the [after-call recovery Agent Note](../architecture/2026-07-10-after-call-compaction-pressure-and-overflow-recovery.md).
+
+```
+assistant/message → tool/result/context/steering
+await serial agent/post-step          ⟵ pressure compaction inside the successful step
+step/end
+
+provider overflow → step/end
+await waterfall agent/request-error  ⟵ forced compaction between attempts
+retry → next numbered step/start      ⟵ derives from the replacement surface
+```
+
+### Retention is turn-agnostic; tool-pairing balance is the only structural guard
+
+Auto-compaction checks after **every successful** step, not once per turn. This is load-bearing for runaway-turn survival: a tool-heavy ReAct turn appends an `assistant/message` + a `tool/result` per step, so the surface grows within a turn. The post-step check can compact early closed tool pairs before continuation opens the next step, and provider-confirmed overflow remains the backstop when a request crosses the limit first.
+
+`compactIfNeeded` retains the smallest tail of whole surface units whose estimated size reaches `retainTokens` and compacts older nodes. A unit is a complete closed step or one no-step message. If the token cutoff lands inside a step, retention expands until the cut is tool-pairing balanced. Balance is checked on surface order, not log sequence, because replacement summaries have new sequence numbers at old surface positions. `dsh-compact` exports the before/after edge helpers; their per-session cache folds only appended surface-tail nodes while `replaceGeneration` is unchanged, does no event reads for log-only growth, and rebuilds current membership and balances after replacement. `compactRegion` rejects boundaries that split a tool call from its result. The in-flight turn receives no special retention.
+
+A runaway turn thus compacts exactly like any other history: its early *closed* steps get summarized while its recent steps stay verbatim. When the only compactable content left is an un-splittable open tail step (its tool-calls have no results yet), compaction declines (`null`) and retries once that step closes.
+
+**Single-unit overflow is out of scope, by design.** If a single retained unit — one closed step, or a large free entry such as a pasted `user/message` — *alone* exceeds the budget, compaction cannot help and the next model call may go out over-budget. Bounding an individual unit's size is a separate concern (output truncation), handled elsewhere; compaction makes no promise about it, and the harness without such a mechanism can still break on a single oversized unit. This is named honestly rather than papered over.
+
+### Head-anchoring: one auto checkpoint, always at the head
+
+Auto-compaction always starts at the surface head, merging the prior checkpoint with newly compacted history so only one automatic checkpoint remains. `shadowedRange` is therefore positional rather than a numeric sequence interval: a newer summary sequence may occupy an older surface position. `shadowedSeqs` records the authoritative surface order. Manual mid-range compaction may leave multiple checkpoints.
+
+### Approximate convergence invariant
+
+`resolveConfig` supplies usable defaults: threshold ratio `0.8`, retained tail `floor(contextWindow × 0.16)`, empty summarization provider/model overrides, `maxTokens: 8192`, `compactionRetries: 1`, `maxOverflowRetries: 1`, and `auto: true`. Optional top-level `thresholdRatio` and `retainTokens` override the policy for the token meter's single context window; retention must remain below the resulting threshold. Convergence remains dynamic because provider output caps can be spent on hidden or surfaced reasoning tokens and summary size is unpredictable. If pressure remains over threshold, `compactIfNeeded()` re-compacts the head checkpoint up to the configured retry count, but each committed summary must be smaller than what it shadows. Overflow bypasses threshold and retained-tail policy for one maximal balanced head reduction, leaving the newest indivisible unit.
+
+### Surface replacement: `compact/*` events are log-only; one `user/message` carries the summary
+
+Because `SurfaceEventType` is closed, the summary cannot ride on a `compact/*` event. The backend instead appends a **single `user/message`** with `surfaceOp: { op: 'replace', start, end }` whose `content` is the (framed) summary and whose `sourceEventSeqs` covers the shadowed entries *and* the bookkeeping events. The `compact/*` events are pure log records (lock + provenance). The surface mutation sits **inside** the lock — `compact/end` is the last event appended:
+
+```
+compact/start    → log-only. Acquires the lock.
+[summarize older range via the backend]
+compact/summary  → log-only. Provenance: raw summary, range, shadowed seqs, token count.
+user/message     → surfaceOp { op:'replace', start, end }. THE surface mutation (framed summary).
+                   deriveMessages() renders it as a user-role message.
+compact/end      → log-only. Releases the lock (carries `error` on a recoverable failure).
+```
+
+`deriveMessages()` then yields `[summary_as_user_message, ...retained_entries]`. Reusing `user/message` is honest rather than a workaround: a summary genuinely *is* user-role context.
+
+### Checkpoint framing + incremental merge (backend-private)
+
+The basic backend wraps the summary as established checkpoint context and tags it for incremental merging on the next cycle. The raw summary remains on `compact/summary`. Framing is backend policy; the seam promises only that one replacement user message carries the possibly framed summary.
+
+### Blocking via a log-recorded lock, plus a crash/recoverable failure taxonomy
+
+The `compact/start … compact/end` bracket is justified, in order of what now does the work:
+
+1. **Crash-detectable orphan + provenance** (primary). Summarization is a slow model call persisted *after* `compact/start`. A crash mid-summarization leaves a `compact/start` with no matching `compact/end` — a detectable orphan. Releasing the lock last (rather than first) converts the crash window from *silent corruption* into that detectable orphan.
+2. **Prevents concurrent compaction.** `compactRegion` refuses to start if the current turn holds an unmatched `compact/start`. (The loop is single-threaded across either awaited automatic seam, so this is also a re-entry tripwire — a thrown "already in progress" signals a real bug.)
+
+Two failure paths, both documented:
+
+- **Crash** (the loop dies mid-summarization): a dangling `compact/start`, no closer. Because `compact/*` are **log-only**, the orphan is **inert** — the surface replacement never landed, so the full, uncompacted history derives correctly. Generic turn-repair (`interruptedTurnClosers`) closes the turn with a synthetic `turn/end`; the orphan sits *before* that `turn/end`, so the turn-scoped in-progress check never sees it and a crash cannot wedge future compaction.
+- **Recoverable** (summarization throws but the loop survives): the backend appends `compact/end` with its **`error`** field set and leaves the surface untouched. Post-step pressure warns and continues; overflow recovery delegates so the original provider error remains authoritative.
+
+`compact/end` keeps its `error?` field (mirroring `tool/result`'s self-contained error — one event tells success from failure without correlating a sibling). There is no separate `compact/error` event.
+
+**Core session repair stays compaction-agnostic — deliberately.** `interruptedTurnClosers` is never taught about `compact/*`. Teaching it would force every future `xxx/start … xxx/end` plugin pair to patch a core module — exactly the coupling the capability-seam architecture exists to avoid. Because the log-only orphan is inert, no special repair is needed: generic turn-repair plus the inertness of an un-landed surface mutation is sufficient.
+
+## Alternatives considered
+
+- **The full algorithm as concrete interface methods** — rejected because it recouples the contract to one retention strategy. Both core methods are abstract; reusable measurement is a separate LLM-family service and `summarize()` is basic's sole hook.
+- **Compaction on `agent/request` or provisional `agent/pre-step` inputs** — rejected because neither proves the final durable request and both couple generic lifecycle to compaction-specific envelope data. Post-step replay plus canonical overflow recovery covers both successful and rejected calls.
+- **A separate `compact/error` event** — rejected: `compact/end` keeps an `error?` field, mirroring `tool/result`'s self-contained error — one event tells success from failure without correlating a sibling.
+- **Teaching core turn-repair about `compact/*`** — rejected: the log-only orphan is inert, and a core module patched for every future `xxx/start … xxx/end` plugin pair is exactly the coupling the capability-seam architecture exists to avoid.
+
+## Consequences
+
+- **Packages**: `packages/compact/compact` supplies the interface and `compact-basic` supplies the backend. `packages/llm/token-meter` owns replay-aware measurement independently. The consumer tier is deferred.
+- **Automatic seams**: `agent/post-step` (`@mode serial`) handles successful-call pressure and `agent/request-error` (`@mode waterfall`) handles final request failures after the failed step closes. Generic `agent/pre-step` remains a four-argument checkpoint with no compaction-only prompt/prefix payload.
+- **`SessionEventMap`** gains `compact/start` / `compact/summary` / `compact/end` by declaration merging (merge-extensible); `SurfaceEventType` is **not** touched. These are session events, not cordis `Events`, so the event-taxonomy gate needs no entry.
+- **`dsh-compact`** owns `toolPairingBalancedBefore(session, seq)` and `toolPairingBalancedAfter(session, seq)`, the cached surface-edge checks that `compactRegion` and `compactIfNeeded` use to avoid splitting a tool-call/result pair. The cache validates current membership by seq and answers both edges from one per-cut balance sequence; stale or missing seqs and orphan results reject. `dsh-session` continues to own the surface `replace` operation, ordered event sequences, and rewrite generation.
+- **`dsh-invariants`** drops its `surface replace: start must be <= end` assertion: a head-anchored compaction lands a high-seq replacement entry at an older range's *position*, so `start > end` numerically is normal and valid (the range is positional, validated by the surface's `indexOf` checks that remain). The turn-enclosure invariant is reused unchanged.
+- **Wiring**: `examples/repl-agent/cordis.yml` loads zero-config `dsh-token-meter` before `dsh-compact-basic`; the service-wide window and compact defaults make the pair usable without repeated numeric policy.
+
+## Testing
+
+- **Unit:** Real Loader and invariant plugins cover whole-unit retention, convergence failure, both `compact/end` outcomes, head anchoring, open-tail refusal, inert crash orphans, forced below-threshold overflow, generation proof, caps, and original-error preservation.
+- **Loop:** Tests pin post-step after durable tool results and before `step/end`, actual `agent/request` routing, closed failed steps, fresh retry numbering, and complete thrown/in-band overflow → compaction → reconstructed retry composition.
+- **With-key e2e:** A real model and bash session with lowered limits triggers compaction, records a complete `compact/start…end` pair, shrinks the surface, and finishes the task.
+- **Snapshot gap:** Runaway-turn compaction cannot yet replay because the summarization call records no `assistant/chunk` events or `sessionId`; interleaved summarization-call replay remains follow-up work.

.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.md

@@ -0,0 +1,71 @@
+# Agent Note: Subagent capability seam
+
+Status: implemented
+
+> The full seam is shipped: the `dsh-subagent` interface and `dsh-tool-subagent` consumer; the two in-process backends (`dsh-subagent-spawn`, `dsh-subagent-fork`); the nested-agent snapshot infrastructure ([per-session snapshot replay](../testing/2026-06-22-subagent-snapshot-replay.md)); and the out-of-process `dsh-subagent-acp` backend ([its Agent Note](2026-06-22-acp-subagent-backend.md)).
+
+## Problem
+
+The harness has a long-deferred seam for **subagents** — an agent delegating work to another agent. The intent was sketched in the `Agent`/`AgentLoop` interfaces ([packages/core/agent/src/types.ts](../../../../packages/core/agent/src/types.ts), [packages/core/agent-loop/src/index.ts](../../../../packages/core/agent-loop/src/index.ts)): a creation option referencing a parent agent (fork = seed the child session with the parent's event log; spawn = fresh session), with the child returned as an `Agent` handle so steering and event subscription work uniformly. This Agent Note realizes that seam; the banner above lists what shipped.
+
+The distinctive requirement — the one that shapes the whole design — is that **multiple subagent implementations must coexist at runtime**. A parent may want a cheap in-process child for a scoped subtask AND an isolated out-of-process child (over ACP) in the same session. The transports we foresee:
+
+- **in-process** — a child concrete `Agent` on the same `Context` (the cheapest, and nearly free given the existing agent factory);
+- **ACP** — act as an ACP *client* driving another agent process (which can be another instance of ourselves);
+- later: **A2A**, the **Codex app-server**, and the **Claude Code Agent SDK** — each the same out-of-process "start a child, prompt it, stream updates, cancel" shape as the ACP backend.
+
+## Alternatives considered
+
+### Why not the bash seam shape
+
+The bash seam ([capability seams](../architecture/2026-06-13-capability-seams.md)) registers exactly one `BashExecutor` per context; loading a second throws. That is correct for bash (one machine, one way to run a command) but wrong here: coexistence is the requirement. So the subagent service is a **named-provider registry** — each implementation registers under a unique name and a caller picks one by name — mirroring the **LLM adapter registry** (`LlmService.registerAdapter`), not the single-service bash executor. The seam is still three-package (interface / implementation / consumer); only the "one vs. many implementations" axis differs.
+
+## Decision
+
+### The three-package seam
+
+A new package group `packages/subagent/`:
+
+| Package | Role |
+|---|---|
+| `@deepseek-ai/dsh-subagent` | interface: `SubagentService` (`ctx.subagents`), `SubagentProvider`, `SubagentRun`, the request/result/capability vocabulary, the `subagent/*` events |
+| `@deepseek-ai/dsh-subagent-spawn` | implementation: a fresh in-process child via `ctx.agents.create` |
+| `@deepseek-ai/dsh-subagent-fork` | implementation: an in-process child seeded with a snapshot of the parent's log |
+| `@deepseek-ai/dsh-subagent-acp` | implementation: an ACP client driving a configured child process |
+| `@deepseek-ai/dsh-tool-subagent` | consumer: the model-facing `subagent` tool over `ctx.subagents` |
+
+### The primitive: async `start → SubagentRun`
+
+A provider exposes `start(request) → Promise<SubagentRun>`. Fulfillment publishes a ready child and transfers its run handle to the caller. One signal covers cancellation before and after readiness; `dispose()` cancels remaining work and awaits quiescence. A rejected start cleans partial resources and emits no lifecycle event. `start` is transport-neutral; `spawn` names only the fresh in-process backend.
+
+### Two kinds of optional capability, discovered two ways
+
+- **Start-time features** (`outputSchema`, `depthLimit`, `toolFilter`, `persona`) ride on a static `provider.capabilities` descriptor. The service checks every requested one BEFORE delegating and **rejects loud** (`SubagentError('UNSUPPORTED_CAPABILITY')`) if the provider lacks it — never accepted-then-ignored. They must be checked before a run exists, which is why they cannot be runtime methods.
+- **Runtime features** (steering via `sendMessage`, follow-up via `resume`) are **optional methods** on `SubagentRun`. The method's presence IS the capability, and TypeScript narrowing is the discovery mechanism: a consumer cannot call an absent method without narrowing first, so there is no silent-degradation path and no separate flags object to keep in sync.
+
+### Fork vs. fresh are separate backends, not a flag
+
+Fresh and forked children are separate providers, not a request flag. `dsh-subagent-spawn` starts an isolated child; `dsh-subagent-fork` seeds a balanced prefix containing only completed parent turns. The in-flight turn is excluded because its subagent call has no result yet and cannot form valid replay history.
+
+### Child isolation and the parent log
+
+Each subagent runs in its **own `Session`** (own id, `parentSession` lineage), persisted independently. The parent's log records only the spawn `tool/call` and its `tool/result` (the child's final output) — the child's internal steps and tool calls stay in the child's own session, never injected into the parent log. This is the only design that is identical across transports: an ACP child's internal events physically cannot be injected into our parent log, so making in-process behave the same keeps the seam transport-agnostic.
+
+### Synchronous collect (first cut)
+
+`dsh-tool-subagent` passes its execution signal to `start()`, awaits the child result, and disposes the run in `finally`. Non-completed outcomes become error results rather than successful partial output. This foreground consumer does not use the run's optional steering method.
+
+### Provider selection is config, not model-facing
+
+`dsh-tool-subagent` binds to exactly one provider name (`Config.provider`); the model sees only `{ description, prompt }`. To expose more than one transport, load the tool plugin more than once, each bound to a different provider and a distinct `toolName` (the tool registry rejects a duplicate name). The *service* holds the multi-provider registry; the *tool* picks one — no provider/type parameter in the schema this cut.
+
+## Testing
+
+Registry and tool tests replace only the nondeterministic child boundary with a package-local scripted provider while exercising the real `SubagentService`, lifecycle, task integration, and model-facing tool. Provider and consumer export shapes retain their Loader regression coverage for the failure described in [postmortem 0001](../../../../docs/postmortem/0001-acp-default-export-drops-inject.md). Registry tests cover reload safety, duplicate names, and start-time capability rejection; nested-agent scenarios replay keylessly through [per-session snapshot replay](../testing/2026-06-22-subagent-snapshot-replay.md); in-process backends also have real-loop unit tests and a with-key e2e.
+
+## Consequences
+
+- **Recursion.** Without a bound, an in-process child can see the delegation tool and recurse. The in-process backends implement the optional absolute depth limit and scoped live-global `toolFilter`; ACP advertises both capabilities off and rejects such a request. The [subagent composition-controls Agent Note](2026-07-12-subagent-persona-tool-filter-and-depth.md) owns their exact semantics and security limits.
+- **Blocking the parent turn.** Foreground collection holds the parent's step open for the child's full duration. Background delegation uses the shared `ctx.tasks` runtime and generic `task_*` tools, the same collection mechanism as background bash; the subagent seam itself remains task-agnostic.
+- **Live progress.** This cut surfaces only lifecycle + final result; a per-chunk child→parent update stream is deferred with the background redesign.
+- **ACP client surface.** Proxying `fs`/`terminal` from the ACP child back to the parent (a shared-workspace mode) is future work; the first cut advertises neither, so the child self-serves in its own process.

.agents/notes/implemented/feature/2026-06-22-acp-subagent-backend.md

@@ -0,0 +1,55 @@
+# Agent Note: ACP subagent backend (out-of-process delegation)
+
+Status: implemented
+
+## Problem
+
+The subagent seam ([the seam Agent Note](2026-06-21-subagent-capability-seam.md)) was built so multiple backends coexist by name on `ctx.subagents`. The in-process backends (`-spawn`/`-fork`) run a child as a second `Agent` on the SAME cordis context — cheap, but the child shares the parent's process, model client, and tools. The seam's whole point was to also support an OUT-OF-PROCESS child reached over a protocol, proving the abstraction generalizes across a process boundary. This Agent Note adds the first such backend: an Agent Client Protocol (ACP) client.
+
+## Decision
+
+`@deepseek-ai/dsh-subagent-acp` registers a `SubagentProvider` that runs each child agent in a SPAWNED SUBPROCESS, driven over ACP as the *client*. It is the direction-inverted twin of the existing server-side bridge `@deepseek-ai/dsh-acp` (the ACP *agent*): the bridge ANSWERS `initialize`/`newSession`/`prompt`; this backend CALLS them and IMPLEMENTS the `Client` callbacks (`sessionUpdate`, `requestPermission`). Pointing the configured spawn command at the `acp-agent` example makes the harness talk to its own process.
+
+### Fresh process per run
+
+Each `start` spawns a new child, runs exactly one ACP session (`initialize` → `newSession` → `prompt`), and `dispose` kills the subprocess and awaits its exit. This is the simplest lifecycle and mirrors the in-process one-child-per-run shape.
+
+### Minimal client stub
+
+The client advertises NO optional capabilities (no `fs`, no `terminal`): the child self-serves file/terminal access in its own process. `session/update` notifications are consumed — the backend accumulates `agent_message_chunk` text as the result output and ignores the rest (thoughts, tool-call cards) in this cut, which surfaces only the child's final answer. `session/request_permission` is auto-answered by a configured policy (`reject` declines every prompt, `allow` approves via the first allow-shaped option) — the first cut surfaces no prompt to a human. Proxying `fs`/`terminal` back to the parent (a shared-workspace mode) remains future work, as the seam Agent Note noted.
+
+### No start-time capabilities
+
+The provider's `capabilities` are all `false`. An out-of-process child cannot honor the parent's `maxDepth` (it has no access to `parent.options.subagentDepth`) or `toolFilter` (it owns its own tool registry), and the first cut does not implement `outputSchema`. The service rejects a request needing any of them before `start` runs. The backend injects only `subagents` (not `ctx.agents`) and ignores `request.parent`.
+
+### StopReason mapping
+
+ACP `StopReason` → harness `SubagentStopReason`: `end_turn`→`completed`, `max_tokens`→`max-tokens`, `refusal`→`refusal`, `cancelled`→`aborted`, `max_turn_requests`→`error` (no clean equivalent — the task did not finish), unknown→`error`. A spawn/transport/RPC failure resolves `error` (or `aborted` if a cancel was requested); `result` never rejects on a child-level failure, per the seam contract.
+
+### Security: scrubbed child environment
+
+The child is a separate process, so it inherits an environment. Credential-shaped ambient vars (`/KEY|SECRET|TOKEN/i`) are NOT forwarded by default — the parent harness's own secrets must not leak into a spawned process implicitly (the same policy the bash executor applies). The child's OWN credentials (it needs a model key) are supplied EXPLICITLY via `config.env`, layered AFTER the scrub, so an intended `DEEPSEEK_API_KEY` survives while an incidental `AWS_SECRET_ACCESS_KEY` does not. Child stderr is inherited to the parent's stderr (diagnostics surface naturally); a spawn-level `error` event (e.g. ENOENT for a bad command) is captured and raced against the ACP drive, so a bad command settles `error` instead of crashing the parent with an unhandled error.
+
+## Testing
+
+- **Keyless unit/integration:** A scripted ACP subprocess exercises real stdio for prompt/output flow, every stop-reason mapping, signal and disposal cancellation (including pre-abort, pre-session race, and torn-pipe cases), both permission policies, ignored non-message updates, missing-command cleanup, provider reload, and namespace exports.
+- **With-key e2e:** The backend spawns the real ACP example; its model answers `PONG`, writes `proof.txt`, and the parent verifies the file.
+- **Snapshot gap:** Each ACP child is a separate process with its own replay session, unlike in-process per-session replay. Deterministic mock-server coverage exists, while `TODO(acp-subagent-replay)` tracks parent replay against a replaying child.
+
+## Alternatives considered
+
+### Why stay on SDK 0.25.1?
+
+The backend needs only `ClientSideConnection`, `ndJsonStream`, `PROTOCOL_VERSION`, and the client protocol types, all supported in 0.25.1. The 0.28 fluent API would require migrating both client and server connection classes across the ACP layer without improving this backend, so that upgrade remains a separate change.
+
+### Why not a persistent child process?
+
+Persistent-process pooling (reuse a warm child across runs) is a performance optimization deferred to future work — it adds session-lifecycle and crash-recovery complexity the first cut does not need; each `start` spawning a fresh child mirrors the in-process one-child-per-run shape.
+
+## Consequences
+
+Every run pays a fresh subprocess (spawn + `initialize` + `newSession`). The parent surfaces only the child's final answer: `session/update` thoughts and tool-call cards are consumed and dropped, and permission prompts never reach a human — the configured policy answers them. The child's environment is credential-scrubbed by default, so its own model key is supplied explicitly via `config.env`.
+
+## Future providers
+
+The same out-of-process spawn/prompt/stream/cancel shape generalizes to other transports named in the seam Agent Note — A2A, the Codex app-server, and the Claude Code Agent SDK — each a sibling provider registered by name. The ACP backend is the proof that the seam supports the boundary; those are mechanically similar.

.agents/notes/implemented/feature/2026-06-24-workspace-context.md

@@ -0,0 +1,87 @@
+# Agent Note: Workspace context instruction files
+
+Status: implemented
+
+## Problem
+
+Repository guidance such as `AGENTS.md` belongs in a coding session's effective context so project conventions, build commands, and review rules arrive without repeated user pasting. The stdio and ACP products need the same behavior, isolated by session cwd: a global system-prompt section leaks one workspace's files into another live ACP session.
+
+Neighboring products establish useful conventions but differ in details. Codex treats `AGENTS.md` as native, Claude Code uses `CLAUDE.md` and familiar system-reminder-style user context, and opencode supports both names with one winner per directory plus lazy nested discovery. The harness needs cross-tool compatibility without loading duplicate or contradictory files from the same scope.
+
+The lifecycle has two distinct classes of content. The initial applicable chain is stable enough to live in the request prefix and benefit from provider prefix caching. Nested files, edits, candidate switches, and removals happen after the session starts and belong in durable append-only history rather than the frozen prefix.
+
+## Decision
+
+The implementation lives in `packages/context/workspace-context` as `@deepseek-ai/dsh-workspace-context`. It is a request-context extension, not a core service or a filesystem backend. `@deepseek-ai/dsh-agent-core` mounts it for both product front doors and forwards its config. The plugin consumes `agent/session-prefix`, `tools/post-execute`, and the optional `ctx.fs` capability.
+
+The plugin does not statically inject `fs`. Providerless product trees therefore boot normally and the plugin no-ops until a filesystem provider exists. All production reads go through that provider. Candidate probes call `lstat` before `resolve`, so a repository-owned final-component symlink is rejected rather than followed outside the workspace. The session-prefix signal and dynamic tool execution signal propagate through resolution, metadata probes, and streaming reads, so cancellation does not wait for an unrelated filesystem scan. Once `lstat` identifies a regular-file winner, a provider exception or disagreement during resolve/stat is classified as unavailable: it is neither interpreted as a deletion nor allowed to fall through to a lower-priority candidate.
+
+### File Names And Precedence
+
+The default per-directory candidate list is `['AGENTS.md', 'CLAUDE.md']`. The list is configurable as `instructionFileCandidates`, and `AGENTS.md` is an ordinary first candidate rather than a hidden priority. In one directory, only the first existing regular-file candidate loads. With defaults, `AGENTS.md` is native and `CLAUDE.md` is a compatibility fallback.
+
+Candidate entries are same-directory file names. Empty entries, `.`/`..`, and entries containing `/` or `\` are ignored. Lowercase names, local variants, and other same-directory names can be opted into explicitly; rule directories and import semantics are outside this contract.
+
+The user-global file is fixed at `$DSH_HOME/AGENTS.md` and is not affected by `instructionFileCandidates`. `$DSH_HOME` defaults to `~/.dsh`, matching the harness-level home role of `~/.codex` or `~/.claude` rather than introducing a plugin-specific home. Tilde expansion and the default live in `dsh-paths` so future harness features share the same convention.
+
+### Baseline Prefix
+
+On the first request of an agent-loop instance, the plugin contributes one user-role message through `agent/session-prefix`. It loads the user-global file first, then finds the project root by walking upward from `agent.session.header.cwd` to a configured root marker (default `.git`), then loads one candidate from each directory from the root to the cwd. A `.git` file and a `.git` directory are both valid markers, covering linked worktrees and submodules. Without a marker, the cwd itself is the root.
+
+The plugin prepends its contribution before `await next()` returns, so session-prefix contributions appear in plugin registration order. In the product spine workspace instructions are registered before a skills catalog and therefore appear first. The loop deep-freezes the composed prefix, logs it in `EpochHeader.messagePrefix`, and reuses it verbatim for that instance. It is request state, not `Session.deriveMessages()` history.
+
+A resumed agent creates a new loop instance and recomposes the baseline from current files, with the new prefix anchored by the resume request header. This permits current baseline content on resume without mutating a prefix already used by an earlier instance.
+
+The baseline is a user-role `<system-reminder>` with `Instructions from: <path>` sections and explicit authority and precedence language. This familiar model-facing frame avoids a harness-specific XML vocabulary. Project paths are root-relative and the user-global path is `~/.dsh/AGENTS.md` for the default home or `$DSH_HOME/AGENTS.md` for a configured home. A literal `</system-reminder>` inside file content is escaped. The package README owns the exact current [prompt shape](../../../../packages/context/workspace-context/README.md#prompt-shape).
+
+### Dynamic Discovery And Refresh
+
+After a successful first-party `read`, `write`, or `edit` call, the `tools/post-execute` listener reconciles the touched descendant chain and every scope already known to the session. A newly reached scope is returned through `additionalContexts` for the next request using an `Additional instructions from: <path>` system-reminder. Under Code Mode, `run_code` defers sub-dispatch contexts onto its outer result, so the same update is appended only after the parent result rather than being injected mid-call.
+
+A content edit appends `Updated instructions from: <path>`, states that the new content replaces the previous content, and includes the complete current file. If precedence changes from one candidate to another, the message also names the previous path and says it no longer applies. If no candidate remains, the plugin appends `Instructions removed: <path>` and states that the previously loaded instructions no longer apply.
+
+Dynamic messages use a raw `context/message` envelope because the plugin owns the complete system-reminder framing. Core context injection therefore supports `envelope: 'raw'`; callers that omit it retain the canonical `<context source="...">` wrapper. `context/message.meta` carries opaque JSON state that is persisted but never rendered to the model.
+
+Shell commands are not discovery triggers. Local bash calls start fresh shells, and inferring reached paths from arbitrary command strings would require shell semantics the prompt plugin does not own.
+
+### Duplicate Suppression And Change Detection
+
+Every dynamic workspace context event stores versioned metadata with `{ action, scope, path, previousPath?, digest? }`, where `digest` is SHA-1 over the loaded content. The model-facing prompt has no HTML comments, hidden markers, or headings that are parsed back into state.
+
+At reconciliation time the plugin scans plugin-owned `context/message` events and derives the latest state for each visible scope. A short per-session pending map begins only after the immutable top-level `tools/result` proves an `additionalContexts` entry survived every post-execute listener, then covers the interval before the loop appends that context to the log. Each entry records the open `{ turn, step }`: an equal durable `context/message` at or after its sequence boundary confirms and removes it, while a matching `step/end` arriving first means the loop discarded its context buffer, so the plugin removes both the pending entry and its version-cache fast path. A nested Code Mode result stages its changes under the parent's opaque execution token so repeated sub-dispatches in one run do not duplicate them; the parent result rolls that provisional state back and commits only contexts retained by outer policy.
+
+An unchanged path and digest is suppressed. A logged removal is a tombstone, so a reappearing candidate becomes a new `set`. Resume works from persisted metadata. If compaction removes an instruction event from the visible surface, that state no longer suppresses a later load, matching the fact that the model can no longer see it. Only changes actually included under the byte budget enter metadata or pending state, so an omitted file remains eligible on a later touch.
+
+The frozen baseline keeps an in-memory path/digest map for comparison. A later successful filesystem touch appends baseline edits or removals as dynamic messages; it never rewrites the prefix. During resumed prefix composition the plugin also reconciles visible dynamic scopes, so nested changes made while the agent was offline can append an update before the first resumed request.
+
+There is intentionally no watcher. Detection occurs at the next successful structured filesystem touch or resumed prefix composition. A provider failure produces no removal; absence is only accepted when all configured candidates in that scope were probed successfully.
+
+### Byte Budget And Bounded Reads
+
+`maxBytes` is required and applies separately to a rendered baseline or one dynamic reconciliation batch; there is no implicit or unbounded render budget. Non-positive and non-finite values disable loading. When content exceeds the budget, broader files are omitted before the most-specific file is truncated. A visible `Workspace instruction budget ...` notice names omitted and truncated paths and byte counts, and output never exceeds the configured bytes.
+
+`maxSourceBytes` is a positive per-file cap with a 1 MiB default. The loader checks reported size before reading and still consumes content through `streamText()` with a running UTF-8 byte count, so missing/stale metadata cannot force an unbounded allocation. An oversized winning candidate is unavailable rather than a reason to fall through to another same-directory name. The plugin deliberately keeps no process-wide cache and never retains instruction prose. It keeps only `{ path, version, digest }` per effective scope in a `WeakMap<Session, Map<scope, state>>`: a matching provider `FsVersion` plus matching effective prompt state skips the read, while a changed version triggers a bounded read and SHA-1 confirmation. SHA-1 remains the cross-provider content identity persisted in visible structured metadata; provider versions are only an in-memory invalidation fast path. Cache transitions for model-visible changes commit only when the corresponding context survives the complete tool-result policy chain, and are invalidated if that accepted context is later dropped with its aborted step before reaching the log.
+
+## Alternatives considered
+
+**Use a global `ctx.systemPrompt.section()`.** Rejected because one Cordis context can host sessions with different cwd values, while repository-owned text is lower-authority context rather than top-authority provider system content.
+
+**Inject the baseline on every `agent/pre-step`.** Rejected because repeated history injection wastes tokens, complicates duplicate state, and prevents a structurally stable provider prefix. Prefix composition gives a frozen, logged, per-instance baseline while dynamic append-only messages handle changes.
+
+**Load both `AGENTS.md` and `CLAUDE.md` in one directory.** Rejected because repositories in transition commonly duplicate guidance across both files. Ordered candidates make precedence explicit and configurable.
+
+**Parse rendered headings or hidden comments to recover loaded state.** Rejected because instruction prose can contain the same text, causing silent false positives. Persisted JSON metadata provides an unambiguous state channel that is invisible to the model.
+
+**Summarize files with a model.** Rejected because instruction files are already curated summaries; another model call is nondeterministic and can erase edge-case requirements. Deterministic full text with byte budgeting is simpler.
+
+## Consequences
+
+Workspace guidance is isolated per session and shared by both product front doors and every tool presentation mode. Initial instructions benefit from stable prefix caching, while nested and changed content remains durable and replayable. The generic session/agent context contract includes optional raw framing and JSON metadata, both propagated through prompt-submit and post-tool `additionalContexts` arrays without flattening entries.
+
+Repository text remains untrusted input. Lower-authority user-role framing, explicit precedence language, delimiter escaping, and symlink rejection reduce risk but do not eliminate prompt injection. Permission and sandbox layers treat workspace files as data rather than authority.
+
+The system is event-driven rather than watch-driven. Edits are not visible at the exact filesystem mutation instant unless that mutation goes through a structured tool; externally changed files are noticed on the next successful structured touch or resume. This keeps the design deterministic and provider-neutral.
+
+## Deferred
+
+Bash-derived path reporting, recursive startup scans, file watchers, lowercase defaults, `.claude/CLAUDE.md`, `.claude/rules/*.md`, import directives, ACP `additionalDirectories`, trust acknowledgements, and model-generated summaries are deferred. Same-directory private variants can be configured today; directory rule systems and imports need their own precedence and trust designs.

.agents/notes/implemented/feature/2026-06-25-ask-user-question.md

@@ -0,0 +1,49 @@
+# Agent Note: Ask-user question capability
+
+Status: implemented
+
+## Problem
+
+The agent sometimes cannot proceed safely from model inference alone: it needs the human to choose a path, confirm a risky/default action, or provide missing information. Before this change, the only way to get that answer was for the model to ask in assistant text and then stop, which broke the normal tool-call loop: the agent had no structured way to pause, no option metadata for UIs, no abort/error taxonomy, and no way for non-stdio front doors to present the question consistently.
+
+This is a user-facing capability, but it also crosses package boundaries. A model-facing tool needs a provider-neutral request vocabulary; each UI surface needs to decide how to show and collect the answer; the agent loop should remain unchanged because a tool call already has the right async shape.
+
+## Decision
+
+Introduce `dsh-user-interaction` as the provider-neutral interface package for `ctx.userInteraction`, colocated with the model-facing consumer `dsh-tool-ask-user` under `packages/ui`. The grouping is intentional: asking a human is a UI-backed product affordance, not part of the providerless core spine. The seam still owns the stable request/answer/error vocabulary, while UI product surfaces provide the concrete provider that collects the answer. The tool registers `ask_user_question`, forwards `{ questions, agent, signal }`, and returns the provider-computed structured answers as the tool result.
+
+The model-facing request vocabulary is deliberately aligned with the product-research schema: `ask_user_question({ questions: [{ id, question, header?, options?: [{ label, description? }], multi_select? }] })`. `id` is supplied per question and echoed in the result so a batch can be routed without relying on question text. `label` is both user-facing display text and the selected value returned to the model; there is no separate `value`, no `recommended`, no `allow_custom`, and no `desc` alias.
+
+Providers return `{ answers: [{ id, selected, custom? }] }`. `selected` is always an array of selected option labels, so single-select and `multi_select` answers share one result shape. `custom` carries a free-text "Other" answer; optionless questions collect `custom` directly. When `custom` is present, it overrides any selected choices and `selected` is empty.
+
+`UserInteractionError` extends `HarnessError`, so failures such as `NO_PROVIDER`, `ASK_ABORTED`, ACP cancellation, or missing session routing survive `ctx.tools.execute()` as machine-routable `{ name, code }` tool errors. This matches the structured-error taxonomy and lets the model or a wrapping plugin distinguish "user cancelled" from a generic thrown exception.
+
+## UI mappings
+
+`dsh-stdio-demo`'s in-package readline module renders each question, shows each option's `description` on the next line, supports comma/space-separated numeric choices for `multi_select`, accepts free-form custom answers, and rejects pending questions on abort, provider disposal, or stdin EOF. A batched request is asked in order and resolved as one answer object. The stdio provider serializes simultaneous requests with an internal queue so only one prompt owns stdin at a time.
+
+`dsh-acp` provides the same seam for ACP sessions. It resolves the calling `Agent` through `ownedRecord`, requiring the forward session-map record at `agent.session.id` to own that exact agent object, and calls ACP `unstable_createElicitation` with a session-scoped form for each question. Single-select options become a `choice` string enum; `multi_select` options become a `choice` array enum; optionless questions use a required `custom` text field. If the client returns both `choice` and non-empty `custom`, the custom answer wins. ACP `decline`/`cancel`, a missing answer, a missing session, and a client without elicitation support all become structured `UserInteractionError`s.
+
+The ACP mapping deliberately uses elicitation, not `session/request_permission`. `request_permission` is still reserved for the separate permission gate: it is a yes/no-or-policy authorization protocol around tool execution. `ask_user_question` is a general information-gathering tool with optional free-form answers, so ACP form elicitation is the closer protocol fit. The bridge's session routing is shared with the future permission gate, but the user intent is different.
+
+## Alternatives considered
+
+**Assistant text followed by a stopped turn.** The model could ask the user in plain assistant text and then stop. That loses the structured option metadata, gives UIs no provider-neutral way to render a choice, and forces the next human answer to arrive as a new user prompt rather than as the result of the operation that needed the answer.
+
+**Core-owned ask-user packages.** The first implementation split the seam and the model-facing tool across `packages/core` and `packages/ui`, but both names describe one UI-backed human-interaction affordance. The seam remains provider-neutral, but it is not providerless core infrastructure like sessions, tools, or the agent registry. Keeping `dsh-user-interaction` and `dsh-tool-ask-user` together under `packages/ui` makes the package map match the product boundary: apps and bridges provide the human-answer provider, and the stdio app opts into the model-facing tool.
+
+**ACP `session/request_permission`.** Permission requests are authorization around tool execution; `ask_user_question` is information gathering with optional free-form answers. Using permission for general questions would collapse two different product concepts and make the future permission gate harder to reason about.
+
+**A loop-level pause primitive.** The agent loop already knows how to await a tool call and resume from a tool result. Adding a new loop special case would duplicate that async shape and make every loop implementation learn about a UI concern.
+
+## Consequences
+
+ACP elicitation is currently marked unstable in the SDK. The fallback is still structured: if a client does not implement it, the tool returns `ASK_FAILED` rather than hanging. A later ACP stabilization may rename or reshape the method; that migration should stay inside `dsh-acp` because the core `ctx.userInteraction` vocabulary is provider-neutral.
+
+The feature gives the model a powerful pause primitive, so prompt guidance matters. The tool description tells the model to ask concise questions and use options when possible. Product policy can later wrap `tools/execute` to restrict when the tool is allowed, but the loop should not special-case it.
+
+`dsh-user-interaction` and `dsh-tool-ask-user` both live in `packages/ui` because they form one product-facing human-interaction capability. `agent-core` does not load either the tool or a provider. `stdio-agent` opts into the seam, its readline provider, and the model-facing tool. `acp-agent` keeps only the `userInteraction` seam/provider by default: ACP elicitation support is still client-dependent, so an ACP leaf must opt into the model-facing tool deliberately once its client can complete elicitation requests.
+
+## Testing
+
+Unit coverage pins provider registration/disposal, duplicate-provider rejection, abort-before-provider, empty-question rejection, structured tool errors through `ctx.tools.execute()`, batched answers, multi-select answers, custom answers, and the model schema including the removal of `value`, `recommended`, `allow_custom`, and `desc`. `dsh-stdio-demo` tests cover option descriptions, queued requests, EOF/abort cleanup, optionless free-form input, invalid option reprompts, duplicate multi-select numbers, and batched question flows. ACP bridge tests drive a real in-memory ACP connection with the real `ask_user_question` tool and verify selected-option, custom-overrides-choice, multi-select, and optionless free-form elicitation paths continue the agent loop.

.agents/notes/implemented/feature/2026-06-29-todo-write-tool.md

@@ -0,0 +1,62 @@
+# Agent Note: The `todo_write` tool — model task list as event-sourced session state
+
+Status: implemented
+
+## Problem
+
+The harness gives the model bash and subagent tools but no way to record a structured task list. A todo list serves two co-equal purposes: it steers the model to plan multi-step work and keep the active task unambiguous (at most one active, exactly one while work remains), and it gives the human a live progress checklist. The ACP protocol has a native `plan` sessionUpdate that editors (Zed) already render, but the bridge never emitted one. Every reference coding agent surveyed (claude-code, opencode, codex, oh-my-pi, pi) ships some form of this; the harness had nothing.
+
+## Decision
+
+Add a model-facing `todo_write(todos: [{ content, status }])` tool whose whole-list state lives on the event-sourced session log as a new `todo/write` `SessionEventMap` variant. Both the stdio UI and the ACP bridge render off the existing `session/event` — the ACP bridge maps the list to a `plan` sessionUpdate.
+
+### Whole-list replace, three-state status
+
+The model sends the ENTIRE list every call; the new list replaces the old (last-write-wins on replay). This is the shape claude-code V1, opencode, and codex `update_plan` all use, and the shape the model is most trained on — no per-item ids, no delta protocol. `status` is exactly `pending | in_progress | completed`: the same triple as codex `update_plan` and, crucially, **identical to the ACP `PlanEntryStatus`**, so the bridge maps it 1:1 with no lossy translation.
+
+### State on the session log, not a service
+
+The list is appended as a `todo/write` event carrying the full `{ todos }` snapshot. The harness is event-sourced — the LLM history, tool calls, and turn structure all live on the log — so the todo list lives there too. This buys durability, replay, and `session/load` reconstruction for free: a reopened session re-derives the current list (the last `todo/write`) and the ACP bridge re-emits the `plan` on load, with no separate persistence backend, no in-memory service to rehydrate, and no extra wiring. An in-memory `ctx.todos` service would have had to reinvent all of that.
+
+### NOT a surface event
+
+`todo/write` is deliberately excluded from `SurfaceEventType`. The surface is the projection that produces the LLM message history (`deriveMessages()`); a todo write produces no conversation message. So it carries no `surfaceOp`, never joins the ordered surface, and never reaches `deriveMessages()` — it is durable, replayable *UI* state that travels alongside the conversation without being part of it. (The dev-mode invariants still require it to sit inside an open turn, which it always does: it is appended mid-step during a tool call.)
+
+### Priority synthesized only at the ACP boundary
+
+ACP's `PlanEntry` requires `content` + `priority` + `status`, but a `TodoItem` has no priority — the model never reasons about it. Rather than burden the schema with a field the model must always supply, the bridge synthesizes a constant `priority: 'medium'` on every entry when it builds the `plan`. Priority is an ACP wire requirement, not a harness concept, so it lives at exactly the boundary that needs it.
+
+### Dropped vs claude-code V1: `activeForm`, id, priority
+
+claude-code V1's item is `{ content, status, activeForm }`; later (V2) it grew ids, dependencies, and ownership — but only to support agent *swarms* (disk-backed, lock-guarded, per-item mutation). This tool keeps the item at the minimum: `{ content, status }`. No `activeForm` (the present-continuous label) — the UI shows `content`; no id — whole-list replace needs no stable identity; no priority — see above. Each dropped field is one less thing the model must produce on every call.
+
+### Single owner — no swarm machinery (YAGNI)
+
+Each list belongs to the calling agent session, and non-agent calls are rejected. There is no shared scope, resolver, or delta protocol. Cross-agent lists would require per-item log deltas and explicit scope selection, so they remain a separate future design.
+
+### Validation: the cheap middle
+
+The schema enforces type/required/enum. Beyond that, `execute` rejects empty or duplicate `content` and more than one `in_progress` task. claude-code leaves single-in-progress to the prompt; oh-my-pi enforces it in code. We take the middle: enforce the cheap invariants that make a plan *coherent* (no blank tasks, no dupes, at most one active), but leave ordering and the discipline of keeping the list current to the model via the tool description. A rejected write returns an `isError` result so the model self-corrects.
+
+## Why no cordis-catalog entry / no `@mode`
+
+`todo/write` is a member of `SessionEventMap`, not a first-class cordis `interface Events` event. The catalog generator (`scripts/gen-cordis-catalog.ts`) scans `interface Events` declarations; a `SessionEventMap` variant rides the existing `session/event` emit and produces no new catalog row. So it carries no `@mode` tag (which the generator requires only on `interface Events` members) — adding one would be meaningless.
+
+## Testing
+
+Four tiers, designed up front:
+- **Unit** — the session event (append/snapshot-clone/last-write-wins/not-on-surface); the tool (schema shape, arg validation via the real `ctx.tools.execute`, value validation, the event append + replacement, no-agent rejection, `presentCall`, HMR-safety); the ACP `todosToPlan` mapping; the stdio render arm.
+- **Real-Loader path** — the plugin run through `Loader.unwrapExports`, asserting the namespace export shape survives (it HAS `inject`, so a stray default would crash at load — postmortem/0001).
+- **Full-loop integration** — a scripted mock model calls `todo_write` through the real agent loop; the `todo/write` event lands and a second call replaces it.
+- **`session/load` replay** — a persisted `todo/write` re-emits the `plan` update when a fresh ACP bridge loads the session.
+- **With-key e2e + snapshot** — a real prompt induces a `todo_write`; the snapshot expected output gains the `plan` notification and the log event.
+
+## Alternatives considered
+
+- **In-memory `ctx.todos` service** — would reinvent durability, replay, and `session/load` reconstruction the log gives for free.
+- **Per-item delta protocol** — only needed for a shared multi-owner list, which is out of scope; whole-list replace is simpler and matches the references.
+- **Tool in `core/`** — `todo_write` is an extension tool registering on `ctx.tools`, not part of the spine; it lives in its own `packages/todo/` group like other tool families.
+
+## Consequences
+
+The todo list is durable, replayable session state: a persisted `todo/write` re-emits the editor's `plan` update on `session/load`, and the log — not plugin memory — is the single source of truth. Whole-list replace means one tool call per update with last-write-wins; there is no delta protocol to reconcile. The event stays off the surface, so a todo update never perturbs the derived model history — the model sees only its own tool call and result.

.agents/notes/implemented/feature/2026-06-30-hook-bridges.md

@@ -0,0 +1,68 @@
+# Agent Note: dsh-hooks-claude + dsh-hooks-codex — the Claude Code / Codex hook bridges
+
+Status: implemented
+
+## Problem
+
+The harness's extension surface is its typed interception seams ([the interception-seams Agent Note](2026-06-30-interception-seams.md)): a "native hook" is just an ordinary cordis plugin subscribing to `agent/session-start`, `agent/prompt-submit`, `tools/pre-execute`, `tools/post-execute`, `agent/turn-continuation`, `subagent/start`, `subagent/end`. But users arrive with **existing** Claude Code (CC) and Codex hook configs — a `hooks.json` (or a settings file's `hooks` key) full of shell-command hooks — and want those to run unmodified. This Agent Note introduces the two **bridge plugins** that translate that external shell-hook protocol onto the typed seams, built on the shared wire-protocol library ([the hook-protocol-lib Agent Note](2026-06-30-hook-protocol-lib.md)).
+
+The framing that shapes the whole design: **a bridge is a compatibility adapter, not a power tool.** Anything a bridge does (block a tool, inject context, force continuation, observe a subagent) a native cordis plugin does more powerfully — typed returns, full `ctx`, no serialization boundary. The bridge's reason to exist is to run the explicitly supported subset of external CC/Codex command hooks. That keeps each bridge thin: parse the config, pick a matcher mode, build the per-event payload, call `runHook` + `mergeHookOutputs` from the shared lib, and map the neutral outcome onto a seam Decision. The package READMEs own the exact current unsupported-event and partial-field inventory against the official protocols.
+
+## Decision
+
+Two independent plugins in the `packages/hooks/` group, each a function/namespace plugin (`name`/`inject`/`Config`/`apply`, NO default export — see [postmortem 0001](../../../../docs/postmortem/0001-acp-default-export-drops-inject.md)) injecting only `bash`:
+
+- **`dsh-hooks-claude`** — the CC dialect. Seven of Claude Code's current hook points: `SessionStart`, `UserPromptSubmit`, `PreToolUse`, `PostToolUse`, `Stop`, `SubagentStart`, and `SubagentStop`. Owns CC-shaped per-event stdin payloads (a base of `session_id`/`transcript_path`/`cwd`/`hook_event_name` plus per-event fields), `CLAUDE_PROJECT_DIR` plus `${CLAUDE_PLUGIN_ROOT}`/`${CLAUDE_PROJECT_DIR}` substitution, and the literal-or-regex matcher mode. `transcript_path` is the persistence locator result or `''`; stdin carries a **trailing newline**.
+- **`dsh-hooks-codex`** — five of Codex's current hook points: `PreToolUse`, `PostToolUse`, `SessionStart`, `UserPromptSubmit`, and `Stop`. It uses an always-regex matcher, Codex-shaped snake_case payloads with `turn_id`/`model`/`permission_mode` extras written WITHOUT a trailing newline, no Codex plugin-env injection or config-time placeholder substitution, and no pre-tool approval or rewrite path. `transcript_path` is the same locator result or `null`; tool payloads carry the real `tool_name` in the reduced `tool_input: { command }` shape.
+
+### Outcome → Decision mapping
+
+Each bridge maps the neutral `MergedHookOutcome` from the shared lib onto the seam's typed Decision:
+
+| Seam | CC | Codex |
+|---|---|---|
+| `agent/session-start` (emit) | additionalContext → `agent.inject()` | plain-stdout output → additionalContext → `agent.inject()` |
+| `agent/prompt-submit` | `deny`→`block`; context-only→delegate+fold | `block`→`block`; context-only→delegate+fold |
+| `tools/pre-execute` | `deny`→`deny`; `ask`→`ask` | `block`→`deny` (no allow/ask) |
+| `tools/post-execute` | `deny`→`block`+feedback; context-only→delegate+fold | same |
+| `agent/turn-continuation` | blocking Stop → `continue` (reason = next-step steering) | same |
+| `subagent/start` (emit) | additionalContext → inject into a live in-process child; a remote child has no local injection target | unsupported by this bridge |
+| `subagent/end` (emit) | observe-only | unsupported by this bridge |
+
+The CC bridge's `ask` result is a real permission path, not a terminal bridge decision: `dsh-tools` resolves it through the optional [approval seam](2026-07-06-approval-seam.md). A composed ACP answerer prompts the owning editor session and `allowed-once` proceeds; without an ApprovalService or answerer, the call fails closed to `deny`.
+
+### Context source is always the plugin (the mislabel guard)
+
+`agent.inject()` defaults a missing `MessageSource` to `{ kind: 'user' }`, so every bridge `inject()` and `HookContext` passes `{ kind: 'plugin', plugin: 'hooks-claude' | 'hooks-codex' }`. Unit coverage pins the resulting `context/message.source` as the plugin rather than the user.
+
+### Adding context is not a veto — delegate, then prepend
+
+A hook that only attaches `additionalContext` (no block/deny) is NOT a decision the bridge should return on its own: returning `allow`/`accept` from a waterfall listener WITHOUT calling `next()` short-circuits every later `agent/prompt-submit` / `tools/post-execute` listener, so a policy/sandbox plugin registered after the bridge would never see the prompt. Each bridge therefore delegates via `next()` before adding its context to the downstream decision. Both seams carry ordered `additionalContexts` arrays, so the bridge prepends its separately sourced entry while preserving every downstream source, envelope, and metadata field; a downstream prompt block still drops all context because the prompt never reaches the model, while post-tool block semantics may explicitly retain contexts. Code Mode ferries the same array through the outer `run_code` result. Only a real `deny`/`block` from the hook itself short-circuits. Tests assert a later listener can still block a prompt a context-only hook allowed and that retained prompt and post-tool contexts remain separate.
+
+### CLAUDE_PROJECT_DIR defaults to the session workspace
+
+Claude Code always exports `CLAUDE_PROJECT_DIR`, and common unmodified hooks reference `$CLAUDE_PROJECT_DIR` for project-relative paths. An explicit `config.projectDir` wins; when it is omitted (the default ACP wiring configures only `configPath`), the bridge defaults the env var per-run to the agent's session workspace — the same `session.header.cwd` the hook already runs in — rather than leaving it empty. So a stock project-relative hook works in the default setup.
+
+### Containment
+
+The config is parsed ONCE at load; a read/parse failure logs and registers nothing rather than crashing boot (a typo'd path must not take the agent down). Only shell-form `type: 'command'` hooks run for CC; `http`, `mcp_tool`, `prompt`, and `agent` handlers are parsed-and-skipped. Codex runs only synchronous command handlers and skips `async: true` or non-command entries. The emit-listener paths (`session-start`, `subagent/start`) run detached, with their `inject` contained in a `.catch` that logs (a throwing inject must not break session boot or the loop).
+
+### Where hooks run, and where their config comes from
+
+Hooks run in the agent's session workspace, so relative paths target the user's project. `configPath` is resolved once against the process launch cwd and applies to every session. Per-session project-local discovery remains deferred under `TODO(per-session-hook-config)`.
+
+## Deferred compatibility gaps
+
+- **Tool-input rewrite.** A CC/Codex `updatedInput` is logged + warned, not honored — input rewrite is a deferred consistency-design problem ([the pre-tool-input-rewrite Agent Note](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md)), because the pre-execution args are read by `tool/call` audit + `assistant/message` history + ACP/tool-bash presentation, so an honest rewrite is a design unit, not a field.
+- **Stop loop-guard** (`TODO(stop-loop-guard)`). Claude Code supplies `stop_hook_active` and overrides a hook after eight consecutive blocks; Codex supplies `stop_hook_active` but documents no equivalent cap. Both bridges always report `false`, so a Stop hook that unconditionally blocks force-continues every step — a hook author must self-limit until state tracking lands.
+- **Hook `continue:false` (hard halt).** A hook can ask to halt the whole run (CC/Codex `continue:false`); the shared merge folds it into `MergedHookOutcome.stop`/`stopReason`, but no bridge acts on it (`TODO(hook-continue-false)`) — the interception seams have no "hard-halt the agent" primitive yet (a Decision blocks/steers a single point, not the run). Deferred with the loop-guard work; the halt request is recorded in the `hook/result` log, and the hook keeps its per-point effect (decision/context) meanwhile.
+- **Config discovery.** The path is explicit in `cordis.yml` and process-level (see above); the full multi-layer CC/Codex precedence walk, per-session project-local discovery, and the trust/hash model are not reimplemented (`TODO(per-session-hook-config)`).
+- **Session-start / subagent-start context is best-effort (`TODO(session-start-gating)`).** Both hooks run detached from startup, so their context is injected when ready but may miss the first request or a short-lived child. Guaranteeing first-request delivery requires an awaited startup seam.
+
+## Alternatives considered
+
+**Concurrent per-point hook execution.** The reference engines run a point's matched hooks concurrently and fold the results. These bridges run them **serially** (`await` per hook inside the match loop) and fold with the same most-restrictive merge. Serial is deliberate: it keeps each hook's `hook/invoked`/`hook/result` pair adjacent and in a deterministic order in the session log, and the fold is order-independent for the decision (`deny > ask > allow`) so the outcome matches. The cost is latency (hook *N* waits for hook *N−1*) and that per-hook timeouts are not overlapped — acceptable for the hook counts real configs use; revisit if a config ever fans out enough for the wall-clock to matter.
+
+## Consequences
+
+Matcher semantics, exit-code handling, and merge precedence live in `dsh-hook-protocol`; each bridge only parses config, builds dialect payloads, and maps outcomes. Per-file coverage includes config branches plus end-to-end mappings through a real loop, `dsh-bash-local`, and shell scripts, while a real-Loader smoke guards the package export shape. Native plugins bypass the wire protocol and return typed decisions directly.

.agents/notes/implemented/feature/2026-06-30-hook-protocol-lib.md

@@ -0,0 +1,30 @@
+# Agent Note: dsh-hook-protocol — the shared Claude Code / Codex hook wire-protocol core
+
+Status: implemented
+
+## Problem
+
+The hooks subsystem ships two bridge plugins: one that runs a user's existing Claude Code (CC) hooks, one for Codex hooks. Studying the reference implementations (`~/repos/refs/claude-code`, `~/repos/refs/codex`) surfaced a decisive fact: **Codex deliberately reimplements a SUBSET of the CC hook protocol.** Its engine reads the same `hooks.json`, uses the same matcher-group shape, the same exit-code/structured-stdout output contract, and the same command-hook execution model — Codex's source even names the engine after Claude's and comments where it "intentionally diverges." So the two bridges would otherwise duplicate the bulk of the protocol.
+
+This Agent Note introduces `@deepseek-ai/dsh-hook-protocol`, a **library** (not a plugin — it registers and injects nothing) holding the genuinely-identical primitives both bridges build on. The split between shared and per-dialect is the design's center of gravity.
+
+## Decision
+
+A new `packages/hooks/` group with `hook-protocol` as a pure library. It owns four primitive families and the `hook/*` session events; each bridge plugin (`dsh-hooks-claude`, `dsh-hooks-codex`) owns what genuinely differs.
+
+**Shared (here):**
+- **Matcher** — `matchesMatcher(pattern, query, mode)`. The ONE axis the dialects differ on is collapsed to the `mode` parameter: `claude` treats a pure `[A-Za-z0-9_|]+` pattern as a literal (pipe = exact-match alternation) and anything else as a regex; `codex` is always an unanchored regex. Match-all on absent/`''`/`'*'`; an invalid regex matches nothing (never throws into the loop).
+- **Execution** — `runHook(bash, hook, options)`. Runs a command hook through the `ctx.bash` seam rather than a bespoke `spawn`: the executor already provides the scrubbed-but-overridable env, process-group kills, and timeout the protocol needs, and `dsh-bash`'s `stdin`/`env` fields (added for exactly this) are the trusted-plugin surface an in-process bridge is allowed to use. It serializes the bridge-built payload to stdin (trailing newline iff CC), honors the hook's `timeoutSec` (else `DEFAULT_HOOK_TIMEOUT_MS`, the 10-minute reference default both dialects share), and never throws (an executor rejection becomes a non-blocking-error `HookOutput`).
+- **Decode** — `parseHookOutput(exit, stdout, stderr)`, the exit-code + structured-stdout codec, producing a dialect-neutral `HookOutput`. Exit `0` → lenient JSON parse of stdout; exit `2` → blocking error with `stderr` as the reason (surfaced as `decision: 'block'` so no caller needs a separate exit-code branch); other → non-blocking error. Parses the CC structured-stdout fields that have a consumer on some path (`continue`/`stopReason`/`decision`/`hookSpecificOutput.{permissionDecision,additionalContext,updatedInput}`/`systemMessage`); the bridge honors only the subset meaningful for its dialect. Fields with no consumer on any path are not parsed at all (CC's `suppressOutput` — hook stdout never enters a transcript here, so there is nothing to suppress; see [the tighten-hook-protocol-contract Agent Note](../simplification/2026-07-04-tighten-hook-protocol-contract.md)).
+- **Merge** — `mergeHookOutputs(outputs)`, folding multiple matched hooks into one most-restrictive `MergedHookOutcome`: permission precedence **deny > ask > allow**, halt sticky on the first `continue:false`, block reasons joined `\n\n`, context/system-messages accumulated in order.
+- **`hook/*` session events** — `hook/invoked` / `hook/result`, declaration-merged into `SessionEventMap` (log-only, like `compact/*` — NOT `SurfaceEventType`s), with `appendHookInvoked`/`appendHookResult` helpers so the invoked/result pairing and turn-enclosure stay consistent across bridges. `appendHookResult` also owns the durable record's semantics — the decision string (the hook's parsed decision, else `'stop'` on `continue:false`, else `'pass'`) and the 500-character `stderrSummary` truncation derive from the `HookOutput` here, not per-bridge.
+
+**Per-dialect (the bridge plugins):** building each event's stdin payload (CC's base+per-event field sets vs Codex's snake_case with `turn_id`/`model` extras), the dialect's env + `${CLAUDE_PLUGIN_ROOT}` substitution (CC) vs none (Codex), and mapping the neutral `HookOutput`/`MergedHookOutcome` onto the harness's seam-specific typed Decisions (`PreToolDecision`, `PromptDecision`, `ContinuationDecision`, `PostToolDecision`).
+
+## Alternatives considered
+
+**One parameterized engine.** Rejected because payload construction and decision mapping genuinely differ by dialect. Matchers, codecs, execution, merge rules, and events remain shared; each bridge keeps its payload and mapping explicit so its wire behavior is readable in place.
+
+## Consequences
+
+Each bridge parses config, builds its dialect payload, invokes the shared runner and merge logic, maps the decision, and appends `hook/*`. Protocol tests cover every matcher mode, exit-code and codec field, runner plumbing, merge precedence, and audit helper at per-file 100%; bridge tests exercise the library's real load path. `updatedInput` is parsed but only logged and warned until the [input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md) lands.

.agents/notes/implemented/feature/2026-06-30-interception-seams.md

@@ -0,0 +1,58 @@
+# Agent Note: Interception seams — the typed-Decision surface a hook programs against
+
+Status: implemented
+
+## Problem
+
+The harness needs a hooks subsystem: users extend or gate the agent at lifecycle points the way Claude Code (CC) and Codex do. The key reframe driving this design is that **"native hooks" are not a package** — a native hook is just an ordinary Cordis plugin subscribing to the canonical lifecycle events. So the real product is a *powerful, well-typed canonical event surface*; the CC/Codex bridges (the `dsh-hooks-claude` / `dsh-hooks-codex` packages) are merely translators that map an external shell-hook protocol onto that same surface. Anything a bridge can do, a plain plugin can do directly — more powerfully (no serialization boundary, full `ctx`, typed returns).
+
+The surface needs distinct contracts for per-prompt policy (CC's `UserPromptSubmit`), session-start observation (CC's `SessionStart`), pre-tool policy, around-dispatch control, post-tool transformation, final-result observation, and continuation with a model-facing reason. Conflating those phases gives plugins mutation channels they do not need and makes finality depend on listener ordering. The [event-domain-semantics Agent Note](../architecture/2026-06-30-event-domain-semantics.md) supplies the three-domain rule and the typed-Decision idiom; this Agent Note applies them to the lifecycle seams.
+
+## Decision
+
+The canonical surface separates transformable policy, around-dispatch control, and observe-only notification. Policy waterfalls return small seam-specific **typed Decision unions**; wrappers return normalized results; notifications receive immutable snapshots and cannot affect the outcome. The set covers the hook points in scope (`session-start`, `prompt-submit`, `pre-tool`, `post-tool`, `stop`-via-continuation) while leaving non-hook execution policy independently composable.
+
+**Agent events** (`dsh-agent`):
+- `agent/session-start(agent, source)` — emit, once before turn 1, carrying a `SessionStartSource` (`startup` for a fresh/forked create, `resume` for a reloaded persisted session; `clear`/`compact` reserved). A pure notification — it CANNOT block startup (a deliberate gap: a bridge logs/injects, it does not gate startup). A listener seeds context via `agent.inject()`.
+- `agent/prompt-submit(agent, content, source, next) → PromptDecision` — waterfall, fired per drained queued message inside the open turn, before the `user/message` append. `allow` (optionally rewriting the prompt `content` or attaching separately sourced `additionalContexts[]`) or `block` (dropping the prompt; the loop appends a durable `prompt/blocked` in its place — see the dispatch note below).
+
+**`agent/turn-continuation`** receives and returns a `ContinuationDecision`. A `{action:'continue', reason?}` may carry model-facing content and source recorded as next-step steering in the same turn — the typed twin of the `/goal` step-end-steer pattern. It is not a `context/message`, so its type does not offer a context envelope or durable context metadata.
+
+### The tool pipeline gives each phase one kind of authority
+
+Every call follows `tools/pre-execute` → guards → `tools/execute` → dispatch → `tools/post-execute` → `tools/result`. The registry snapshots caller input, materializes and freezes arguments, and assigns an opaque token. Nested calls carry only the parent token. Identity remains immutable; only `signal` may change around dispatch. The log, UI, and tool body therefore agree on what ran.
+
+- **`tools/pre-execute`** is the extensible waterfall gate. Its `PreToolDecision` allows, denies, or asks. Deny skips `tools/execute` and core dispatch. Ask resolves through the optional approval seam: only `allowed-once` continues through guards and dispatch; rejection, cancellation, an unavailable channel, a missing approval service, or an agent-less call becomes a normalized denial. Every outcome still reaches post-policy and final observers.
+- **`ctx.tools.guard()`** installs synchronous scope-aware policy after the whole pre-execute waterfall. A guard may deny or abstain, never force-allow, so listener ordering cannot resurrect an operation that a final invariant forbids.
+- **`tools/execute`** is the around-dispatch waterfall for timeout, retry, and metrics plugins. A wrapper delegates to core dispatch with `next()`, may add, replace, or remove only `exec.signal` before doing so, and receives the already-normalized result of a thrown or unknown tool; returning its own valid result short-circuits dispatch.
+- **`tools/post-execute`** is the inspect/transform waterfall. Its `PostToolDecision` accepts, blocks with feedback, optionally replaces content, or attaches `additionalContexts`. The returned decision is the supported transform channel; after the waterfall, the registry materializes the complete outcome once before final observation.
+- **`tools/result`** is the synchronous contained notification after every transform, lossless-JSON materialization, and the outer error boundary. It receives the same frozen execution identity and an immutable snapshot of the authoritative result; observer failures are contained per listener and cannot change or reject `ToolRegistry.execute()`'s returned outcome.
+
+Core dispatch and the tool body sit inside normalization boundaries, so tool, listener, malformed-result, non-JSON result, and identity-shape failures resolve as JSON-safe `isError` results rather than escaping the turn. A post-execute listener can therefore inspect a thrown tool, and a final observer sees exactly what the caller receives and the session log can persist.
+
+**`TurnEndReason.rejected`** (`dsh-session`): a turn whose entire prompt batch was blocked by `prompt-submit`.
+
+### Three load-bearing loop decisions
+
+1. **Open the turn before prompt policy.** A fully blocked batch becomes a zero-step `rejected` turn, preserving enclosure and giving ACP a durable terminal event. Every veto also records `prompt/blocked` with the original prompt and reason, so mixed batches retain blocked inputs. Every allowed `additionalContexts` entry is injected into the open turn.
+
+2. **Post-tool `additionalContexts` and asynchronous injections enter the active-batch FIFO and append when that batch settles.** `content`/`feedback` shape the result `execute()` returns, but each context is a separate `context/message`, and a single step or composite tool can produce many. Appending context immediately would interleave `result(c1) → context → result(c2)` or place nested context before its outer result, breaking tool-call/result adjacency. `ToolRunContext.deferContext()` therefore collects nested-dispatch context through failures, `execute()` surfaces the ordered array on `ToolExecutionResult`, and the loop accepts it into the same FIFO as `agent.inject()` calls made during execution. The FIFO appends after every recorded result when the batch settles, including before an interrupted turn closes. An accepted outer call preserves deferred contexts before decision contexts; an outer block discards deferred contexts and exposes only contexts explicitly supplied by the blocking decision.
+
+3. **A forced `continue` `reason` is enqueued through the steering channel**, so the next step's top-of-loop drain records it as steering for the continued turn — next-*step* steering within the SAME turn, not a next-*turn* prompt (matching the existing `hasSteering` force-continue override).
+
+### Pre-tool input rewrite is a separate consistency decision
+
+`PreToolDecision` cannot rewrite arguments. History and the audit call are logged before execution, and ACP presentation reads the same input, so the registry seals arguments before policy. A valid rewrite must update history, audit, presentation, and execution before identity is created; that contract belongs to the [input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md).
+
+### Boundaries
+
+The seam package does **not** declare `hook/*` session events (the durable hook-invocation log); those belong to `dsh-hook-protocol`, because a native plugin uses typed decisions without an external hook log. The native-plugin integration test (`packages/core/agent-loop/tests/interception.spec.ts`) composes the seams through the real loop with no `hook/*` protocol. Compaction (`PreCompact`/`PostCompact`), Notification, and Codex `PermissionRequest` remain outside this decision. The [approval seam](2026-07-06-approval-seam.md) resolves `ask` decisions through `ctx.approval`, while terminal monotonic stopping is owned separately by `agent/turn-stop`.
+
+## Alternatives considered
+
+- **Shipping pre-tool INPUT rewrite as part of this seam set** — deferred as the over-reach signal; the section above carries the consistency problem (audit, history, and presentation all read `tool/call.arguments` logged before execution), and [the pre-tool input-rewrite proposal](../../proposed/feature/2026-06-30-pre-tool-input-rewrite.md) owns the design.
+- **Declaring the durable `hook/*` SessionEvents alongside the seams** — rejected: a native plugin uses the typed Decisions with no hook log at all (the worked example proves it), so the durable log belongs to [the hook-protocol library](2026-06-30-hook-protocol-lib.md), not the seam surface.
+
+## Consequences
+
+The canonical interception surface is uniformly typed without giving every extension the same power: hooks return decisions, execution wrappers wrap, terminal guards only deny, and final observers only observe. The loop owns session-start, prompt-submit, post-tool context buffering, and continuation; `dsh-tools` owns identity sealing and the five-phase execution pipeline. Their contracts are documented in [architecture.md](../../../../docs/architecture.md), package READMEs, [core interception decisions](../../../../docs/core-data-structures/core.md#interception-decisions), and [tool structures](../../../../docs/core-data-structures/tools.md). The ACP bridge maps `rejected` turns to its `cancelled` codec value, while hook-driven snapshots verify the observable bridge behavior end to end.

.agents/notes/implemented/feature/2026-06-30-session-store-fork-api.md

@@ -0,0 +1,41 @@
+# Agent Note: SessionStore fork API
+
+Status: implemented
+
+## Problem
+
+The event-sourced session log already has the primitive a fork needs: create a new session with a seed event prefix, then derive model history from that seeded log exactly as replay does. That primitive is intentionally low-level: `ctx.sessions.create(id, { seed, meta })` accepts any valid seed, but ordinary live-session branching needs policy around which prefix can be copied, which metadata is stamped on the child, and how errors are classified.
+
+The semantic hazard is the fork boundary. A valid user-visible fork seed must be contiguous and turn-enclosed. Forking inside an active turn would copy an open `turn/start`, possibly an open `step/start`, and possibly dangling tool calls. That violates turn-enclosure and provider-transcript invariants, and it creates a misleading child history that appears to have participated in an unfinished parent turn. The existing [subagent seam](2026-06-21-subagent-capability-seam.md) deliberately solves a different problem: tool-triggered subagent forks usually happen while the parent turn is open, so `dsh-subagent-fork` clips the seed to the parent's last completed-turn prefix. A general session fork should not silently clip; it should either fork the requested boundary or reject it.
+
+## Decision
+
+`dsh-session` owns ordinary live-session forking directly on `ctx.sessions`. There is no separate `dsh-session-fork` package or `ctx.sessionFork` service: the API has no independent backend, event vocabulary, lifecycle, or persistence behavior, and all durable work delegates to the existing session store and persistence backends.
+
+The store exposes one operation:
+
+```ts ignore-check
+type SessionForkSource = Session | SessionId
+
+class SessionStore extends Service {
+  fork(source: SessionForkSource, boundary?: number, childSessionId?: SessionId): Session
+}
+```
+
+`boundary` is the inclusive source event `seq` to copy through. When omitted, it defaults to the source session's current last event; on an empty source, omitted `boundary` creates an empty child. Fork-specific validation only checks that the requested boundary exists and is a `turn/end`. The selected prefix is then deep-cloned into the child seed. The child inherits the source session's `cwd`, stamps `parentSession` to the source id, and sets `seedLength` to the copied prefix length. When `childSessionId` is omitted, `SessionStore` generates one using its existing id policy.
+
+An empty prefix is forkable; any non-empty boundary must be a safe existing sequence at `turn/end`, regardless of reason. Typed errors distinguish missing sources, stale objects, duplicate child ids, and invalid boundaries. Broader log validation and crash repair remain with their existing owners.
+
+## Alternatives considered
+
+**Separate `ctx.sessionFork` service.** This was the first implementation, but review showed it overfit the capability-seam pattern. The code had no swappable backend, no extra event surface, no independent ownership lifecycle, and no durable behavior beyond `ctx.sessions.create({ seed, meta })`. Keeping a separate package would make callers discover and install a second service just to perform policy around a session-store primitive.
+
+**Two functions: `snapshot()` plus `fork()`.** This preserved a reusable seed/metadata computation, but the only supported consumer created a session immediately. It also made the surface feel more abstract than the concrete operation users need. A single `fork()` with an explicit `boundary` keeps the API direct while still supporting previous-point forks.
+
+**Silently clip open turns to the last completed boundary.** That is correct for `dsh-subagent-fork`, where delegation often starts while the parent turn is open and the child should inherit only the completed prefix. It is wrong for ordinary user/session branching because it hides that the requested fork point was not actually a valid boundary and silently drops the parent turn tail.
+
+## Consequences
+
+The public surface stays small and discoverable: live session branching is part of `ctx.sessions`, next to `create({ seed })`, rather than a standalone service or a two-step helper pair. Persistence continues to work through existing `session/created` and `session/flush` behavior: a forked child starts life with seeded events, so existing backends persist that seed once and preserve `parentSession` / `seedLength` in the header.
+
+The v1 scope still excludes ACP `session/fork`, unloaded persisted-session forking, model-facing tools, and subagent refactors. If a future ACP method is added, it should advertise the capability only after it has transcript/snapshot coverage; this Agent Note adds no editor-facing updates, so no ACP snapshot is required now. Fork-child replay remains covered by the existing [seed-boundary testing Agent Note](../testing/2026-06-22-fork-child-replay-seed-boundary.md), while this API gets focused `dsh-session` unit tests plus JSONL persistence coverage.

.agents/notes/implemented/feature/2026-06-30-subagent-observe-enrich.md

@@ -0,0 +1,29 @@
+# Agent Note: Subagent lifecycle enrichment — lastAssistantMessage (observe-only)
+
+Status: implemented
+
+## Problem
+
+The hooks subsystem ([interception seams Agent Note](2026-06-30-interception-seams.md)) lets a plugin observe and gate the agent at lifecycle points. Claude Code and Codex both expose **SubagentStart / SubagentStop** hooks, and CC's carry the subagent's final message. The harness already emits `subagent/start` and `subagent/end` lifecycle events ([the subagent capability-seam](2026-06-21-subagent-capability-seam.md)), but their payloads were minimal (`provider`, `id`, and on end `stopReason`) — not enough for a hooks bridge to report WHAT a subagent produced without separately reaching for the live run.
+
+This Agent Note enriches the end payload. It is deliberately **observe-only**: no control-flow change and no waterfall. A run-affecting subagent-stop decision (continuation, injection that changes the run) is a separate, larger redesign and stays out of scope.
+
+## Decision
+
+**Add `lastAssistantMessage` — the child's final output — to `SubagentRunEndInfo`.** On the settle path it is the readonly typed `SubagentResult.output`, so an observer sees what the child produced without holding the run. On an infrastructure rejection where no `SubagentResult` exists, it is absent and the event reports `stopReason: 'error'`. Providers and listeners are trusted same-process collaborators and honor the borrowed immutable payload contract.
+
+Both events stay plain **`emit`s**. Async `SubagentService.start()` attaches result observation to the ready provider run, emits `subagent/start`, and then returns the run; an in-process listener can therefore reach the published child via `ctx.agents.get(info.id)`, while a remote provider need not have a local registry entry. A rejected provider start emits neither event. The callbacks remain observe-only and per-listener containment keeps one bad subscriber from stranding a live run or starving later listeners.
+
+## Alternatives considered
+
+**An `agentType` subagent-kind label** (the harness analogue of CC's `subagent_type`) on the request + both lifecycle payloads — an earlier draft shipped it; dropped in review because it is a Claude-Code concept that does not fit our own seam (nothing here interprets it, and the only consumer was a CC-dialect bridge). The CC bridge instead feeds Claude Code's own default matcher value `"general-purpose"` for its SubagentStart/Stop `agent_type` matcher, so this Agent Note ships ONE enrichment: `lastAssistantMessage`.
+
+**A control-flow `subagent/end`** — deferred; see below.
+
+## Why observe-only, and what is deferred
+
+A control-flow `subagent/end` (an awaited waterfall returning a stop/continue decision, like the other interception seams) would require: reshaping `subagent/end` from emit to waterfall, restructuring `SubagentService.start` to await listeners before settling, and implementing the `resume` capability in the in-process provider so a "continue" can actually re-run the child. That belongs to the background/steering subagent redesign the [capability-seam Agent Note](2026-06-21-subagent-capability-seam.md) already defers (the same redesign that unifies long-running-tool handling across subagents and bash). This Agent Note ships the observe-only enrichment a hooks bridge needs today; `FIXME(subagent-continuation)` / `TODO` anchors mark where the control-flow version would land if and when that redesign happens.
+
+## Consequences
+
+A hooks bridge (or a native plugin) can now forward the child's `lastAssistantMessage` to a SubagentStop handler by subscribing to the existing emits — no new control-flow surface. The vocabulary addition is documented in [docs/core-data-structures/subagent.md](../../../../docs/core-data-structures/subagent.md) (the events prose) and the two subagent READMEs; the catalog is regenerated. No production behavior changes — the events fire exactly as before, with one more (optional) field on the end payload — so no snapshot or e2e change is needed.

.agents/notes/implemented/feature/2026-07-05-dynamic-workflows.md

@@ -0,0 +1,78 @@
+# Agent Note: Dynamic workflows — a script-driven multi-agent orchestration seam
+
+Status: implemented
+
+## Problem
+
+The harness can delegate ONE task to ONE child (`dsh-tool-subagent`), but work that fans out across many independent pieces — an audit over many files, a migration, multi-angle research, adversarial verification of findings — forces the model to orchestrate turn by turn: every intermediate result lands in the parent context, the plan lives nowhere durable, and coordination costs a model round-trip per step. Claude Code ships this capability as [dynamic workflows](https://code.claude.com/docs/en/workflows): the model writes a JavaScript orchestration script, a runtime executes it, and the script — not the conversation — holds the loop, the branching, and the intermediate results.
+
+## Decision
+
+A workflow capability family at `packages/workflow/` in the bash seam shape (interface / implementation / consumer), plus the structured-output foundation it needs on the subagent seam.
+
+### The script contract (Claude Code-compatible)
+
+A workflow call contains JSON `meta` (`name`, `description`, and optional `whenToUse`/`phases`) and a JavaScript `script` body with top-level `await` that returns a JSON value. Metadata is validated as data and never evaluated. The body receives `agent(prompt, options)`, `parallel(thunks)`, `pipeline(items, ...stages)`, `phase(title)`, `log(message)`, and `args`. Pipeline stages receive `(prev, item, index)` with no cross-stage barrier; failed children and ordinary stage errors resolve the affected item to `null` and skip its remaining stages. Claude Code's determinism restrictions are deferred with journaling, so compatible bodies may use clock and randomness after moving their meta header into the parameter.
+
+One deliberate strictness DIVERGENCE from CC: hook misuse — unknown or deferred options (`effort`/`isolation`/`agentType`), malformed arguments, schemas outside the supported subset, tripped caps, seam start failures — throws a `WorkflowError` with `fatal: true`, and the combinators RE-THROW fatal errors instead of nulling the item. Without this, a typo'd option dissolves into a `null` indistinguishable from a child failure — the accepted-then-ignored failure mode this repo bans. One addition: the tool's `args` parameter is a JSON OBJECT (a bare list is wrapped as a field) so the wire schema stays honest.
+
+### The seam (dsh-workflow)
+
+`ctx.workflows` is an abstract `WorkflowService` in the bash shape — one engine per context, no named-provider registry (engines are deployment swaps, not co-residents). `start(request)` throws synchronously for a script that cannot begin; a returned `WorkflowRun`'s `result` NEVER rejects (failures resolve as `stopReason: 'error' | 'cancelled'`). The `workflow/*` events are observe-only emits carrying DATA SNAPSHOTS (id + meta; `workflow/end` omits the result value), per-listener contained, mirroring `subagent/start`/`subagent/end` — control stays with the run's holder. Vocabulary details: [core-data-structures/workflow.md](../../../../docs/core-data-structures/workflow.md).
+
+### The engine (dsh-workflow-workerthread): one worker thread per run
+
+**Trust premise**: workflow scripts have the same trust as the model's bash access. The engine contains buggy scripts and guarantees settled results, JSON-safe values, and cancellation quiescence; it does not defend against hostile code. A vm context and worker thread are not security boundaries: a script can escape to Node APIs with process-wide authority. Sandboxing requires a separate-process or isolated-vm engine behind this seam.
+
+**Why `node:worker_threads`**: each run gets one unpooled worker. A vm context limits the documented script surface, while message-port RPC bridges `agent()` to host-side child loops. The worker prevents synchronous script work from blocking the host, provides a serialization boundary, and permits forced termination after cancellation. `isolated-vm` was rejected because of its maintenance state and deployment requirements.
+
+The host validates metadata and parses the body before publication. Private enum-keyed payload maps define the wire protocol; pending starts, published child records, one cancellation signal, worker-death reaping, result precedence, and disposal quiescence preserve the subagent run contract across it. The [agent-scope runtime-design Agent Note](../architecture/2026-07-12-agent-scope-runtime-design.md#workflow-children-are-pending-starts-or-published-records) owns those race algorithms.
+
+The engine exposes an in-process `MessageChannel` test path because main-process V8 coverage cannot see worker execution.
+
+**Meta is data**: the schema-validated `meta` field reaches the seam as JSON and is only shape-validated. The host never evaluates a metadata literal, which would let script-controlled accessors run outside the worker's isolation.
+
+**Value boundary**: `materializeFromRealm` copies outbound values and rejects functions, symbols, nested `undefined`, exotic prototypes, cycles, sparse arrays, and non-finite numbers. Data-property copies make `"__proto__"` safe; getters are read normally and a throwing getter fails loudly. `args` crosses through `workerData` and is cloned again before exposure. Realm functions are invoked rather than copied, and thrown values use a total renderer so `result` cannot reject. Hook errors are host-realm `WorkflowError`s, so scripts branch on `name` or `code` rather than `instanceof Error`, as documented in the engine README. Concurrency, total-agent, item, timeout, and grace limits are validated config.
+
+### The consumer (dsh-tool-workflow)
+
+A `workflow` tool mirroring `dsh-tool-subagent`'s synchronous shape: start, await, `try/finally` dispose, abort-bridge `exec.signal`, non-`completed` → `isError`. Render intent: a `generic` card titled by the call's `meta.name` parameter (presentation is a pure function of args). The tool description IS the model-facing authoring spec. The usage policy ships with the tool as its own `tool:<toolName>` prompt section (explicit-ask-only guidance — tool guidance lives in tool plugins, never in the deployment persona); the harness has no ultracode-style effort gate.
+
+### The foundation: structured output on the subagent seam
+
+`SubagentStartRequest.outputSchema` is implemented by `dsh-subagent-inprocess` for both in-process backends. Each structured child receives its own scoped capture tool, instruction, and enforcement registrations on `child.ctx`; concurrent children can use different schemas without sharing mutable policy, and disposing the child removes the entire attachment.
+
+An output schema makes a schema-valid committed capture mandatory for successful child completion. The scoped runtime presents the capture tool and instruction, commits only a successful final outcome—including the enclosing `run_code` outcome for an SDK call—denies later side effects after capture becomes pending, and stops the child without another model step after commit. A validation failure remains a retryable tool error; clean completion without a committed capture settles as an error.
+
+`StructuredOutputSchema` is the raw enforceable JSON-Schema subset in `dsh-tools` (single-string `type`, `properties`/`required`/`additionalProperties`, `items`, scalar `enum`/`const`), and unsupported keywords fail loudly because that wire data becomes the capture tool's parameters verbatim. The [agent-scope runtime-design Agent Note](../architecture/2026-07-12-agent-scope-runtime-design.md#structured-output-commits-only-authoritative-outcomes) owns the assembly, commit, guard, and terminal-stop correctness algorithms.
+
+## Testing
+
+Worker-side logic runs through an in-process `MessageChannel` so V8 coverage measures it. Unit tests cover script helpers, fatal and nullable failures, JSON boundaries, caps, cancellation, child ownership, and structured output through real loops. A built-bin smoke runs the separately bundled `lib/worker.cjs` under plain Node, a with-key e2e drives real child agents, and model-facing workflow behavior is snapshot-covered through its owning example.
+
+## Deferred (documented non-goals of this cut)
+
+- **Background collection** (start tool → run id → completion notice → collect), designed alongside bash/subagent background unification.
+- **Journaling + resume** (`resumeFromRunId`, cached agent() prefixes) — implementing it reintroduces CC's determinism bans as a script-contract tightening (scripts may read the clock today).
+- **Saved/bundled workflows** (a `.deepseek/workflows/` registry, slash-command surface) and **script persistence to a run directory** (the tool-call event already records the script durably).
+- **Nested `workflow()`**, **token `budget`**, and the `effort`/`isolation`/`agentType` agent options (each rejects loud with a message naming it deferred).
+- **An overall run wall-clock timeout** — cancellation always frees the caller (result settles within the grace), so a cap on total run time is a policy knob for the background redesign, not a correctness need here.
+- **Engine hardening beyond worker threads**: an isolated-vm or separate-process engine behind the same seam (actual sandboxing; memory limits).
+- **ACP progress UI** over the `workflow/*` events (a `/workflows`-style view); the events exist for it.
+- **ACP-backend structured output** and **`toolFilter`** (both still capability-gated `false`).
+
+## Alternatives considered
+
+- **Hostile-value containment in the host** (trap-free proxy rejection, accessor-never-invoked descriptor walks, realm-side pre-rendering of thrown values, realm-built promises/arrays/error clones with structural fatal recognition): rejected because every defense targets an author the trust premise accepts, while the thread's serialization boundary already makes cross-realm values total by construction.
+- **In-process `node:vm` execution**: mechanically simplest — no RPC, no thread — but `start()` blocks the caller for the script's initial synchronous slice, a synchronous spin past the first await cannot be killed in-process (the vm `timeout` covers only that first slice), and `dispose()` could only abandon an unsettling script on the host loop. The worker-thread engine keeps the same vm-context script surface while unblocking the host and making termination real.
+- **Background execution as the default** (CC's shape): deferred; foreground-synchronous matches `dsh-tool-subagent`'s cut, and background semantics should be designed ONCE across bash/subagent/workflow rather than per-tool.
+- **Workflow-layer JSON parsing for `agent({schema})`**: duplicating a seam concern at one consumer while the seam's capability flag stayed dishonestly `false`.
+- **Meta embedded in the script as `export const meta = {...}`** (CC's exact format): keeps scripts self-contained and CC scripts drop-in, but obtaining meta requires evaluating model-written text on the host. Even an empty timed vm context cannot bound script-controlled getters when the host reads the resulting object. A JSON parameter removes the scanner, evaluation, and host-spin hole; the cost is that a CC script's meta header must move into the parameter (the body stays drop-in).
+- **`SchemaSpec` as the outputSchema type**: the author-facing DSL cannot express what arrives as data and cannot be validated against without conversion loss.
+- **A schema-object library (zod, or the repo's schemastery) for the structured-output subset**: the schema is wire data — plain JSON that crosses the vm realm boundary in `agent({schema})` and lands verbatim in the forced tool's parameters — exactly where live schema objects cannot sit; consuming raw JSON Schema at runtime would need a third-party converter on top (zod core only emits JSON Schema, not the reverse), and it would put a second schema language beside schemastery's config role.
+- **ajv for value validation**: it validates FULL JSON Schema, so the subset gate — the module's actual point, since every accepted keyword must be one the harness enforces — would remain hand-written regardless; it compiles validators through `new Function`; and it would be dsh-tools' first runtime dependency, all to replace the ~70-line value walker while the path-qualified, every-violation error reporting stays custom either way.
+- **Provider JSON mode instead of the capture tool:** it guarantees valid JSON, not schema conformance, and its interaction with tool calling is unclear. The capture tool preserves in-turn validation retries. Provider-side strict tool schemas can later narrow the accepted subset without changing this design.
+
+## Consequences
+
+Fan-out plans now live in rerunnable scripts, and `outputSchema` provides authoritative structured child results. Each run pays worker startup and message-port RPC costs, but host startup stays non-blocking, cancellation can terminate the worker, and serialization enforces the value boundary. Worker threads are not a security boundary. Invalid options fail rather than degrading to Claude Code's `null`; consumers retain control through the run handle while observers receive snapshots only.

.agents/notes/implemented/feature/2026-07-05-skill-system.md

@@ -0,0 +1,53 @@
+# Agent Note: Skill system — progressive disclosure instructions for agents
+
+Status: implemented
+
+## Problem
+
+Agent products have converged on a skill pattern: keep the request prompt small by listing only available instruction bundles, then load the full body when the model decides a task matches. Codex, Claude Code, OpenCode, and Kimi Code differ in details, but all separate discovery metadata from complete instructions so a workspace can carry reusable behavior without paying the full prompt cost on every turn.
+
+DeepSeek Harness uses the same primitive so project-specific review, plugin-authoring, and tool-usage guidance lives next to the workspace or the user's agent configuration instead of being hard-coded into the loop.
+
+## Decision
+
+`@deepseek-ai/dsh-skill` is the pure provider registry (`ctx.skills`), `@deepseek-ai/dsh-skill-local` is the shipped local filesystem provider, and `@deepseek-ai/dsh-tool-skill` owns the session-prefix catalog and model-facing loader tool. `dsh-agent-spine-demo` loads the registry, local provider, and consumer by default so stdio and ACP apps get the same behavior while embedded or remote providers contribute skills without changing the registry or consumer. Its `skills` config forwards `registry`, `local`, and `tool` branches to those owners.
+
+Provider plugins register synchronously during `apply()`. Provider membership is direct effect-owned state: registration and disposal invalidate completed catalogs synchronously, and discovery reads the current provider map on demand rather than observing registry-change events. Provider catalogs return ranked candidates from awaited `list()` calls, where remote providers perform initialization, authentication, and discovery while honoring the lookup abort signal. The registry validates each candidate, resolves same-name skills first-wins by rank, provider registration order, and provider-local order, then sorts summaries by skill name for deterministic consumers. It caches only completed catalog snapshots and retries when a provider/runtime revision changes during discovery, so an unload cannot freeze a stale, unresolvable skill into a session prefix. Runtime `ctx.skills.register(...)` remains a convenience for embedded in-process skills and uses project-over-user priority; `runtime` is reserved as the registry-owned provider name.
+
+The local provider scans cwd-sensitive project roots, custom roots, and user roots in first-wins rank order: project `.dsh`, project `.agents`, `customSkillDirs`, user `.dsh`, then user `.agents`. The user `.dsh/skills` scan skips `.system` so a system-owned directory is not treated as normal user content. DeepSeek Harness does not ship built-in system skills; embedded or remote providers supply additional skills when configured.
+
+Each skill is either `<name>/SKILL.md` or `<name>.md` with YAML frontmatter. `name` and `description` are required; `whenToUse`, `disableModelInvocation`, and `metadata` are optional. Names are kebab-case. YAML frontmatter is parsed with the `yaml` package instead of `js-yaml` or a hand-written parser: `yaml` is the already-declared modern parser for this package's limited frontmatter needs, and a narrow parser would either reject valid YAML users expect to work or grow into an unreviewed YAML subset.
+
+Local skill filesystem I/O goes through `ctx.fs` when a filesystem service is loaded: project-root lookup probes `.git` with `resolve` and `stat`, root discovery uses `listDir`, and skill reads use `readText`. The Node filesystem remains a fallback for minimal contexts that mount `dsh-skill-local` without the fs seam. Missing roots, unreadable or malformed skill files, and transient provider `list()` failures degrade to warn-and-skip so one bad source does not make every agent request fail; malformed candidates still fail fast because they are provider contract violations.
+
+`dsh-tool-skill` contributes one user-role `<system-reminder>` catalog through [`agent/session-prefix`](2026-07-07-session-prefix.md). The catalog contains sorted skill name and description only; it excludes bodies, paths, sources, providers, and routing hints. Descriptions are whitespace-normalized, XML-escaped, and capped by `catalogDescriptionMaxLength`, whose default is `500` and minimum is `3`. The session-prefix seam freezes the request-only catalog per loop instance and records it in the request header, preserving reconstructability without adding it to durable history. Full skill bodies are never included in the catalog.
+
+The `skill({ name })` tool loads one full skill for the current agent cwd and returns a tool result containing `<skill_content name="...">`, `<skill_resources>`, and `<skill_instructions>`. `resourceBase` supplies a directory, URL, or opaque provider-managed base for explicitly referenced scripts, references, and assets; resources load only as needed, without directory enumeration. An unresolved name reports that the skill is unknown or no longer available; invalid names and skills marked `disableModelInvocation` retain distinct tool errors. The tool result is the model-visible disclosure path.
+
+The data structures and catalog/tool contract are documented in [skills.md](../../../../docs/core-data-structures/skills.md), with service signatures in the generated [services catalog](../../../../docs/cordis-catalog/services.md).
+
+## Alternatives considered
+
+**Inject full skill bodies into every system prompt.** Rejected because it destroys progressive disclosure and makes every request pay for instructions that may not apply.
+
+**Expose skills only as slash commands.** Rejected because model-initiated loading is the core capability; slash/ACP command advertisement does not change discovery.
+
+**Put local filesystem scanning directly inside `ctx.skills`.** Rejected because coding agents, web agents, and future plugin ecosystems need different skill sources. A provider registry mirrors the subagent seam: the registry owns conflict resolution and consumers, while implementations own loading.
+
+**Use a system-prompt section.** Rejected because the rendered system prompt is a single string, while the catalog is a user-role `<system-reminder>` message with request-only lifecycle requirements. [`agent/session-prefix`](2026-07-07-session-prefix.md) is the selected mechanism: it places the catalog ahead of derived history and records the composed message in the request header.
+
+**Materialize built-in DSH authoring skills under `~/.dsh/skills/.system`.** Rejected because bundled skills do not write user home on startup, and embedded or remote providers supply configured skills.
+
+**Recursively discover nested `**/SKILL.md`.** Rejected. Flat files and one-level directory bundles cover the configured roots while keeping duplicate handling and catalog order easy to reason about.
+
+**Hand-parse frontmatter.** Rejected because the accepted schema includes an open `metadata` object. A narrow parser would either reject valid YAML users expect to work or grow into an unreviewed YAML subset.
+
+## Consequences
+
+The agent-core spine includes one session-prefix contributor, one local provider, and one model-facing tool. Skill discovery is cwd-sensitive, so callers that create agents with different session cwd values can observe different project skill overrides by design.
+
+The catalog is deterministic for a fixed root set and runtime registration revision, but disk changes are not watched; discovery is memoized until runtime registration invalidates the cache or the process restarts.
+
+## Deferred
+
+Forked skill contexts (`context: fork`), direct user/slash invocation (`user-invocable`), parameter declarations and hints (`arguments` and `argument-hint`), and per-skill tool constraints (`allowed-tools` and `disallowed-tools`) are outside the shipped contract. The registry, local provider, and model-facing tool do not parse, advertise, or enforce these fields.

.agents/notes/implemented/feature/2026-07-06-approval-seam.md

@@ -0,0 +1,137 @@
+# Agent Note: The approval seam — one-shot permission decisions over a waterfall of answerers
+
+Status: implemented
+
+## Problem
+
+Two callers need to put one question — "may this specific action proceed?" — to a human: `tools/pre-execute`'s `ask` decision (including the Claude-Code hook bridge's `permissionDecision: ask`) and the [sandbox Agent Note](2026-07-06-sandbox.md)'s post-denial one-shot escalation retry. A shared seam keeps them from inventing separate outcome vocabularies, UI routing, cancellation, and audit trails, while guaranteeing that a deployment with no UI can never grant an unanswerable request.
+
+The routing problem is ownership: an approval prompt must reach the editor session that owns the asking agent (the ACP bridge multiplexes N sessions over one connection), fail closed for agents nobody owns (in-process subagents, tests), and stay out of deployments that compose no UI (headless, CI).
+
+## Decision
+
+One package, `dsh-user-approval` (`packages/ui/user-approval`), owning the vocabulary and the `ctx.approval` service — the MECHANISM. The POLICY — who answers, and whether a session is asked at all — lives outside it: answerers are `approval/request` waterfall listeners registered by the plugins that own the channel (the ACP bridge; future terminal UIs; test scripts), and a per-session policy tier can decide before any human is involved. Consumers (`dsh-tools`' ask routing, the sandbox escalation gate) resolve a question to a closed outcome and derive their own tool results from it. Deliberately ONE package, not the capability-seam three (see Alternatives).
+
+### How a deployment uses it
+
+One `cordis.yml` entry mounts the seam. Not loading it is the fail-closed opt-out: consumers deny unanswerable requests with zero approval code registered.
+
+```yaml
+- id: approval
+  name: '@deepseek-ai/dsh-user-approval'
+  # config:
+  #   policy: never   # deployment default for sessions without an override; 'ask' when omitted
+```
+
+The entry alone provides mechanism, not a channel: with no answerer composed, every ask resolves `unavailable` and the asking tool call denies — fail-closed needs no configuration. Composing the ACP app (`@deepseek-ai/dsh-acp-demo`, as in [the acp-agent example's default tree](../../../../examples/acp-agent/README.md)) completes the loop: its bridge registers an answerer that prompts the owning editor session via `session/request_permission`, so a hook's `ask` or an escalation request surfaces as a one-shot Allow/Reject prompt attached to the already-streamed tool call. `policy: never` is the unattended stance — every ask auto-rejects deterministically, stated in the system prompt, no human in the loop. `policy` is validated against the closed list at plugin load; anything else throws.
+
+What a composed deployment observes: `allowed-once` lets exactly that call proceed; rejection, dismissal, and channel absence deny with three distinct reasons the model can tell apart; a successful in-turn request lands a durable `approval/asked`/`approval/decided` pair on the asking agent's session log; nothing about a grant persists past the call that asked. An idle request or audit append failure rejects instead of returning an unaudited decision.
+
+One ask under this composition, verbatim from the sandbox example's recorded `escalation-approved` scenario — the model requests a sandbox escalation, the gate asks, the bridge prompts the owning editor, the user clicks Allow once:
+
+```
+tool/call        bash {"command": "printf 'escalated\n' > escalated.txt && cat escalated.txt",
+                       "sandbox_permissions": "workspace-write",
+                       "justification": "the user asked to write escalated.txt in the workspace"}
+approval/asked   {"toolName": "bash", "callId": "call_00_…",
+                  "reason": "escalate sandbox to workspace-write: the user asked to write escalated.txt in the workspace"}
+  → session/request_permission {"toolCall": {"toolCallId": "call_00_…"},
+                  "options": [{"optionId": "allow-once", "name": "Allow once", "kind": "allow_once"},
+                              {"optionId": "reject-once", "name": "Reject",     "kind": "reject_once"}]}
+  ← the user picks "Allow once" on the prompt the editor attaches to the streamed bash call
+approval/decided {"outcome": "allowed-once"}
+tool/result      "escalated" — this one call ran under the wider mode; the grant died with it
+```
+
+The `escalation-rejected` twin ends in `{"outcome": "rejected"}` instead: nothing executes, and the model's result carries the asker's verbatim fail-closed text (`the user rejected escalating this command to "workspace-write"`). A hook's `permissionDecision: ask` rides the identical wire; only the asker and its deny texts differ (§ Ask routing in dsh-tools). Headless, the same request skips the prompt entirely and settles `unavailable`.
+
+### Design detail
+
+#### The seam: mechanism and policy split
+
+After validation and a successful `approval/asked` append, the service resolves the `approval/request` waterfall to `allowed-once`, `rejected`, `cancelled`, or `unavailable`. It borrows the readonly request identity and signal, treats abort as `cancelled`, contains answerer failures and invalid returns as `unavailable`, discards late answers, and appends the paired `approval/decided` event. Pre-commit audit failures reject; post-append observer failures cannot undo an authoritative event. `allowed-once` authorizes only the asked action, and `request()` rejects outside an open turn so the audit pair remains inside the durable commit boundary.
+
+Answerers are `approval/request` waterfall listeners. Zero listeners fall through to `unavailable`; a recognizing listener occupies the first-wins decision slot, while an unrecognized agent must delegate with `next()`. Listeners dispose with their fibers, so an unloaded channel fails closed. Because sibling registration order is not deterministic, a deployment composes one terminal answerer and reserves `prepend` for decide-or-delegate gates.
+
+`ApprovalRequest` carries the asking `agent`, `toolName`, optional exact `callId`, human-readable `reason`, and optional `signal`. It uses the `CallId` brand without importing `dsh-tools`, which depends on this seam. Tool arguments stay on the already-streamed call that a UI references by `callId`.
+
+#### Ask routing in dsh-tools
+
+`ToolRegistry.execute()` resolves `ask` before dispatch: `allowed-once` proceeds, while rejection, cancellation, and channel absence produce distinct deny reasons. Opportunistic `ctx.get('approval')` consumption lets an absent or unmounted service fail closed without gating the registry fiber. Agent-less execution also fails closed because it has neither an audit session nor a UI owner.
+
+#### The per-session policy tier
+
+The seam also owns the session-scoped `'ask' | 'never'` policy described by [the sandbox Agent Note](2026-07-06-sandbox.md). Effective policy is folded from logged switches over the deployment default. `'never'` resolves to `rejected` inside `request()` before any answerer can run; `'ask'` dispatches and otherwise falls through to `unavailable`. The prompt states only deterministic `'never'`, switch narration is coalesced, and every request still records the audit pair.
+
+#### The ACP answerer
+
+The ACP bridge answers only for an exact agent object owned by its forward session map. It attaches `session/request_permission` to the existing `callId`, advertises one-shot allow/reject options, maps cancellation separately, and never grants an unknown option. Foreign or call-less requests delegate; a failed client RPC becomes `unavailable`. Hooks and `tools/pre-execute` decide whether a call asks at all.
+
+The answerer routes through the bridge's exact-agent ownership check described by [the ACP support Agent Note](2026-06-14-acp-agent-client-protocol.md), implementing the per-session permission ownership required by [the multi-session Agent Note](2026-06-14-acp-multi-session.md).
+
+#### Audit, and what the model sees
+
+`approval/asked` and `approval/decided` are durable log-only events; the model sees only the ordinary tool result derived from the outcome. Successful completion commits one `decided` per `asked`, including cancellation and contained answerer failure. Idle requests append neither event; a pre-commit failure rejects, while failure of the second append can leave an already-committed `asked` unmatched.
+
+#### Entities and dependencies
+
+`dsh-user-approval` depends on Cordis plus the session, agent, and branded-call contracts; `dsh-tools` and `dsh-acp` consume it. The sandbox executor stays independent because `dsh-tool-bash` owns escalation requests. The fixed dispatch-and-audit service remains one package; replaceable answerers live with their channel owners. Static capability grants and `subagent-acp` child-side permission answers remain separate concerns.
+
+### Testing
+
+Unit tests pin outcomes, first-wins delegation, containment, cancellation, scoped routing, audit pairing, the unbypassable `'never'` policy, tool deny reasons, and ACP ownership/outcome mapping through a real scripted bridge.
+
+Snapshots record allowed and rejected sandbox escalation through `session/request_permission`, plus the `'never'` prompt and policy-switch notice. Unscripted permission prompts cancel and fail closed.
+
+## Deferred
+
+- **`allow_always` grant storage** — honoring a persistent grant means designing storage, scope identity (call? path? prefix? session? time window?), and revocation; until designed, only the one-shot options are advertised ([the sandbox Agent Note](2026-07-06-sandbox.md) § Escalation records the open scope question).
+- **A recorded hook-driven `ask` through a composed answerer** — the human-prompt wire is recorded through the sandbox example's escalation branches. The hook matrix's `hook-cc-pretool-ask` pins the no-ApprovalService fallback denial, while the hook-producer-plus-answerer composition remains on the unit tier.
+- **Routing a child agent's approvals to the parent session** — `subagent-acp`'s child auto-answers its own `permission` requests; surfacing them to the parent's editor is its own design.
+
+## Alternatives considered
+
+- **A single registered provider instead of waterfall listeners** — rejected: a `registerProvider()` surface forces every composition question — allowlist pre-filters, external hook deciders, scripted test answers, a policy gate in front of a human — inside one provider implementation. The waterfall gets composition, fail-closed absence, and HMR disposal from machinery the runtime already has; the seam's JSDoc pins the single-decision-slot convention instead of inventing a provider registry.
+- **An inline `tools/pre-execute` permission gate in the ACP bridge** — rejected: prompting for every bridge-owned call hardwires the asking POLICY into the UI plugin, cannot serve a second asker (sandbox escalation happens after execution starts, with no pre-execute moment), and leaves hook-produced `ask` decisions without a shared mechanism.
+- **The generic user-interaction seam (`ctx.userInteraction`)** — rejected as the approval mechanism: the two share a skeleton (route by agent, block for a human, handle absence), but approval's contract is narrower in every dimension that matters: a closed outcome vocabulary instead of free text, a protocol-native prompt attached to a tool call instead of a generic form, mandatory fail-closed absence, and audit events. Approval therefore does not ride the shipped `packages/ui/user-interaction` / `ask_user_question` elicitation path — an elicitation form is not a permission prompt, and a free-text answer is not a closed outcome; sharing provider plumbing stays open if the two ever converge.
+- **Static optional injection in `dsh-tools`** — rejected: the vendored cordis `Inject` type has no optional flag — the object form maps service names to intercept config, and a declared inject gates the fiber. `ctx.get('approval')` is the documented opportunistic-consumption pattern (the `tool-bash` owner-token lookup, the loop's persistence probe), reads presence per call, and degrades correctly across HMR without extra machinery.
+- **The capability-seam three-package split** — rejected: interface/implementation/consumer fits a seam whose implementation is swappable (bash-local vs bash-sandbox). Here the service body is fixed mechanism and the variable part is listeners that live with their owners — splitting would manufacture an implementation package with nothing in it ("don't split preemptively").
+- **Offering `allow_always` now** — rejected: the protocol can express it, but honoring it means designing grant storage, scope identity, and revocation (§ Deferred). Advertising an option the harness cannot honor manufactures doomed grants.
+
+## Consequences
+
+The implemented contract is pinned by the suites in Testing:
+
+- `allowed-once` dispatches one action; every other outcome denies with a distinct reason, and `'never'` rejects before prompting.
+- Missing, foreign, agent-less, throwing, invalid, and disconnected answer paths fail closed.
+- Successful requests route by exact agent ownership and append one replayable, model-invisible audit pair; idle and pre-commit failures reject.
+- ACP ownership keeps prompts inside their session, while a deployment without the service emits no prompt or audit events.
+
+Costs and accepted limits:
+
+- **Two decide-eager answerers race for the slot.** Sibling-plugin listener order is not deterministic, so the seam cannot referee competing terminal answerers — mitigated by convention (one terminal answerer per deployment; `prepend` only for decide-or-delegate gates) rather than a priority mechanism the event bus does not have.
+- **Production exercise rests on one composition.** `ask` has two producer families — the hook bridges through `tools/pre-execute`, and sandbox escalation through its own gate — with the wire recorded in the sandbox example's snapshot suite, so the seam's real-world coverage is that one composition until more deployments compose it.
+- **Ownership keys on `Agent` object identity.** The answerer resolves the forward session-map record at `agent.session.id`, then requires that record to own the exact agent object; every current path hands the same object through the loop and the seams, but a future boundary that clones or proxies agents would make the bridge delegate and fail closed — safe, but silently UI-less — and would need a different ownership contract.
+
+## FAQ
+
+- **What happens in a deployment with no answerer at all (headless, CI)?** Every ask falls through the empty waterfall to `unavailable` and the tool call denies with the "no approval channel is available" reason. Fail-closed is the zero-listener default, not a configuration.
+- **Can a grant persist — "always allow this"?** No. `allowed-once` authorizes the single asked-about action and the service stores nothing between requests; `allow_always` is deliberately not advertised until grant storage is designed (§ Deferred).
+- **What does the model see of an approval?** Only the tool result the asker derives from the outcome — the audit pair never enters the transcript. The three non-grant reasons are distinct, so the model can tell a human "no" from a dismissed prompt from a missing channel.
+- **Who decides whether a call asks in the first place?** Policy producers: a hook returning `permissionDecision: ask`, any `tools/pre-execute` listener, or the sandbox escalation gate. The seam and the bridge only route and answer; neither injects its own judgment about what deserves a prompt.
+- **What happens when the user dismisses the prompt, or the turn aborts mid-ask?** Dismissal maps to `cancelled` with its own deny text. An already-aborted signal settles `cancelled` without dispatching; an abort during the ask discards the late answer. When both audit appends commit, either path records one pair, never two.
+- **What if the client answers with an option the harness never offered?** Any selection other than the offered `allow_once` maps to `rejected` — an unknown optionId from a non-conforming client can never grant.
+- **How do subagents' approvals route?** An agent no answerer owns delegates through the whole waterfall and fails closed — in-process subagents are deliberately unanswerable. `subagent-acp`'s child-side auto-answer is separate; routing a child's asks to the parent's editor is deferred (§ Deferred).
+- **What does `policy: 'never'` actually change at runtime?** The service resolves every ask for that session to `rejected` before dispatching any answerer (in-service, so no registration order can bypass it); the system prompt states the policy; switches are narrated at boundaries; each successful auto-rejection records the audit pair.
+- **What happens across a hot reload, or when the UI plugin unloads mid-session?** Answerers dispose with their owning fiber, so the next ask degrades to `unavailable` instead of hanging on a dead channel; remounting re-registers the answerer with no catch-up state.
+- **Where does the user see what they are approving?** On the tool call itself: the prompt attaches to the already-streamed call via `callId` — arguments included — and adds the asker's human-readable `reason`; the request carries no argument copy of its own.
+
+## Prior art
+
+In-repo precedents this design copies or contrasts with:
+
+- The `fs/write-intent` gate (`packages/fs/fs/`) — the documented single-occupancy decision-slot waterfall semantics (first answer wins, delegate via `next()`) the answerer contract reuses.
+- `hook/invoked`/`hook/result` — the log-only audit-pair precedent `approval/asked`/`approval/decided` follows; [the hook-bridges Agent Note](2026-06-30-hook-bridges.md) ships `permissionDecision: ask`, the first producer.
+- [The interception-seams Agent Note](2026-06-30-interception-seams.md) — the `tools/pre-execute` `allow`/`deny`/`ask` vocabulary whose `ask` this seam services.
+- [The ACP support Agent Note](2026-06-14-acp-agent-client-protocol.md) — the exact-agent ownership check against the forward session map that the answerer routes through; [the multi-session Agent Note](2026-06-14-acp-multi-session.md) — the per-session permission-ownership blocker this implements.
+- The opportunistic `ctx.get()` consumption pattern (`tool-bash`'s owner-token lookup, the loop's persistence probe) — how `dsh-tools` consumes the seam without gating its fiber on it.

.agents/notes/implemented/feature/2026-07-06-explicit-tool-order.md

@@ -0,0 +1,49 @@
+# Agent Note: Explicit model-facing tool order
+
+Status: implemented
+
+## Problem
+
+Model-facing tool order followed plugin registration order, which depends on concurrent module loading for otherwise independent plugins. That race produced different request headers in CI and snapshot recordings. Because order affects request bytes, caching, and the durable header, it needs an explicit deterministic policy.
+
+## Decision
+
+The system-prompt assembly owns the canonical model-facing tool order, exactly where it already owns section order. `toolOrder?: string[]` on `dsh-system-prompt` is the optional explicit policy:
+
+- A listed tool that is registered takes its listed position.
+- A listed name with no registered tool is a configuration error. Shape errors (rest entry missing or duplicate names) fail from the service constructor; an unregistered name rejects every `assemble()` — the earliest moment the registered tool set exists to check against (tool plugins register after the service constructs), and the only universal one (registrations can change at any time; cordis has no "all plugins loaded" event). Under the shipped loop the first turn fails before any model request — see the consequences below for the exact blast radius.
+- A registered tool absent from the list is inserted at the `'<unlisted-tools>'` rest entry (`TOOL_ORDER_REST`), in lexicographic name order among the other unlisted tools.
+- No collected tool may use `TOOL_ORDER_REST` as its `ToolSchema.name`; the assembly rejects that reserved name before ordering.
+- The list must contain the rest entry exactly once and no duplicate names.
+- When `toolOrder` is unset, the canonical order is plain lexicographic name order (code-unit comparison, locale-independent), so determinism requires no configuration.
+
+`assemble()` canonicalizes provider tools before the `system-prompt/assemble` waterfall, removing registration-order variance at its source. The waterfall starts from this deterministic list; unchanged order then flows into the request header, frozen request, and reconstruction checks without loop-specific ordering logic.
+
+Scope is deliberately narrow: this fixes the REGISTRATION-ORDER race, not plugin behavior. A `system-prompt/assemble` listener may still add, remove, or rearrange tools — same as it may edit sections after their sort — and owns the determinism of what it emits; the waterfall contract already demands deterministic listeners (the reconstructability invariant would catch a listener that diverges between build and replay).
+
+Config plumbing follows the `persona` precedent, and `toolOrder` sits beside it: the app configs (`dsh-stdio-demo`, `dsh-acp-demo`) accept the key and forward it through `dsh-agent-spine-demo` (whose schema is the intersection of the owners' schemas) to the `SystemPrompt` child. One schemastery footnote is load-bearing: a schemastery array defaults to `[]`, but an omitted `toolOrder` must stay ABSENT (= lexicographic) rather than become an explicitly-configured empty list (invalid — it lacks the rest entry), so every schema on the chain forces the default to `undefined`.
+
+## Alternatives considered
+
+- **Registration order (the status quo)** — a concurrent-import race, host-dependent (the CI flake above), invisible in review.
+- **A linearization of the plugin dependency graph** — the relation is partial and independent tool plugins are incomparable; the flake happened with the partial order fully satisfied.
+- **Per-plugin `weight` on each tool contribution** — scatters the order across plugins yet still needs a global numbering convention nobody owns (the section `order` bands show that coordination cost being paid by hand).
+- **Sorting in `ToolRegistry.schemas()` (the registry layer)** — equally deterministic, but the registry is a membership store consumed by more than the assembly; ordering is a prompt-composition concern, and the assembly already owns the composition policy for sections.
+- **A `LlmService` config + `orderTools()` method the loop calls before logging the header** — works, but adds a public service method and a loop edit solely to apply a policy at a distance; every future request composer must remember the call. Canonicalizing where the list is born makes an unordered list unrepresentable, with zero new surface.
+- **Normalizing inside `llm.stream()`** — runs after the header event is logged (the flake survives) and rebuilds the deep-frozen envelope, silently disarming the reconstruction invariant.
+- **An exhaustive list (no rest entry)** — every newly loaded tool plugin would break boot; the mandatory rest entry keeps unlisted tools deterministic and their position explicit.
+- **A boot-time validation pass (a `SystemPrompt.assertToolOrderSatisfied()` called by `dsh-app-boot` after `loader.await()`)** — would turn the misconfiguration into a startup death instead of a first-turn failure, but costs a public service method plus a structural coupling from the generic boot glue to one service, and cannot replace the assembly-time check anyway (embedded callers never run app boot; registrations change after boot). No existing event can host the check either: cordis v4 has no ready-like event, `loader/entry-init`/`internal/status` fire mid-load (racy against tool registration, the very entropy this Agent Note kills), and the agent lifecycle events are no earlier than the assembly. One enforcement point at `assemble()` was judged worth the later failure moment.
+
+## Consequences
+
+- Every registry-built assembly starts with a deterministic tool order on every host; absent an expert listener that deliberately changes it, every `request/header` event and model request inherits that order. The CI-vs-local registration-order flip is structurally gone, and the default is lexicographic.
+- The initial `PromptAssembly.tools` is canonical, so waterfall listeners start from the model-facing order; provider registration order is observable nowhere before that cooperative seam.
+- The snapshot suite's single pinned request-header fixture (`text-turn`) carries the new canonical tool order; every other ACP snapshot keeps the header bulk scrubbed as `{{system}}`/`{{tools}}`, per the pinned-header design.
+- A pure tool reordering between steps is logged like any other header change: a full `request/header` snapshot with reason `'change'`. Stable canonical order prevents registration timing from creating such changes in the ordinary path.
+- The `toolOrder` key rides the app → `agent-core` → `SystemPrompt` forwarding chain, so deployments set it next to `persona` in the app config; `dsh-llm` and the agent loop are untouched.
+- A misspelled or unloaded tool name in `toolOrder` fails the turn at prompt assembly, not the boot: the loop assembles inside the turn (after `turn/start`, before `step/start`), so the rejection reaches the turn's outer catch — the turn closes balanced with an `error` reason carrying the message, `agent/error` mirrors it, no step opens, no `request/header` is logged, no request reaches the adapter, and the agent returns to idle. Every turn fails identically until the config is fixed; the process itself stays up (matching the repo rule that explicit config references must not be silently ignored — the enforcement point is the assembly because no earlier universal moment exists).
+- A tool provider that returns the reserved rest-entry name has the same prompt-assembly failure shape as an unknown listed name. This keeps the sentinel from becoming an ambiguous real tool and preserves the "never drops a tool" ordering contract.
+
+## Testing
+
+System-prompt tests cover lexicographic default order, listed/rest placement, provider-order independence, shared names, invalid lists, unknown or reserved names, the canonical pre-waterfall list, and the rule that listener-added tools are not re-sorted. Loop tests pin identical logged and dispatched order across registration permutations, forwarding through agent-core and both apps, deep-frozen requests, and balanced turn failure with no step, header, or adapter call for an unknown configured name. Snapshot replay keeps the full canonical list only in the pinned `text-turn` header; other fixtures continue to use `{{tools}}`.

.agents/notes/implemented/feature/2026-07-06-sandbox.md

@@ -0,0 +1,210 @@
+# Agent Note: The subprocess sandbox — confinement seam, native runners, escalation, and per-session modes
+
+Status: implemented
+
+## Problem
+
+A coding agent needs this product path: bash subprocesses — and the hook commands that ride them — execute under a restricted file sandbox by default; if and only if the sandbox actually denies an operation, the model may request one user approval for that same operation and, once granted, retry it once with wider permissions. An every-tool boundary is deliberately NOT the claim: fs/web/todo execute in-process where an `execve` wrapper is meaningless (§ In-process tools), and the cross-family boundary is staged follow-up work (§ Deferred phases). Without a shared vocabulary, every tool reinvents approval fields, denial parsing, retry matching, and permission-state hints.
+
+The harness is an SDK, so confinement must be a capability developers COMPOSE: whether to sandbox, and which backend per platform, belongs in the leaf `cordis.yml` as a first-class entry — not inside one executor's private machinery. And the first-choice runner, `bwrap`, is unusable on exactly the hosts a sandbox matters most (minimal containers, disabled unprivileged userns, LSMs that deny `mount`), so a fallback runner has to ship with the SDK rather than be assumed on the host.
+
+Confinement alone leaves two gaps. A denial with no escalation path is terminal — the model can only give up, which pressure-cooks operators into configuring `workspace-write` or `danger-full-access` globally and defeats the sandbox. And the model-visible knobs (the sandbox mode, the approval policy) change over an agent's lifetime — an ACP user flips a per-session setting, an operator edits `cordis.yml` while the process is down — while the model must never act on a stale belief about them: what IS the state on every request, what changed while the agent lives, and what changed while nobody was watching all need answers.
+
+## Decision
+
+One seam, one per-platform chain of local backends, one consumer, and two levers on top: a per-call escalation path and per-session runtime modes. Everything below composes from the leaf `cordis.yml`; nothing touches `agent-loop`. The scope is deliberately bounded: the phases this Agent Note names but does not design — per-session workspace root, cross-family fs enforcement, the `subagent-acp` consumer, more environments, a Windows chain — are listed under § Deferred phases, each a follow-up design, not a config knob.
+
+### How a deployment uses it
+
+Four `cordis.yml` entries turn an unconfined coding agent into the sandboxed product path; [`examples/acp-agent`](../../../../examples/acp-agent/README.md) uses this composition by default:
+
+```yaml
+- id: sandbox
+  name: '@deepseek-ai/dsh-sandbox-local'   # the per-platform runner provider (ctx.sandbox)
+- id: bash
+  name: '@deepseek-ai/dsh-bash-sandbox'    # the confined executor, replacing dsh-bash-local behind ctx.bash
+  config:
+    mode: workspace-write                  # the deployment default every session starts from
+    workspaceRoot: !!js process.cwd()      # the boundary workspace-write may write under
+- id: approval
+  name: '@deepseek-ai/dsh-user-approval'        # the escalation gate's channel (the approval Agent Note)
+  config:
+    policy: ask
+- id: permission
+  name: '@deepseek-ai/dsh-permission'      # one product-facing select over both mechanism knobs
+```
+
+The swap is invisible to every consumer of `ctx.bash`: the bash tools, hook commands, and background tasks run exactly as before, spawned through the wrapped argv the provider returns. Deleting the `sandbox` and `permission` entries and replacing `bash` with `@deepseek-ai/dsh-bash-local` is the opt-out — execution is unconfined again and the escalation fields vanish from the tool schema, because they are capability-gated on the mounted executor, not on configuration. Omitting only `approval` keeps confinement but fails every escalation closed with its own error text; `permission` also requires the approval seam and a confining executor, so a partially composed preset layer fails loud at load.
+
+Misconfiguration fails loud: `mode` outside the closed vocabulary is rejected at plugin load, and a host with no usable backend throws the structured `SANDBOX_UNAVAILABLE` — at `confine()` before the command ever spawns — rather than degrading to unconfined execution. `runnerCommand` on `dsh-sandbox-local` is the operator's explicit assertion of a bwrap-compatible runner (chain and probes skipped); it doubles as the deterministic fake-runner seam for keyless tests.
+
+Denied file effects return a `[sandbox: file access denied under <mode> mode]` marker and instructions not to work around the denial. A confining executor adds paired `sandbox_permissions` and `justification` fields for one approved retry that must be strictly wider than the session's effective mode. A grant widens only that retry; rejection executes nothing, returns `the user rejected escalating this command to "<mode>"`, and permits no re-ask. The prompt does not announce sandbox mode, avoiding preemptive refusal. When `dsh-permission` is composed, ACP exposes one `Permissions` select whose presets write both knob events; unmatched knobs appear as switch-away-only `custom`. Only a switch to the deterministic `'never'` approval policy is stated in the prompt and narrated.
+
+### Design detail
+
+#### Scope grounding
+
+OS subprocess confinement applies to the bash executor, including hook commands, and later to ACP subagent children. Filesystem, web, and other tools execute in-process and require policy at their own seams; an argv wrapper cannot confine a function closing over `ctx`. The existing bash request/spec split carries per-call overrides, while `tools/pre-execute` and the approval seam own the human decision.
+
+#### The seam: `ctx.sandbox`
+
+`dsh-sandbox` owns the vocabulary and the `SandboxProvider` contract: `confine(argv, policy)` returns the argv to spawn INSTEAD of the caller's own — wrapped so the process and everything it spawns run confined — plus the `enforcement` completeness the selected backend achieves, its denial dialect (`denialSignatures`, the stderr substrings that backend's kernel prints on a denied file effect), and its runner-failure dialect (`runnerFailureSignatures`, how the runner ITSELF failing — and therefore the command never running — identifies itself); with no usable backend it throws the fail-closed `SANDBOX_UNAVAILABLE` error, never a silent unconfined passthrough. The vocabulary: `SandboxMode` (`read-only` / `workspace-write` / `danger-full-access`, FILE effects only — network and process visibility are not claimed), `SandboxEnforcement` (`full` / `partial`), `SandboxPolicy` (mode + workspace root).
+
+Policy rides each CALL, not the provider: two consumers may confine under different policies at the same instant (bash under `read-only` while a confined child agent keeps its state directory writable), and an approved escalated retry is a new call with a wider policy — inexpressible under a config-fixed provider mode.
+
+The seam confines SAME-WORLD subprocesses only: a backend shares the host's filesystem and kernel. Containers, microVMs, and remote executors are NOT backends of this seam — they replace whole capability implementations (`ctx.bash`, `ctx.fs`) as environment-coherent groups, because an agent whose bash runs in a container while its fs tools write the host lives in two split worlds.
+
+Left open, for the phase that needs them: whether network restriction arrives as a separate `network_mode` or merges into `sandbox_mode` once a runner enforces both, and whether `SandboxPolicy` grows extra writable-root grants now (the launcher already speaks `--rw <path>`) or only when escalation needs them.
+
+#### Local backends and the shipped launcher
+
+`dsh-sandbox-local` selects one platform runner per provider lifetime and caches the verdict. Linux functionally probes `bwrap` then Landlock; macOS uses Seatbelt. Unsupported platforms and unusable runners fail closed. Each wrap carries backend-specific denial and runner-failure signatures so `dsh-bash-sandbox` can distinguish a denied file effect from a broken sandbox. `runnerCommand` skips selection as an operator assertion of a bwrap-shaped runner, but missing or unexecutable commands still classify as sandbox failure and never run the payload unconfined.
+
+The launcher is a ~300-line C program (plain C11 over the raw Landlock UAPI — no libraries beyond a statically linked musl, so the audit surface is that one file plus the kernel's stable syscall contract): `--ro <path>` / `--rw <path>` grants, `--`, the wrapped argv; it installs the ruleset on itself and `exec`s (rulesets are inherited across `execve`, and it sets `no_new_privs` before restricting); `--probe` enforces a maximal ruleset in a short-lived child and exits 0 only when the kernel actually enforces; launcher failures exit 125 without exec'ing.
+
+The Landlock launcher ships through [`node-addon-landlock-run`](https://www.npmjs.com/package/node-addon-landlock-run), with platform binaries selected by npm. That package owns path resolution, probing, and CLI flags; the harness maps sandbox modes to grants. Versioning the entry point with its binaries keeps probe parsing and launch syntax aligned.
+
+FIXME: Revisit the separate-repository boundary and try to maintain the launcher source and its platform package family inside this monorepo, so the native release surface and harness contract evolve together.
+
+Backend profiles share the mode contract but differ in necessary host grants. Landlock and Seatbelt allow only `/dev/null` in read-only mode; workspace-write also permits their required host temp roots. Each wrap carries backend-specific denial signatures. Landlock reports partial enforcement on older ABIs that cannot govern every operation, while successful bwrap and Seatbelt profiles report full enforcement.
+
+#### The bash consumer
+
+`dsh-bash-sandbox` extends `LocalBashExecutor` and hands `ctx.sandbox` the exact `['bash', '-c', command]` argv it is about to spawn. A denial is an orthogonal result fact, conservatively classified from the active runner's stderr dialect. A runner failure outranks denial: foreground execution throws `SANDBOX_UNAVAILABLE`; a settled `BashProcess` stamps `sandbox.runnerFailed`, and the bash producer renders it through generic `task_output`.
+
+The model's view is result facts only: the static tool description explains the denial marker (`[sandbox: file access denied under <mode> mode]`), encourages attempting commands that may be denied, and forbids retrying around a denial; when the escalation fields are advertised, a denied result additionally carries the escalation hint itself, so the sanctioned same-turn retry is prompted at the decision point rather than depending on the model recalling the description (§ Escalation). No prompt section states the sandbox mode (§ Per-session modes).
+
+#### Escalation: one approved wider retry after a denial
+
+`BashExecRequest.sandboxMode` is an optional per-call input; resolved specs make the field explicit. `BashExecutor.sandboxMode` advertises whether the mounted executor can honor it, so only a confining composition exposes escalation. The seam accepts any explicit mode; the tool owns the wider-only escalation rule. Non-sandboxing executors remain honestly unconfined.
+
+`SandboxBashExecutor.resolve()` stamps the effective mode — escalation grant > session override > configured default — so `run()`/`start()` read the spec, never the config. Per-process wrap facts are keyed by the returned `BashProcess`; `onProcessDone()` classifies stderr and stamps that handle before `done` resolves, so overlapping processes retain their own modes and runner dialects.
+
+When a confining executor is mounted, `bash` advertises paired `sandbox_permissions` and `justification` fields. The schema exposes the full closed escalation vocabulary because effective mode is per-session; execution rejects any target that is not strictly wider than that call's effective mode. Approval resolves before execution. `allowed-once` stamps the granted mode onto only that request, while `rejected`, `cancelled`, `unavailable`, a missing approval service, or a missing agent all fail closed with distinct results. No grant is persisted.
+
+Escalation is a same-turn retry of the denied command with the narrowest sufficient `sandbox_permissions` and a `justification`; the approval prompt is the consent step. It must be grounded in an actual denial, except when the session already observed the same denied access, and a disabled or rejected approval ends that command. The retry, approval decision, and result use existing tool and approval events. `dsh-tool-bash` owns the ask because the executor seam has neither the agent nor call id required for user interaction.
+
+Left open: what a durable grant's scope identity is beyond the sandbox mode — exact call, path, command prefix, session, or time window — before an `allow_always` option can be advertised.
+
+#### Per-session modes: the session log as the store
+
+```
+effective(session) = findLast(the session's own knob events)?.value ?? the composition-config default
+```
+
+The default is composition config (`cordis.yml`) — operator-owned, process-wide. A runtime switch is a SESSION-SCOPED override recorded as one log-only event in that session's own log. Restart immunity (resuming a session replays its log, so overrides come back with zero catch-up machinery) and multi-session isolation (one editor tab's `workspace-write` cannot disturb another's `read-only`) both fall out by construction, and no external config store exists anywhere.
+
+**One event per knob, owned by its domain** — the merge-extensible `SessionEventMap` idiom every existing event family already follows (`approval/*` in `dsh-user-approval`, `hook/*` in the hooks packages):
+
+```ts
+interface SessionEventMap {
+  'bash/sandbox-mode': { mode: 'read-only' | 'workspace-write' | 'danger-full-access' }
+  'approval/policy': { policy: 'ask' | 'never' }
+}
+```
+
+Each owner exports the same three-piece kit: the event declaration, a pure fold (`effectiveSandboxMode(events)` / `effectiveApprovalPolicy(events)` — a `findLast`, typed to the domain's closed union), and THE write path (`setSandboxMode(session, mode)` / `setApprovalPolicy(session, policy)` — a switch IS its event; nothing mutates state out of band). No shared owner service, no generic facts map, no registry: a third knob copies the ~40-line pattern into its own package. Execution follows the fold on both sides — the bash tool's per-call stamp reads it as the middle rung of the § Escalation precedence chain, and the approval seam's `'never'` gate is [the approval Agent Note](2026-07-06-approval-seam.md)'s side of the same pattern.
+
+Sandbox mode is not narrated in the prompt; denial results report the mode when it matters, avoiding preemptive refusal based on a standing label. Approval policy is different: only `'never'` is stated because automatic rejection otherwise looks like a user decision. Policy-change notices are coalesced and delivered by the next pre-step, with log-derived fallback after restart. The notice source is inferred from event position: a knob event after the last request header is user-driven; unlogged drift is operator or config driven.
+
+**The editor surface** is protocol-native [Session Config Options](https://agentclientprotocol.com/protocol/session-config-options) — the spec's replacement for session modes (slated for removal in ACP v2), already SDK-typed. When `ctx.permission` is composed, the bridge advertises one `permission` select (category `mode`) in `session/new` and `session/load`; its options are the deployment's preset table, and its `currentValue` is `PermissionService.current()` over the session log plus composition defaults. The shipped `workspace-write` and `danger-full-access` presets each bundle a sandbox mode with an approval policy and write through to both domain setters; a knob combination outside the table is reported as switch-away-only `custom`. `session/set_config_option` validates and switches through the permission service, then returns the complete refreshed state (the spec contract).
+
+**Turn enclosure is the commit boundary.** A switch during an open turn appends immediately. An idle switch remains pending on the bridge record and is appended at the next prompt submission, before assembly or execution; last write wins per knob. Openness comes from log boundaries rather than `agent.status`, and setters do not append from inside a `session/event` listener because that would reorder later observers. Until anchoring, responses overlay the pending value. A crash discards it, and reload returns the durable fold.
+
+#### In-process tools
+
+fs/web/todo execute in-process, so their sandbox semantics are policy at their seams: the fs intent gates deciding by the shared mode vocabulary (§ Deferred phases, cross-family) make `read-only` a real boundary instead of a bash-only approximation — until then the contract says so honestly. No generic per-tool sandbox runtime: a host-mediated tool leaves the process only by returning declarative effects the host validates, which is a rewrite, not a wrapper.
+
+FIXME: Revisit this tool-local boundary. The follow-up design needs to determine whether sandboxing becomes a global harness capability that applies uniformly to every tool, instead of expressing in-process enforcement independently at each tool seam.
+
+### Testing
+
+- **Unit:** pin platform selection and profiles, fail-closed runner classification, per-call facts, escalation validation and outcomes, permission preset folding and write-through, narrator coalescing, ACP advertisement and validation, and turn-enclosed config writes.
+- **Keyless real-runner:** exercise bwrap, Landlock, and Seatbelt against real filesystem effects at provider and bash-consumer layers; packed-install coverage proves the registry launcher remains executable. The real ACP composition pins permission switching and rejects unknown presets. CI rejects a silent all-skip.
+- **With-key:** drive a real model, runner, bridge answerer, and disk effect through granted and rejected escalation; unavailable credentials or runners self-skip.
+- **Snapshot:** pin the permission config-option wire, preset and knob events, prompt deltas and notices, and both scripted approval branches. Snapshot mode starts unconfined so unrelated fixtures remain platform-independent; policy scenarios switch explicitly. Real denial stderr stays on platform tests because its dialect is runner-specific.
+
+## Deferred phases
+
+Each phase gets its full design when picked up, validated against the code at that time, and lands with unit, real-API e2e, and snapshot coverage at the tiers it touches.
+
+- **Per-session workspace root** — the executor's write boundary stays config-fixed for its lifetime while each ACP session has its own cwd; a per-session root rides the same per-call policy carrier once designed.
+- **Cross-family boundary** — the fs intent gates decide by the shared mode, making `read-only`/`workspace-write` real boundaries beyond bash.
+- **Second consumer** — `subagent-acp` optionally confines child agents (per-call policy; unconfined default — a child agent must write its own persistence).
+- **More environments** — an environment-coherent capability group example (e.g. bash+fs against one container).
+- **Windows chain** — `PLATFORM_CHAINS.win32` is reserved and empty (fail-closed); filling it means a confinement runner from the AppContainer/restricted-token family, shipped from its own repository on the `node-addon-landlock-run` template, plus its profile dialect and denial/runner-failure signatures.
+
+## Alternatives considered
+
+- **Command-string heuristic preflight** — rejected: cannot understand expansion/subprocesses/symlinks; the strict attempt (run it, let the kernel decide) is the only trustworthy denial signal.
+- **Functionally probe even a platform's sole backend** — rejected: probing arbitrates between candidates; with one there is nothing to decide, and probe cost taxes the first confined command of every session (prohibitive for heavy future backends). The runner's own exec-time fail-closed refusal plus `runnerFailureSignatures` classification carries the safety property instead.
+- **Commit the built launcher binaries** — rejected: a binary in a diff is unreviewable and churns history; reviewed source + native CI builds + the launcher repo's byte-pinned publish rehearsal keep bytes out of every tree.
+- **Compile the launcher on install** — rejected: pushes a C toolchain onto every consumer; a fallback that exists only where a compiler happens to be is not a fallback.
+- **Cross-compile both architectures from one builder** — rejected: requires carrying a pinned cross toolchain (rustup targets, zig, or a container image) solely to rebuild two ~70 KB binaries; per-architecture native runners already exist and each builds its own platform package (the `node-addon-require-builtin` model, the launcher repo's own pipeline).
+- **No fallback (bwrap or fail closed)** — rejected: concentrates failure on the hosts a sandbox matters most, degrading to `danger-full-access` by resignation.
+- **Keep the mechanism inside `dsh-bash-sandbox`** — rejected: blocks the existing second consumer, makes future phases read mode out of a bash plugin's config, and cannot express escalation.
+- **Config-fixed mode on the provider** — rejected: one mode per process; cannot serve concurrent consumers with different policies nor the one-shot widened retry.
+- **One interface spanning containers/VMs too** — rejected: `confine(argv)` presupposes a shared filesystem; environment isolation is capability-sibling backends deployed as coherent groups.
+- **Generic ToolRuntime wrapping any tool** — rejected: mechanically false for in-process tools (closures over `ctx`); the declarative-effects rewrite is unjustified for fs/web/todo.
+- **Ask inside the executor (`dsh-bash-sandbox`)** — rejected: no `agent` to route through, no `callId` to attach the prompt to; adding them teaches a transport seam about sessions and UIs — the tool layer holds both and owns the model-facing vocabulary.
+- **Auto-retry inside the same tool call** — rejected: a hidden re-entry the log cannot reconstruct: one `tool/call` would have produced two executions with different policies — the retry is a NEW logged call with its own arguments and result facts.
+- **Advertise the escalation fields unconditionally** — rejected: under `dsh-bash-local` they are a dead lever — advertising an option the harness cannot honor manufactures doomed grants; capability-gating costs one registration-time read.
+- **A default-relative escalation ladder (advertise only the modes wider than the executor's registration-time default)** — rejected: per-session overrides make the default the wrong baseline — a session switched narrower than the default loses exactly the lever it needs, and under a `danger-full-access` default the fields vanish entirely while a `read-only`-overridden session stays confined with no escalation path. The enum pins the closed target vocabulary; strict widening is a per-call execution check against the session's effective mode.
+- **Per-session dynamic tool schemas** — rejected: schemas are registry-global by design (one assembly vocabulary, the pinned-header snapshot contract), and re-registering per session would buy only what the execution-time strict-wider check already guarantees, at the cost of a per-session schema surface and header churn on every switch.
+- **Hard-match the retry to a prior denial** — rejected: command-string identity is fragile (quoting, `workdir`, env prefixes, a pipeline retried as its failing stage) — false-rejects honest retries or is trivially satisfied; the real boundary is the human seeing command + justification. Revisit only if `allow_always` grant storage ever needs machine-checkable scopes.
+- **A generic `env/state` facts map with an owner service** — rejected: approval and sandbox compose independently, so neither's state may drag in a third package; single-key folds are one `findLast` each, dissolving the owner service; no invariant spans the knobs, so atomic multi-key patches bought nothing.
+- **Narrate via `agent/user-message` + a bus event** — rejected: it presupposes a turn-entry seam that does not exist (the real seam is `agent/prompt-submit`), and pre-step's position serves both the coalesced turn-entry notice and the mid-turn immediacy bound with one listener.
+- **A standing prompt statement of the sandbox mode (+ a switch narrator)** — shipped first, then removed on live evidence: with `Bash commands run under the "read-only" file sandbox.` in every request, the model refused to ATTEMPT denied-then-escalatable work (five of twelve turns in the first manual session ended with zero tool calls), turning the sandbox into a soft lockout. The denial marker names the mode at the moment it matters and the escalation fields carry the recovery; the approval knob keeps its statement because an auto-rejection is behaviorally indistinguishable from a human "no".
+- **Track "last told" with its own bookkeeping events** — rejected: the `request/header` fold already records the exact prompt the model saw; parsing the closed candidate sentences back replaces a second bookkeeping stream — events are needed only where they ARE the store.
+- **ACP session modes instead of config options** — rejected: the preset is already one deployment-defined config-option select, and modes are slated for removal in ACP v2.
+
+## Consequences
+
+What shipped pins — the tiers in Testing hold each:
+
+- A denied command retried with `sandbox_permissions` + `justification` prompts the user through the composed answerer chain; a grant runs THAT call under the wider mode (result facts say so) while every other call keeps its own effective mode; every non-grant outcome produces its distinct error text and executes nothing.
+- The escalation fields exist exactly when the mounted executor confines; a request that is not strictly wider than the call's effective mode fails closed with its own text and prompts no one; a deployment with no ApprovalService fails escalating calls closed and leaves plain calls untouched.
+- The system prompt never states the sandbox mode (an approval `'never'` policy is the one stated knob), and the whole exchange — headers, knob events, notices, approvals, results — reconstructs from the session log alone, with no event types beyond the two knob events.
+- N idle-time flips produce at most one anchored event per knob (a net-zero sequence anchors none — a no-op push from a client echoing current selections records nothing); an approval-policy switch is narrated in at most one coalesced notice; a mid-turn sandbox switch is honored by the next call's stamp.
+- A resumed session's overrides apply and are reported to the editor with no special-casing; a default changed while the process was down is narrated before the session's first new request, attributed to the operator.
+- Two concurrent sessions never see each other's state, notices, or config options.
+- `agent-loop` is untouched — everything rides `systemPrompt.section`, `SessionEventMap` merging, `agent.inject()`, `agent/pre-step`, `agent/prompt-submit`, and the ACP handler surface.
+
+Costs and accepted limits:
+
+- **The one-wrapper illusion is given up knowingly.** A `tools/pre-execute` wrapper plus prompt conventions does not solve sandbox approval — the correct design costs structured denials, native runner probes, per-call policy carriage, and consistent cross-family enforcement, and this design pays it.
+- **`read-only` is not yet a cross-family boundary.** Until the fs intent gates decide by the shared mode, the claim holds for bash only; the contract says so honestly (§ In-process tools).
+- **Windows has no backend.** Its chain slot is reserved empty — fail-closed, never a fallthrough; filling it is a deferred phase.
+- **The Seatbelt rung leans on Apple's deprecated-but-shipped `sandbox-exec` CLI.** As darwin's sole candidate it is selected without probing, so a future removal surfaces at execution as the runner-failure classification — re-thrown `SANDBOX_UNAVAILABLE`, the command never runs; fail closed, never open.
+- **Landlock confinement is only as complete as the running kernel's ABI.** Reported as `enforcement: 'partial'` rather than refused — the deliberate trade that keeps the fallback available on older-kernel hosts.
+- **The launcher arrives as a registry dependency.** Trusted through its own repository's release pipeline (reviewed C source, native CI builders, byte-pinned publish rehearsal) plus this repo's version pin — the real-kernel e2e legs are what vouch for behavior through the installed bytes.
+- **The model may over-ask.** Escalating without denial grounding, or picking `danger-full-access` where `workspace-write` suffices: the description steers and the enum forces the ladder, but the human prompt is the actual gate; the `approval/asked` reasons make over-asking auditable, and a `prepend` policy answerer can auto-reject patterns a deployment never wants.
+- **The advertised target set is static while the effective mode is per-session** (schemas are registry-global) — a session already at the widest mode is still offered the fields. Harmless by construction: the strict-wider check at execution, not the enum, is the safety boundary — a non-widening request fails with its own text and never prompts anyone.
+- **A granted escalation is not a working sandbox.** An unavailable backend still fails closed even for a granted escalation to a confining mode — at `confine()` when the platform has no chain or every probe fails, at execution when an unprobed sole runner refuses (classified as a sandbox failure, not a command failure) — while a granted `danger-full-access` run never touches the provider at all: there the grant, not the probe, is the authority.
+- **An idle switch lives in bridge memory until the next prompt submission anchors it.** A crash in that window reverts it (reported on `session/load`), and a session that never submits another prompt never persists it — accepted, with a loop-owned idle commit turn left as future work if durability becomes required.
+- **The approval narrator's restart baseline parses prompt prose.** The closed candidate sentence is owned by the writing module itself, so a wording change is a coordinated writer+parser edit in one file; a session whose headers predate the section silently adopts the current policy without a notice.
+- **The approval section is still a dynamic prompt surface** (a `'never'` switch breaks provider prompt-prefix caching for that session). Accepted: policy switches are rare, and a model acting on a stale `'never'` is worse. The sandbox knob no longer touches the prompt at all.
+- **The model may hold a stale belief about the sandbox mode** (nothing announces a switch). Accepted deliberately: the next attempt's marker or success corrects it, and the observed failure mode of announcing — preemptive refusal — is worse than one wasted retry.
+
+## FAQ
+
+- **A command came back with `[sandbox: file access denied under read-only mode]` — did it fail?** It RAN, and the kernel refused a file effect: the denial is a result fact orthogonal to exit code. The teaching forbids retrying around it; the one sanctioned move is the same command retried once with an escalation request.
+- **How is a BROKEN sandbox told apart from a failing command?** Runner failure outranks denial in classification: a failed run matching the wrap's `runnerFailureSignatures` means the command NEVER ran — foreground re-throws the structured `SANDBOX_UNAVAILABLE` with the runner's stderr line, a background task stamps `sandbox.runnerFailed` and renders its own marker. A broken sandbox can never read as a failing command, and the command never runs unconfined.
+- **What happens on a platform with no backend — Windows today?** `confine()` throws the fail-closed `SANDBOX_UNAVAILABLE` and the command never spawns; `win32` is a reserved EMPTY chain, pinned by test to fail closed identically until a Windows runner fills it (§ Deferred phases).
+- **`bwrap` is installed on my host but unusable (disabled unprivileged userns, an LSM denying `mount`) — what happens?** The chain probe is functional — it builds and enforces a real profile rather than checking `--version` — so a present-but-unusable `bwrap` fails its probe, selection falls to the registry-installed Landlock launcher, and the verdict is cached for the provider's lifetime.
+- **Does the sandbox restrict network or process visibility?** No — `SandboxMode` claims FILE effects only; the bwrap profile deliberately does not unshare pid, and no backend claims network. Whether network restriction becomes its own knob is left open in § The seam.
+- **Which tools actually run confined?** OS subprocesses through `ctx.bash` — the bash tools, and hook commands transitively. fs/web/todo execute in-process, where an `execve` wrapper is mechanically meaningless; their `read-only` semantics arrive with the cross-family deferred phase, and until then the contract says bash-only honestly.
+- **Does a granted escalation persist?** No. The grant is consumed by the exact foreground or background call that asked; every neighboring call keeps its own effective mode. A later background denial surfaces through `task_output` and may ground a new exact-command retry.
+- **When does an editor's mode switch take effect?** Mid-turn: appended immediately, honored by the very next call's stamp. Idle: held on the bridge's session record, anchored at the next turn's `agent/prompt-submit`, with N flips coalescing to at most one event (none if net-zero); a crash before anchoring reverts it and `session/load` reports the truth. The model is not told — its next command simply behaves under the new mode.
+- **What survives a restart — and what if the operator changed the config default while the process was down?** Overrides replay from the session log (`effective = fold ?? config`), so a resumed session keeps its modes with zero catch-up machinery; a default that drifted offline changes behavior the same way a switch does (the approval policy, being stated, is additionally narrated with operator/config attribution).
+- **What does `enforcement: 'partial'` on a result mean?** The selected backend enforces the subset its kernel ABI governs — e.g. Landlock before ABI v3 does not govern path truncate — and says so structurally instead of refusing the host; the probe's report line distinguishes the cases. The bwrap and Seatbelt profiles govern every promised file effect by construction, so they always report `full`.
+
+## Prior art
+
+In-repo precedents this design copies or contrasts with:
+
+- [The capability-seams Agent Note](../architecture/2026-06-13-capability-seams.md) — the interface/implementation/consumer split and the "don't split preemptively" timing rule the second consumer satisfied.
+- The `dsh-bash` request/spec split ([the bash vocabulary catalog](../../../../docs/core-data-structures/bash.md)) — the per-call carrier template `sandboxMode` rides, and the explicit-`resolve()` defaulting convention.
+- [The approval seam Agent Note](2026-07-06-approval-seam.md) — the channel escalation asks through; its answerer waterfall, audit pair, and one-package rationale are recorded there.
+- [Event-sourced sessions](../architecture/2026-06-11-event-sourced-sessions.md) and [the turn-enclosure invariant](../architecture/2026-06-15-turn-enclosure-invariant.md) — the log-as-store foundation the per-session modes fold over, and the commit boundary the anchoring design obeys.
+- [The interception-seams Agent Note](2026-06-30-interception-seams.md) — the `tools/pre-execute` vocabulary the escalation gate deliberately does not reuse (an escalating call has no pre-execute moment of its own).

.agents/notes/implemented/feature/2026-07-07-mcp-client-plugin.md

@@ -0,0 +1,212 @@
+# Agent Note: MCP client plugin — connect to external MCP servers and bridge their tools
+
+Status: implemented
+
+## Problem
+
+The harness had no way to consume tools from the MCP (Model Context Protocol) ecosystem. MCP is the emerging standard for tool servers — GitHub, filesystem, databases, code search, and hundreds of community servers expose tools via MCP. Users want to point the harness at one or more MCP servers and have their tools appear as native model-facing tools, without writing per-server glue code.
+
+The `ToolRegistry` already accepts raw JSON Schema tool definitions (documented in `dsh-tools` README: "Raw JSON-Schema tool definitions (from MCP servers) are still accepted by `ToolRegistry.register()` directly"), and the extension cookbook sketches the intended pattern ("MCP | one plugin per server: discover tools → `ctx.tools.register()`"). The infrastructure was ready; the bridge plugin was missing.
+
+## Decision
+
+### Package
+
+A single package `@deepseek-ai/dsh-mcp-client` at `packages/mcp/mcp-client/`. No capability-seam three-package split — there is no foreseeable second MCP client implementation, and the convention is "don't split preemptively" ([capability seams Agent Note](../architecture/2026-06-13-capability-seams.md)).
+
+### SDK
+
+Use the official [`@modelcontextprotocol/sdk`](https://github.com/modelcontextprotocol/typescript-sdk) (`Client`, `StdioClientTransport`, `StreamableHTTPClientTransport`). The harness does not implement its own JSON-RPC — consistent with how ACP delegates to `@agentclientprotocol/sdk`.
+
+### Scope
+
+MCP Client only (no server side — ACP already covers the "expose harness as an agent" role). Bridge **Tools** only — Resources and Prompts are deferred (they require harness-side consumption mechanisms that don't exist yet, and design space is large).
+
+### Plugin shape
+
+Namespace plugin (named exports `name`/`inject`/`Config`/`apply`, no `export default`). `inject: ['tools']`. Each MCP server is one plugin instance in `cordis.yml` — the same package loaded N times with different configs, like `dsh-tool-subagent`.
+
+### Configuration
+
+Flat discriminated union on the `transport` field:
+
+```typescript
+interface StdioConfig {
+  transport: 'stdio'
+  serverName: string          // required namespace, ^[A-Za-z0-9_-]{1,32}$
+  command: string
+  args?: string[]
+  env?: Record<string, string>
+  cwd?: string
+  toolCallTimeoutMs?: number  // default 60_000
+}
+
+interface StreamableHttpConfig {
+  transport: 'streamable-http'
+  serverName: string          // required namespace, ^[A-Za-z0-9_-]{1,32}$
+  url: string
+  headers?: Record<string, string>
+  toolCallTimeoutMs?: number  // default 60_000
+}
+
+type Config = StdioConfig | StreamableHttpConfig
+```
+
+`serverName` is the stable local identity that namespaces this server's tools in the model-facing name (below). It is deliberately user configuration, NOT the remote `serverInfo.name`: the remote name is untrusted input, is not unique across deployments (prod and staging instances of one server report the same name), and may change on server upgrade — none of which may silently rename model-facing tools. A duplicate `serverName` across live instances is a configuration error: the later instance fails at load with an actionable message, never silent shadowing or skipping. A short `serverName` (`gh`) is also the knob for shortening public names.
+
+Example `cordis.yml` usage:
+
+```yaml
+- id: mcp-github
+  name: '@deepseek-ai/dsh-mcp-client'
+  config:
+    serverName: github
+    transport: stdio
+    command: npx
+    args: ['-y', '@modelcontextprotocol/server-github']
+    env:
+      GITHUB_TOKEN: !!js process.env.GITHUB_TOKEN
+
+- id: mcp-web
+  name: '@deepseek-ai/dsh-mcp-client'
+  config:
+    serverName: web
+    transport: streamable-http
+    url: http://localhost:3000/mcp
+    headers:
+      Authorization: !!js `Bearer ${process.env.MCP_TOKEN}`
+```
+
+The model sees `mcp__github__create_issue`, `mcp__github__search_code`, `mcp__web__search`.
+
+### Lifecycle
+
+Boot-time from `cordis.yml`. HMR (`@cordisjs/plugin-hmr`) provides hot-swap: editing the yml entry triggers dispose of the old instance (disconnects, unregisters tools) and creation of a new one (connects, discovers, registers). No runtime-dynamic API for now. Public names are pure functions of `(serverName, rawName)`, so an HMR swap that keeps `serverName` recreates identical model-facing names — session history and permission rules stay valid — and adding or removing an unrelated server never renames an existing tool.
+
+### Tool discovery and registration
+
+Every MCP tool has two names:
+
+- `rawName` — the exact MCP `Tool.name`, used only on the wire (`tools/call`).
+- `publicName` — the globally unique model-facing name registered in the `ToolRegistry`:
+
+      mcp__<serverName>__<rawName>
+
+This server-qualified shape is the de-facto standard among multi-server agent clients — every surveyed end-user product qualifies MCP tools by server ([Claude Code](https://code.claude.com/docs/en/agent-sdk/mcp#tool-naming-convention) `mcp__github__list_issues`, [Codex](https://openai.com/index/unrolling-the-codex-agent-loop/) `mcp__weather__get-forecast`, [Gemini CLI](https://geminicli.com/docs/tools/mcp-server/#3-tool-naming-and-namespaces), [VS Code](https://github.com/microsoft/vscode/blob/ab9ec62c6a61e429a9abd612ff220c3f4834c9ea/src/vs/workbench/contrib/mcp/common/mcpServer.ts#L217-L260), [Cline](https://github.com/cline/cline/blob/52fdbb1d72f7324a28142a7ba7678d4b53c902f4/sdk/packages/core/src/extensions/mcp/name-transform.ts#L20-L35), [Roo Code](https://github.com/RooCodeInc/Roo-Code/blob/b867ec9145750d0ae1ff7f02d35406e9bf2a0b16/src/utils/mcp-name.ts#L117-L140), [Goose](https://github.com/block/goose/blob/b3a012cbdde854b0fe14f95b1c48543bf6517c0a/crates/goose/src/agents/extension_manager.rs#L1391-L1441), [OpenCode](https://github.com/anomalyco/opencode/blob/d199b1bff90282a4f9cd6251b5fc7b16875a52f6/packages/opencode/src/mcp/catalog.ts#L117-L120)); the exact `mcp__<server>__<tool>` spelling follows Claude Code and Codex. The `mcp__` marker keeps MCP registrations out of the native tools' namespace and gives permission/telemetry rules a stable shape (`mcp__*`, `mcp__github__*`).
+
+1. On connect: drain `client.listTools()` pagination, derive every tool's `publicName`, then register each as a raw `ToolDefinition` via `ctx.tools.register()`. The MCP JSON Schema and description pass through unchanged (no `defineTool` DSL conversion); only the model-facing `name` is replaced.
+2. Listen for `notifications/tools/list_changed` → re-run the same sync (dispose previous generation, register new). Deterministic names mean unchanged tools keep their names across re-syncs.
+3. The executor closes over `rawName`; the public name is never sent to the server and never parsed to recover the raw name.
+4. No `presentCall`/`presentResult` — the ACP bridge's generic-card fallback handles rendering.
+5. Tools are transparent in the system prompt — no "[via MCP]" annotation beyond the name itself.
+
+### Public name normalization
+
+MCP allows tool names up to 128 characters including `.`; the DeepSeek function-name contract allows `[A-Za-z0-9_-]` and at most 64. Public names are normalized deterministically: invalid characters become `_`, and when replacement or truncation changed the name, a 12-hex-char SHA-256 hash of the `(serverName, rawName)` identity is appended so distinct MCP identities can never collapse into the same public name:
+
+```typescript
+function publicToolName(serverName: string, rawName: string): string {
+  const joined = `mcp__${serverName}__${rawName}`
+  const normalized = joined.replace(/[^A-Za-z0-9_-]/g, '_')
+  if (normalized === joined && normalized.length <= 64) return normalized
+  const hash = sha256(`${serverName}\0${rawName}`).slice(0, 12)
+  return `${normalized.slice(0, 64 - 13)}_${hash}`
+}
+```
+
+### Name conflict handling
+
+MCP guarantees tool-name uniqueness only [within one server](https://modelcontextprotocol.io/specification/2025-11-25/server/tools#tool-names); cross-server collisions are the norm, not the exception (a [Microsoft Research survey](https://www.microsoft.com/en-us/research/blog/tool-space-interference-in-the-mcp-era-designing-for-agent-compatibility-at-scale/#namespacing-issues-and-naming-ambiguity) of 1,470 servers found 775 colliding tool names; `search` alone appears in 32 servers, and the official GitHub server publishes bare `create_issue`). The always-on namespace makes collisions structurally impossible instead of handling them at collision time:
+
+- Two servers publishing `search` coexist as `mcp__github__search` and `mcp__web__search`.
+- A native harness tool named `search` is unaffected.
+- Duplicate `serverName` config fails the later instance at load (see Configuration).
+- A server listing the same tool name twice is an invalid tool list: the sync throws and the previous generation stays registered.
+- A registry conflict during the swap can only mean a foreign tool squats on this server's `mcp__<serverName>__` namespace: the partial generation is rolled back (zero tools from this server) and the error is logged loudly.
+
+Tools are never silently skipped; which tools are available never depends on plugin load order.
+
+### Naming invariants
+
+1. Every MCP tool has the stable identity `(serverName, rawName)`; every active identity has exactly one public name.
+2. Public names are deterministic, globally unique, and satisfy the DeepSeek 64-char `[A-Za-z0-9_-]` contract.
+3. MCP `tools/call` always receives the original raw name.
+4. Connecting, disconnecting, or re-syncing an unrelated server never renames an existing tool.
+5. Registration order never determines which tool is available.
+
+### Tool execution
+
+A unified `execute` handler for all tools from one MCP server:
+
+1. Resolve `rawName` (the executor closes over it) and call `client.callTool({ name: rawName, arguments }, { signal: exec.signal })` with the configured timeout — the public name is never sent to the server.
+2. Map the result:
+   - Multiple `text` content blocks → join with `'\n'` into a single `TextBlock` (required: `flattenText` uses `join('')` without separator, so multiple blocks would lose inter-block boundaries).
+   - `image` content blocks → discard with a `ctx.logger.warn` (the harness has no image content block type; [drop-image Agent Note](../simplification/2026-07-04-drop-image-content-block.md)).
+   - `isError: true` → map to the harness `isError` result path (`{ content: [...], isError: true }`).
+3. Cancellation: `exec.signal` (from the agent loop's cancel) is passed through to the MCP SDK's `callTool`, which sends `$/cancelRequest` to the server.
+
+### Subprocess environment (stdio transport)
+
+Replicate the `buildChildEnv` + `SENSITIVE_ENV_PATTERN` scrub from `dsh-subagent-acp`: filter ambient env (strip credential-shaped vars matching `/KEY|SECRET|TOKEN/i`), then merge `config.env` on top. Explicit env overrides survive the scrub.
+
+### Disconnection / crash
+
+No auto-reconnect. If the MCP server process exits or the transport closes:
+
+1. The effect disposes → all registered tools are unregistered (fiber-scoped disposers).
+2. Subsequent model calls to those tools → `ToolNotFoundError` → `isError: true`.
+3. Recovery: user edits `cordis.yml` (triggers HMR reload) or restarts the harness.
+
+This matches the ACP subagent pattern: "crash = terminal, report error, clean up, don't retry."
+
+## Alternatives considered
+
+### MCP Server side (expose harness tools to external MCP clients)
+
+Deferred. The ACP bridge already exposes the harness as an agent server. Adding an MCP server layer would duplicate that with a different protocol, and the primary user need is consuming external tools, not exposing them.
+
+### Capability-seam three-package split (interface / impl / consumer)
+
+Rejected. There is no foreseeable alternative MCP client implementation — MCP has one protocol, one SDK. The convention is "don't split preemptively" until a second implementation appears.
+
+### Auto-reconnect with exponential backoff
+
+Rejected for v1. Adds complexity (partial-availability state where tools are registered but temporarily non-functional), and stdio process crashes usually indicate a configuration problem that retrying won't fix. HMR already provides the manual recovery path. Can be added as a future `reconnect: boolean` config if needed.
+
+### Bridge Resources and Prompts
+
+Deferred. Resources need a harness-side mechanism to decide WHEN to inject content (system prompt? on demand? model-triggered?). Prompts need a "prompt template" concept the harness lacks. Both require their own design; Tools are the high-value, low-risk starting point.
+
+### Raw model-facing tool names with an optional `toolPrefix`
+
+Rejected — this was the original proposal, built on the premise that "most MCP servers already use semantic prefixes in their tool names (e.g. `github_create_issue`)". The premise is false: the official GitHub server publishes `create_issue`, the reference filesystem server `read_file`, Sentry `search_issues` — and the Microsoft survey above shows collisions are common at ecosystem scale. Collision-time prefixing (or warn-and-skip) also makes the available tool set depend on plugin load order, and a tool could be silently renamed when an unrelated server is added — invalidating session history and permission rules mid-conversation. No surveyed multi-server agent product ships raw names.
+
+### Server-only namespace (`github__create_issue`, no `mcp__` marker)
+
+Rejected for v1. It prevents cross-server collisions but does not separate MCP registrations from native harness tools, and it forfeits MCP-wide policy shapes (`mcp__*`). The marker costs 5 characters; the `mcp__<server>__<tool>` spelling matches Claude Code and Codex, maximizing model familiarity. If the ToolRegistry later grows source-aware namespaces, dropping the literal marker can be revisited as a naming-policy change.
+
+### Deriving the namespace from the server-announced `serverInfo.name`
+
+Rejected. The remote name is untrusted, non-unique across deployments, and changeable on upgrade; tool identity and permission rules must not silently follow it. The namespace is local configuration.
+
+### Preserve multiple TextBlocks in tool result
+
+Rejected. `flattenText()` in the DeepSeek serializer uses `join('')` (no separator) when flattening `ContentBlock[]` to wire format. Multiple text blocks would silently lose inter-block boundaries — a correctness bug. All existing tools return a single TextBlock; the MCP bridge follows suit.
+
+## Testing
+
+Coverage is named per tier; each behavior lives at the cheapest tier that can express it.
+
+- **Unit** (`tests/mcp-client.spec.ts`, `tests/apply.spec.ts`, mocked MCP SDK): the `publicToolName` algorithm (clean, normalize, truncate-and-hash, determinism, distinct-identity separation), raw-vs-public wire discipline, cross-server and native-tool coexistence, duplicate-`serverName` load failure and reservation release, invalid-tool-list rejection, generation swap/rollback, failed-re-sync retention, result mapping, cancellation, config schema validation. 100% per-file coverage gates the package.
+- **E2E** (`tests/mcp-client.e2e.ts`, keyless): the real MCP protocol against the in-repo fixture server, `@modelcontextprotocol/server-everything`, and `@modelcontextprotocol/server-filesystem` over stdio, and against an in-process `StreamableHTTPServerTransport` server over Streamable HTTP — discovery under the namespace, dotted-name normalization end to end, execution round-trips, duplicate-`serverName` rejection, disposal.
+- **Snapshot**: deliberately none. MCP tools introduce no new transcript surface — they register as raw `ToolDefinition`s and render through the ACP bridge's generic-card fallback, which the bridge's unit suite already pins (`packages/ui/acp/tests/stream-update.spec.ts`). Adding an MCP server to the snapshot example's `cordis.yml` would mutate the pinned `text-turn` system-prompt fixture (forcing a with-key re-record of every recorded expected output) and make every replay depend on spawning an external MCP server process — for zero new rendering behavior. If a later change gives MCP tools their own render intent, that change names its snapshot coverage then.
+
+## Consequences
+
+- A `cordis.yml` entry per MCP server is the entire integration cost: `serverName: filesystem` + a stdio command (or a Streamable HTTP URL) puts `mcp__filesystem__read_file` in the model's tool list, callable, with the raw `read_file` on the wire.
+- Public names are part of session history and permission/config surfaces; the naming algorithm is a v1 contract pinned by tests, and changing it after release is a breaking change.
+- The `mcp__<serverName>__` qualifier costs tokens on every name. Accepted: descriptions and JSON schemas dominate tool-definition tokens, and the qualifier buys stable identity, collision isolation, and MCP-wide policy shapes (`mcp__*`, `mcp__github__*`).
+- **MCP SDK stability**: the `@modelcontextprotocol/sdk` is still evolving; breaking changes require updating the bridge. The version is pinned, and the SDK is widely adopted (Claude Desktop, Cursor, VS Code) so breaking changes are unlikely to be silent.
+- **Tool schema quality**: MCP servers may expose poorly-described tools (vague descriptions, incomplete JSON schemas). The harness passes them through as-is — garbage-in-garbage-out; that is the server author's responsibility, not the bridge's.
+- **Stdio process management**: a misbehaving MCP server that ignores signals could wedge dispose. The Cordis fiber disposal has bounded quiescence; a stuck transport eventually times out at the framework level.
+- Crash recovery is manual (HMR edit or restart) — accepted for v1; a `reconnect` config remains open as future work.

.agents/notes/implemented/feature/2026-07-07-session-prefix.md

@@ -0,0 +1,41 @@
+# Agent Note: The session prefix — request-only messages in front of the derived history
+
+Status: implemented
+
+## Problem
+
+A plugin often owns a session-stable opener the model must always see — a skills catalog, an AGENTS.md digest, a workspace baseline. Before this seam the harness offered two homes, and both are wrong for that content. The system prompt is one rendered string: message-shaped content (a user-role `<system-reminder>` envelope, a multi-message primer) does not fit it, and providers weight conversation messages differently from system text. Durable history (`agent.inject()`, a `context/message` at session start) makes the opener permanent: every `deriveMessages()` consumer replays it, the compaction retention walk owns it, forks bake it in stale, and a resume cannot refresh it — a catalog captured at session birth outlives the world it described.
+
+The obvious third option — let a plugin edit the request's `messages` on the way out — is banned by [the reconstructable-requests Agent Note](../architecture/2026-07-05-reconstructable-requests.md): every loop-built request is a pure function of the session log, so whatever channel carries the opener must log exactly what it sends. What was missing was a request-only message channel with a durable record.
+
+## Decision
+
+`agent/session-prefix` is a waterfall on the agent event map ([`packages/core/agent/src/types.ts`](../../../../packages/core/agent/src/types.ts)): listeners receive a frozen empty seed and return an extension (the canonical contribution is a prepend, `[mine, ...await next()]`, which yields registration order on the wire). The loop ([`packages/core/agent-loop/src/loop.ts`](../../../../packages/core/agent-loop/src/loop.ts)) fires it once per loop instance, lazily before the instance's first `agent/pre-step`; the composed list is deep-cloned, deep-frozen, cached on the instance, and placed in front of the ENTIRE derived history — directly after the provider's system slot — on every request the instance sends ([wire order](../../../../docs/core-data-structures/core.md#the-request-envelope-llmcallconfig-and-the-logged-header)).
+
+Three properties carry the design:
+
+- **Request-only, header-logged.** `deriveMessages()` never returns the prefix; its one durable record is `EpochHeader.messagePrefix` on the instance's anchoring `request/header` snapshot — the channel the reconstructable-requests Agent Note already owns for the request's non-history half, so no new session event exists. The dev invariant ([dsh-invariants](../../../../packages/support/invariants/src/index.ts)) recomputes `messagePrefix + boundary derivation` against every loop-built request; an unlogged prefix cannot reach the wire.
+- **Frozen per instance.** Reuse is structural, not disciplined: the cached product cannot change mid-session, so the provider's prompt cache holds by construction and the prefix extends the cacheable region at zero marginal cost per step. A process restart or `ctx.agents.resume()` is a new instance: it recomposes, and any drift lands attributably on the `'resume'` header snapshot. This is the routing rule the seam creates: session-frozen openers ride the prefix; content that changes mid-session rides the append-only history channels (`agent.inject()` or tool/prompt-submit `additionalContexts` — [the interception-seams Agent Note](2026-06-30-interception-seams.md)), each a durable `context/message` paid once and prefix-cached thereafter.
+- **Exact in the durable request envelope.** Composition precedes the instance's first `agent/pre-step` and request boundary. The first routed request logs the current prefix on its header, so post-step token pressure reads the exact prefix together with the actual prompt, tools, and routed model; no compaction-only parameter is carried through the generic pre-step seam. A composition interrupted by cancel/dispose is discarded, never cached: an abort-aware listener's degraded fallback cannot leak into later requests, and the next turn recomposes under a live signal.
+
+Because composition runs before the boundary snapshot, a composing listener's session append joins the CURRENT request's derived history. Compaction structurally cannot touch the prefix (or the system prompt): it rewrites surface nodes, and header state never enters the surface.
+
+## Testing
+
+[Interception tests](../../../../packages/core/agent-loop/tests/interception.spec.ts) pin compose-once reuse without changed headers, prepend order, empty-prefix omission, immutability, composition before pre-step, and the prefix on the routed header; [cancellation tests](../../../../packages/core/agent-loop/tests/cancel.spec.ts) pin discard and recomposition. Session, invariant, token-meter, and compaction tests cover header round trips, request reconstruction, and durable prefix-aware pressure accounting. Snapshot normalization preserves prefix counts, while the [pinned-header scenario](../testing/2026-07-06-pin-request-header-content-in-one-scenario.md) owns content and the default example remains prefix-free. The provider-independent seam needs no dedicated e2e; the with-key [request-cache e2e](../../../../packages/core/agent-loop/tests/request-cache.e2e.ts) covers its cache economics.
+
+## Alternatives considered
+
+- **Per-request `before`/`after` slots recomputed every step** (the shape first proposed: a waterfall firing on every request, contributing frozen `before` messages ahead of the history and fresh `after` messages behind it) — rejected. A per-step `before` recompose invites drift that must be logged as a full changed header, and an `after` slot sits behind the growing history, so its tokens re-pay on every request and everything after it is uncacheable. Measured against the alternatives, every current update pattern is served cheaper by a durable append (paid once, cache-read thereafter), and the only content with no home was the session-stable opener — which wants freezing, not recomputation.
+- **A system-prompt section** (`system-prompt/assemble`) — rejected for this content: the assembly renders to the single `system` string, so message-shaped openers do not fit, and the system prompt is deliberately re-assembled per step (with a full changed header when it changes) while the opener wants instance-frozen semantics.
+- **A durable history opener** (`inject()` at session start) — rejected: permanent history is the failure mode in the problem statement — replayed everywhere, compactable, stale across resumes.
+- **Compose per turn instead of per instance** — rejected: a turn-boundary recompose either desyncs silently from the log or forces a changed header, and it busts the provider cache exactly as often as it fires; the legitimate refresh point is the instance boundary, where the `'resume'` snapshot already records drift attributably.
+- **Carry prompt/prefix through `agent/pre-step` for provisional pressure** — rejected because it couples a generic lifecycle seam to one consumer and still misses later request routing and tools; post-step replay reads every request-envelope field from its durable routed header.
+- **A dedicated session event carrying the prefix** — rejected: the header events are the request's non-history record by design; a second event would be a second home for the same fact and another codec to keep total.
+
+## Consequences
+
+- `agent/pre-step` stays a generic `(agent, turn, step, signal)` checkpoint. Compaction receives no prefix parameter; `ctx.tokenMeter` folds the prefix from the canonical routed header at post-step.
+- A contributor whose content changes mid-session is not re-read until the next instance — by design. A deployment needing mid-session catalog updates routes the change notice through the append-only history channels and pays one durable `context/message`.
+- The dropped `after` slot leaves no request-only channel near the request tail; nothing in the repo needs one, and adding it back would re-open the every-step re-pay cost the design exists to avoid.
+- An empty composition is canonical absence: no-contributor deployments log no extra header bytes and their requests are the bare derivation.

.agents/notes/implemented/feature/2026-07-08-background-subagent-tasks.md

@@ -0,0 +1,62 @@
+# Agent Note: Background subagent tasks
+
+Status: implemented
+
+## Problem
+
+The [subagent seam](2026-06-21-subagent-capability-seam.md) returns a `SubagentRun`, but the model-facing tool originally collected every run synchronously. Independent, slow delegations therefore held the parent call open or ran serially.
+
+Subagents need the same start, collect, list, stop, ownership, notification, and cleanup behavior as other long-running tools without adopting process-stream semantics. The child session remains the detailed trace; the parent needs the final answer and task status. A background child also outlives its starting tool call, so its cancellation and owner-disposal contracts must be explicit.
+
+## Decision
+
+Each `dsh-tool-subagent` instance may expose `run_in_background`, controlled by `enableRunInBackground` and enabled by default. A disabled instance omits the parameter and rejects a forced background argument at execution. Provider selection remains deployment configuration, so one instance still registers one distinctly named tool for one provider.
+
+Background subagents use the [generic background task runtime](../architecture/2026-06-20-generic-long-running-tool-runtime.md). Collection, listing, cancellation, completion notices, and prompt guidance come from `task_output`, `task_list`, and `task_kill`; there are no subagent-specific companion tools.
+
+Foreground calls retain their synchronous contract: await provider startup and `run.result`, return final text only for `completed`, map other terminal reasons to an errored tool result, and always dispose the run before returning.
+
+For a background call, the tool validates the parent and refuses an already-aborted execution signal before calling `ctx.tasks.start()`. The task runtime preflights the control surface and owner cleanup before invoking the producer starter. That starter creates an independent `AbortController` and begins `ctx.subagents.start()`; after the id is returned, the tool-call signal no longer owns the child.
+
+The task registration maps the subagent seam as follows:
+
+- `kind` is `subagent`, `label` is the model-supplied description, and `owner` is the parent agent.
+- `cancel(reason?)` aborts the task-owned controller. The same signal covers pending provider startup and the ready child.
+- `done` awaits provider startup, the child result, and `run.dispose()`. Completed runs return final text, aborted runs become `killed`, and other stop reasons become `failed`. Startup, result, and disposal failures become failed outcomes rather than rejected task promises.
+- `readOutput` is absent. While live, `task_output` returns status only; after settlement, it returns final output idempotently. Intermediate child activity remains in the child session.
+
+## Lifecycle
+
+A background subagent belongs to its parent agent and is not durable across owner closure. The task runtime attaches cleanup to the exact owner's scope. Agent disposal cancels the task and awaits startup rollback or child disposal before `AgentHandle.dispose()` resolves, preventing leaked child agents and sessions.
+
+Completion notices target the exact owner captured at start. If owner teardown has already disposed the injection target, the notice is dropped; cleanup, not notification, is the lifecycle guarantee.
+
+## Model guidance
+
+The generic task prompt teaches the shared habit: retain ids, continue independent work instead of busy-polling, collect relevant tasks before answering, and kill irrelevant work. The subagent schema adds only that background mode returns a task id and that `task_output` collects the result. Authorization and owner cleanup enforce the runtime boundary independently of prompt compliance.
+
+## Alternatives considered
+
+### Subagent-specific wait, output, and stop tools
+
+Capability-specific tools would duplicate the task protocol, teach another collect-and-stop habit, and complicate multiple provider instances. The generic runtime provides the required behavior without changing the tool's one-provider-per-instance shape.
+
+### Survival after owner closure
+
+Survival requires persistent task state, child-session recovery, a late-result delivery channel, and policy for abandoned owners. Owner-scoped cleanup gives process-local work a clear lifetime. Durable jobs require a separate design.
+
+### No owner checks for isolated clients
+
+Agents and logs may be session-scoped, but the task registry and predictable ids are runtime-global. The generic owner fence therefore applies to subagents like every other producer.
+
+### Incremental child transcript output
+
+Streaming child history into the parent would blur the log boundary and make provider behavior diverge. This surface exposes final output only; richer observation belongs to session or UI tooling.
+
+## Testing
+
+Unit coverage pins stop-reason mapping, dispose-before-report behavior, startup and result failures, pre-aborted refusal, detachment from the starting call's signal, cancellation before and after provider readiness, collection through the real task tools, the no-surface preflight fence, missing-runtime failure, and per-instance schema gating. Snapshot coverage pins the model-facing schemas.
+
+## Consequences
+
+The parent can fan out slow delegations and collect them through the same task controls used by bash. Child work no longer occupies the starting tool call, but it can consume resources until collected, killed, or owner-disposed. Prompt guidance encourages collection; owner cleanup provides the hard lifetime boundary. Deployments that require synchronous delegation can disable background mode per tool instance.

.agents/notes/implemented/feature/2026-07-08-repeat-tool-guard.md

@@ -0,0 +1,74 @@
+# Agent Note: Repeat-tool-call guard plugin
+
+Status: implemented
+
+## Problem
+
+A model stuck in a loop re-issues the same tool call with byte-identical arguments — re-running a failing grep, re-reading an unchanged file, polling a command that already gave its answer — and each round trip burns tokens, wall-clock, and (for paid APIs) money without adding information. The harness has nothing that notices: the loop has no step budget, no plugin tracks call repetition, and the model only escapes when it happens to vary its own behavior. The failure mode is real and cheap to detect — [pi-repeat-tool-guard](https://github.com/Kingwl/pi-repeat-tool-guard) ships exactly this as a pi coding-agent extension: count consecutive identical calls and, past a threshold, append a `<system-reminder>` telling the model to stop repeating itself and change course.
+
+The harness already has every seam the pi extension uses, and better ones: [the interception-seams Agent Note](2026-06-30-interception-seams.md) gives `tools/post-execute` a sanctioned way to attach model-facing context to a finished call, the loop buffers and injects that context with call/result adjacency preserved, and injected context is a logged `context/message` — so a native guard satisfies the model-visible ⟺ logged rule with no new session event. What was missing was only the plugin itself.
+
+## Decision
+
+The guard is a loop-hygiene plugin, not a model-facing tool. It counts consecutive calls to the same tool with identical canonical arguments and injects advisory reminders at configured thresholds. It never delays, blocks, or rewrites a call; the model decides whether to retry differently or finish.
+
+The plugin is `@deepseek-ai/dsh-repeat-tool-guard` at `packages/guard/repeat-tool-guard/`, opening the `guard/` group for loop-hygiene plugins (single-package groups have precedent: [the todo-write Agent Note](2026-06-29-todo-write-tool.md) shipped `todo/tool-todo`). It registers two listeners and holds state in a `WeakMap` keyed by the live `Agent` object — the tool registry is a context-level singleton whose waterfalls interleave every agent's calls (subagents run on the same context), so per-agent keying is correctness, not polish; weak object keys also make a disposal-only cleanup listener unnecessary.
+
+- **`tools/post-execute` (waterfall)** — the one detection point. The listener receives `(exec, result)` together, so counting and reminder delivery need no cross-event pending map (the pi extension needs one only because its `tool_call`/`tool_result` hooks are separate events). It always delegates via `next()` and, when a threshold is hit, prepends a reminder to the downstream decision's `additionalContexts` — the observe-and-enrich posture [the hooks bridges](2026-06-30-hook-bridges.md) already use, honoring the waterfall contract. Counting happens here rather than in `tools/pre-execute` because post-execute also runs for denied calls (`ToolRegistry.execute` routes a deny through the same pipeline), and a model hammering a denied call is exactly the loop worth breaking.
+- **`agent/prompt-submit` (waterfall)** — pure reset hook: delegate via `next()`, clear the submitting agent's chain. A user interjection changes the context; repetition across it is not a loop.
+
+### Detection semantics
+
+The chain key is `(tool name, canonical arguments)`; a call identical to the previous tracked call increments the agent's consecutive counter, a different tracked call resets it to 1. Canonicalization is a deep key-sort plus `JSON.stringify`: `ToolExecution.arguments` is by construction the loop's `JSON.parse` output (or the raw string fallback for malformed argument JSON, which is itself a comparable value), so the pi original's bigint/circular/`undefined` handling has no inputs here and is deliberately dropped.
+
+Two deliberate rules, both documented in [the package README](../../../../packages/guard/repeat-tool-guard/README.md) because they are behavior a reader would otherwise guess at:
+
+- **Untracked calls are transparent to the chain.** A call excluded by `include`/`exclude` neither increments nor resets the counter, so `grep X → todo_write → grep X` still counts as two consecutive `grep X` when `todo_write` is excluded. This is what makes exclusion useful — bookkeeping tools interleaved into a loop must not launder it — and it is the pi extension's (undocumented) semantics, kept on purpose and written down.
+- **Calls without an agent are ignored.** A direct `ctx.tools.execute()` caller (tests, non-loop consumers) has no model to remind and no live agent object to key on.
+
+### Reminder delivery
+
+Reminders ride `additionalContexts` as their own entries (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}` — the label is load-bearing per `HookContext`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit, and the loop appends buffered contexts as `context/message`s after the step's results, which the session renders as tagged synthetic-user envelopes and derived history replays. Thresholds escalate: the first configured threshold gets a short "you are repeating yourself, analyze the previous result" nudge; each later threshold gets the detailed form naming the tool, the repeat count, and the canonical arguments (head-truncated at `argumentsPreviewChars`, default 500 — a looping `write`-sized payload must not ride into the next request unbounded; the chain key always compares the full canonical string), and stating that the calls made no progress. The pi original hardcodes the gentle text to the literal count 3; the guard keys it to `thresholds[0]`, fixing that bug in the port. A downstream hook bridge contribution remains a separate array entry, so both plugins retain their source, envelope, and metadata.
+
+### Config
+
+```yaml
+- id: repeat-tool-guard
+  name: '@deepseek-ai/dsh-repeat-tool-guard'
+  config:
+    thresholds: [3, 5, 8]        # default; consecutive counts that trigger a reminder
+    include: []                  # tool-name patterns to track; empty ⇒ all tools
+    exclude: [todo_write]        # tool-name patterns transparent to the chain
+    argumentsPreviewChars: 500   # default; cap on arguments quoted in the detailed reminder
+```
+
+`thresholds` is validated at load and throws on an empty list, a non-integer, a value below 2, or a duplicate — misconfiguration fails loud, replacing the pi original's silent fall-back to defaults. `include`/`exclude` entries support `*` wildcards. Patterns are predicates over whatever tools exist at call time, not references to a registry entry, so an entry matching no currently registered tool is NOT an error — unlike `toolOrder`'s referent check, `exclude: [mcp_*]` must stay valid in a deployment that loads no MCP tools.
+
+## Testing
+
+- **Unit:** A real loop with a scripted adapter covers counting and reset rules, untracked transparency, disposal cleanup, per-agent isolation, canonical argument key order, escalation, denied calls, no-agent execution, wildcard escaping, invalid config, and downstream block or replacement decisions at per-file 100% coverage.
+- **Snapshot:** The keyless `repeat-tool-guard` scenario makes five identical `todo_write` calls and pins the gentle third-call and detailed fifth-call reminders in both ACP output and the session log. The plugin is loaded in the live example but remains inert in other scenarios.
+- **E2e:** None; the plugin is deterministic and provider-independent, and its seam contracts are covered by their owners.
+
+## Alternatives considered
+
+- **Append the reminder into the tool result** (`accept` with replaced `content` — the pi extension's mechanism, which patches result content because that is the only channel its API offers) — rejected: it makes the logged `tool/result` lie about what the tool returned, and `additionalContexts` is the separate sanctioned channel for post-execute commentary, with loop-level buffering that preserves call/result adjacency.
+- **Count in `tools/pre-execute` with a pending-reminder map** (the pi two-phase shape) — rejected: post-execute alone sees `(exec, result)` together and also fires for denied calls, so one listener with no cross-event state covers strictly more attempts with less machinery.
+- **Escalate to `block` at the highest threshold** — rejected for the initial scope: a blocked call punishes legitimate identical repeats (polling a long-running terminal, re-checking a file the agent expects to change), and an advisory reminder keeps the model in control. Revisit with evidence; the decision shape (`PostToolDecision`) already supports it.
+- **A per-deployment external hook via the CC/Codex bridges** (a `PostToolUse` script) — rejected as the answer: it works for one deployment, but a shipped, unit-tested, `cordis.yml`-configurable plugin is the harness-native form, without per-call subprocess cost.
+- **A loop-level step or repetition budget in `agent-loop`** — rejected: "plugins, not loop changes"; a hard step budget is a blunter, orthogonal control that would need its own proposal.
+- **Fuzzy/near-identical detection** (normalized paths, similar-but-not-equal arguments) — rejected: exact match after canonicalization is cheap, deterministic, and explainable to the model; similarity thresholds invite false positives and need evidence before they earn complexity.
+- **Placing the package in `core/`** — rejected: core is the product spine; a behavioral guard is an optional leaf plugin, and the `todo/` precedent is a small dedicated group per plugin family.
+
+## Consequences
+
+- The reminder is advisory by design: idempotent polling patterns that repeat identical calls on purpose still receive nudges past the thresholds, and the pressure valves are config (`thresholds`, `exclude`) plus reminder text that explicitly allows finishing when enough evidence has been gathered. Each trigger costs reminder tokens on the next request; thresholds bound the frequency.
+- Chain state is in-memory only: a session resumed from persistence starts with a fresh chain, so a loop spanning a resume draws its reminders later than a live one — accepted, the guard is a heuristic nudge, not a logged invariant, and persisting counter state would buy little for real complexity.
+- When multiple post-execute producers attach context on one call, each contribution stays a separate `HookContext`; ordering follows waterfall nesting and each entry retains its own provenance.
+- Implementing the snapshot tier surfaced a hidden assumption in the suite kit: the fixture guard equated "authored model scenario" with "override-driven". The `Scenario` table now carries an explicit `overridden` flag, and the sidecar's presence is checked BOTH ways against it (an unregistered stray sidecar would silently replace the derived script) — the suite kit is stricter than it was before this plugin existed.
+
+## Deferred
+
+- Compaction does not reset chains: a compacted history changes what the model sees, but the repetition risk usually survives compaction.
+- Escalating to `block` at a high threshold is not implemented; `PostToolDecision` already supports it if evidence arrives.
+- Subagent chains stay isolated per agent; no sharing mechanism exists until a concrete case appears.

.agents/notes/implemented/feature/2026-07-08-self-referential-cordis-toolset.md

@@ -0,0 +1,80 @@
+# Agent Note: The self-referential cordis toolset
+
+Status: implemented
+
+## Problem
+
+Everything in this harness is a cordis plugin, but the agent running inside that plugin runtime cannot see or touch it: it cannot enumerate the services and events around it, cannot extend itself with a new tool mid-session, and cannot compose capabilities it invents. Handing the model that power is worth exploring — a self-referential agent that inspects and modifies its own runtime — but it raises three correctness problems at once, and the design is about answering them rather than the raw "let the model run code" mechanic.
+
+First, model-written registration must be validated where it happens: a malformed tool schema has to fail at registration, not when a later request tries to assemble it into a prompt. Second, model-written code has to call service APIs whose source it has never seen — guessed method signatures and, worse, guessed return-value shapes cost many steps of blind probing. Third, everything the model mounts must be fully disposable, by the model on demand and by the ordinary plugin lifecycle when the host plugin reloads, or a long session accretes orphaned listeners and tools.
+
+## Decision
+
+The toolset ships as [`@deepseek-ai/dsh-tool-cordis`](../../../../packages/cordis/tool-cordis/README.md) — a new top-level `packages/cordis/` group — and is demoed by [`examples/cordis-agent`](../../../../examples/cordis-agent/README.md). It gives the model three tools over the live cordis runtime it is running inside: inspect it, mount model-written plugins into it, dispose them again.
+
+The vm isolates accidental global pollution, and the context façade hides framework internals. Neither restricts the authority of exposed services: a mount can call `ctx.bash` to run commands with the host executor's privileges and can reach the real filesystem and web services. This is an opt-in development tool with bash-equivalent trust, not a security boundary or product default.
+
+### The three tools
+
+| Tool | Contract |
+|---|---|
+| `cordis_inspect` | Read-only report over the live runtime, one Markdown section per `what` value (omit `what` for all sections). An exact `name` with `what: "api"` or `what: "events"` narrows to one source-documented target. Never mutates. |
+| `cordis_mount` | Evaluates `code` (the body of an async JavaScript function) in a `node:vm` sandbox; the code must `return` a cordis plugin, which is mounted as a child of the `cordis-dynamic` group fiber and tracked under a fresh id (`dyn-1`, `dyn-2`, …). |
+| `cordis_unmount` | Disposes one dynamic mount by id and returns only after disposal reaches quiescence — every registration the plugin made is unwound, not merely requested to stop. |
+
+`cordis_inspect` sections: `services` (every provided ctx service and the owning fiber, non-active owners flagged), `plugins` (a flat list of every loaded plugin with its lifecycle state, from `ctx.registry` — what capabilities are loaded, deliberately not the tree shape), `tools` (what the model can call), `dynamic` (the mount table: id, name, state, provided services, awaited services), `api` (live service signatures + the type shapes they reference, from the generated catalog), and `events` (harness events with dispatch mode and signature). Broad `api` and `events` reports omit full JSDoc to stay compact; an exact `name` returns one service or event with its original method/declaration JSDoc. A name is invalid with other sections, unknown targets fail, and an API target must be live. The model-facing tool descriptions carry the operational rules the model needs at call time; [the generated tool catalog](../../../../docs/tool-catalog.md) is their exhaustive rendering.
+
+### Sandbox semantics
+
+Mount code runs as an async-function body in a fresh vm realm. Its documented surface steers file, network, process, and timer access through Cordis services so mounts remain inspectable and disposable. Host-realm helpers still make Node escape possible, consistent with the trusted posture. `vmTimeoutMs` bounds only synchronous evaluation.
+
+Sandbox globals are deliberately small: a tagged write-through `console` (`[cordis:<id>] …` on the host stdout/stderr, so a listener that fires long after the mount call still lands somewhere the user sees), the `harness.defineTool` / `harness.registerTool` registration pair, the encoding primitives fresh vm contexts lack (`btoa`/`atob` as host closures over `Buffer` — a sanctioned exception, `Buffer` itself is never exposed — plus `TextEncoder`/`TextDecoder`), and callable traps over the withheld Node APIs (`require`, `setTimeout`/`setInterval`/`setImmediate`/`clearTimeout`/`clearInterval`, `fetch`) that throw a redirect naming the cordis alternative. Only function-shaped globals are trapped; `process` and `Buffer` stay `undefined` so a `typeof` feature probe stays inert rather than detonating a throwing accessor.
+
+Mount code crosses the vm boundary through three controls. Dual-realm `instanceof` recognizes both host and vm objects. `harness.defineTool` normalizes results into host-realm JSON and validates the `ToolExecuteReturn` shape before logging. The mounted plugin receives a whitelist context façade, not a raw or pass-through `Context`; framework plumbing and context-valued returns are rejected. Service reads require a declared `inject`, preserving Cordis activation and unload semantics. `ctx.tools.get` exposes only the schema view, so mounted code cannot bypass `ToolRegistry.execute` by calling a definition directly.
+
+The boundary normalizes unambiguous JSON-Schema forms into `SchemaSpec`, including object wrappers, `integer`, and optional fields. Invalid vocabulary fails with the accepted alternatives. Parse, TypeScript, missing-return, Node-API, and duplicate-tool errors include the relevant source line or corrective contract without narrating implementation internals.
+
+### The dynamic group and mount lifecycle
+
+All dynamic mounts are children of one `cordis-dynamic` group beneath the tool plugin, so ordinary fiber disposal handles reload and unload. Mounting awaits settlement; startup failure disposes the fiber before returning an error. A settled pending mount remains visible with its missing injections. `cordis_unmount` awaits the mount fiber's disposal.
+
+### Cross-mount composition via provide/inject
+
+Mounts relate to each other through ordinary cordis service semantics, with their ids as the lifecycle handles: mount A calls `ctx.provide('foo', value)`, mount B declares `inject: ['foo']` and activates the moment `foo` exists; mounted first, B stays pending and names the missing service; unmounting A sends B back to pending (its registrations unwound) and a later re-provide re-runs B's `apply` through a fresh sandbox façade; a duplicate provide fails loud with the owning fiber named. One realm caveat: a service value provided by a mount is a vm-realm object — method calls on it work from anywhere, but consumers must not assume host prototypes on it.
+
+### The generated API catalog
+
+`cordis_inspect` serves API and event data from a generated catalog rather than a duplicated table. The generator reuses the Cordis catalog AST scan and emits service summaries, signatures, original service-method and event JSDoc, event modes, referenced type declarations, and the inherited context surface. Ambiguous type names are omitted and oversized declarations are marked as truncated.
+
+Freshness is gated like every generated artifact: `pnpm run verify-cordis-api` (in `doc-sync`) regenerates in memory and fails on any diff, so a JSDoc or public-signature edit cannot ship without regenerating the catalog the model reads. At runtime the inspect tool intersects the catalog with the live runtime rather than dumping it: broad reports render live catalogued services as summary + signatures, live services without a catalog entry (mount-provided ones) as name + owning fiber, catalogued services with no live provider tersely, and then the referenced type shapes. Exact-name reports render one live service or event with the original JSDoc immediately before each signature; keeping that detail opt-in avoids charging its token cost on exploratory listings.
+
+### Configuration, rendering, and observability
+
+The plugin exposes one config field, validated by schemastery and documented in [the config catalog](../../../../docs/config-catalog.md): `vmTimeoutMs` (default 5000), the millisecond bound on the synchronous portion of mount-code evaluation. Tool names, the `cordis-dynamic` group name, and the `dyn-` id prefix are structural vocabulary and stay fixed. All three tools render as `generic` cards per [the tool cookbook](../../../../docs/cookbook/adding-a-tool.md) (`cordis_inspect` a `read`, `cordis_mount` an `execute` carrying the code as `rawInput`, `cordis_unmount` a `delete`), with no `presentResult` overrides.
+
+Model-visible ⟺ logged holds with no new session event type: a mount or unmount is visible only through its own `tool/call` / `tool/result` pair, which the loop logs, and the changed tool set a mount induces is logged by the full changed request header the loop emits when schemas change between steps. There is deliberately no `cordis/mount` provenance event — it would duplicate what the tool-call pair records. Dynamic mounts are process-lifetime, not session state: resuming a persisted session rehydrates the conversation but does not re-mount plugins.
+
+## Alternatives considered
+
+**A structured per-capability registration tool instead of `cordis_mount`.** The most tempting alternative is a `cordis_register_tool` with explicit `name` / `description` / `parameters` / `code` fields (and siblings `cordis_register_listener`, `cordis_register_service`, …) rather than a single "mount a plugin" primitive. It was rejected because its one real win — no plugin boilerplate for the single commonest case — does not pay for its costs, while a single mount primitive answers every capability at once.
+
+| Dimension | Structured per-capability tools | Single `cordis_mount` |
+|---|---|---|
+| Schema correctness | `parameters` is still a model-written JSON object needing SchemaSpec validation, merely one step earlier | The same validation runs at the sandbox boundary, with the same instructive errors |
+| The code field | An `execute` body is still model-written JS in a vm; the realm and service-call correctness problems are unchanged | One sandbox, one normalization path, one guarded registration |
+| Capability coverage | Tools only; listeners, services, `inject` relations each need another structured tool — a surface that grows without bound | One vocabulary (a cordis plugin) covers every effect, present and future |
+| Cross-mount composition | Not expressible in a tool-registration payload | Native `provide`/`inject`, ordinary cordis semantics |
+| Inspectability | Registers something the plugin list cannot show as a plugin | What the model mounts is exactly what `cordis_inspect` renders |
+| Model ergonomics | Wins for the single most common case (no plugin boilerplate) | Mitigated by the canonical recipe in the mount description plus boundary errors that teach the fix |
+
+The correctness investment therefore goes where it pays for every capability at once: the generated API catalog surfaced through `cordis_inspect`, and sandbox-boundary validation whose error messages teach the correct call. A structured registration tool remains addable later as sugar that synthesizes mount code; nothing here forecloses it.
+
+**A hand-maintained service/event reference in the tool.** The first cut of the inspect tool carried a hand-written table of service method signatures. It was replaced by the generated `api-catalog.ts` because a hand table drifts from the JSDoc the moment a signature changes and nothing gates the drift, whereas the generated artifact is freshness-checked against the same AST the docs use.
+
+**A new `cordis/mount` session event.** A durable provenance event recording each mount (source, name) has clear precedent (`hook/invoked`, `compact/start`). It was declined for v1: mount and unmount are already visible as `tool/call` / `tool/result` pairs and the tool-set change is already logged as a full changed request header, so a dedicated event would only duplicate the record. It remains addable if an audit use case needs mount provenance separable from the tool call.
+
+**A hardened / capability-restricted sandbox.** Trapping Node built-ins and handing mount code a whitelist façade rather than the raw context might suggest an intent to sandbox for safety. It is explicitly not that: the traps and the façade narrow the *surface* mount code sees — steering it onto cordis services and away from leak-prone Node built-ins and framework internals — for correctness and to close the unguarded-context escape, but the capabilities the façade exposes (`ctx.bash`, `ctx.fs`, `ctx.web`) reach the real runtime, so it is not a security boundary. A real one (separate process, permission prompts) was out of scope for a dev/opt-in toolset and would fight the entire point — handing the model the live runtime.
+
+## Consequences
+
+The toolset is a deliberate opt-in with a fully-privileged `ctx`, so a deployment adopts it as consciously as a bash tool. Several facts follow that the tool descriptions warn the model about directly: a waterfall listener (e.g. `tools/pre-execute`) that returns without calling `next()` vetoes the chain, so a mounted listener can lobotomize the agent's own tool dispatch ([waterfall semantics](../../../../docs/cordis-primer.md#cordis-waterfall-semantics)); mount code runs inside a tool call of the current turn, so awaiting anything that resolves only after the turn deadlocks; `vmTimeoutMs` bounds synchronous evaluation only; and mounts do not survive session resume.

.agents/notes/implemented/feature/2026-07-09-bash-backed-grep-glob-discovery.md

@@ -0,0 +1,166 @@
+# Agent Note: Bash-backed grep and glob discovery tools
+
+Status: implemented
+
+## Problem
+
+The harness needs model-facing `glob` and `grep` tools, but making them `ctx.fs` provider methods turns a local product convenience into a universal filesystem backend contract. Local workspace discovery is naturally a process-backed `rg` workflow; remote or virtual filesystem backends may expose their own search API, may not share a local `ripgrep` view, or may not support discovery at all. The v1 should not require every filesystem backend to implement search before the file read/write/edit seam has proven that need.
+
+Search output also has two distinct budgets. The tool needs enough raw `rg` output to compute a stable logical result, but the model should receive only a bounded preview plus a recovery path when the formatted result is larger than the inline budget. The generic spill policy only sees the final tool result, so it cannot recover matches that a search tool already omitted. Search therefore needs tool-owned retention and best-effort formatted-result spill.
+
+## Decision
+
+`glob` and `grep` are model-facing tools in `@deepseek-ai/dsh-tool-fs-search`, backed by the bash seam, not by new `ctx.fs` provider methods. The package registers model-facing filesystem discovery tools, but execution uses `ctx.bash.resolve(request)` followed by `ctx.bash.run(spec)` with fixed `rg` command templates assembled by the tool. The tool layer owns schemas, argument validation, shell quoting, result parsing, result formatting, retention, formatted-result spill handoff, and timeout declaration. The bash executor owns request defaulting/capping, subprocess execution, process-group termination, environment scrubbing, raw output capture, and backend substitution across local, sandboxed, or remote bash implementations.
+
+The tools do not use `ctx.bash.start()` and do not create model-visible background tasks. They run as ordinary foreground tools from the agent loop's perspective: the tool call returns only after the `rg` command exits, times out, is aborted, or fails. `defineTool({ timeoutMs })` declares the cooperative tool-call budget, `@deepseek-ai/dsh-timeout-policy` enforces it through `exec.signal`, and the tool forwards that signal into the bash request before `resolve()` / `run()`. The bash backend's own timeout remains a second safety cap; whichever aborts first wins.
+
+The tools align `path` with Claude Code's search tools while binding resolution to the bash workdir, not to `ctx.fs`. The tool derives the bash request workdir from `exec.agent?.session.header.cwd`, mirroring `dsh-tool-bash` and `dsh-tool-fs`; when no session cwd exists, it omits `request.workdir` so the bash implementation applies its configured cwd or process cwd through `resolve()`. For `grep`, `path` is an optional ripgrep target and may be a file or directory; omitted means the resolved bash workdir. For `glob`, `path` is an optional directory search root; omitted means the resolved bash workdir. Relative `path` values resolve against that workdir. Returned paths are displayed relative to the resolved bash workdir when possible and are intended to be follow-up-readable only in co-located deployments where the bash workdir and filesystem `read` root are the same workspace. v1 documents that deployment requirement but does not perform runtime cross-service validation. Remote or virtual filesystem search is deferred until there is a shared workspace/root contract or a provider-specific search backend.
+
+The package does not inject `fs`. It injects `tools`, `systemPrompt`, and `bash`; it deliberately reads `spillStore` with `ctx.get('spillStore')` instead of static inject because formatted-result spill is optional. Existing `@deepseek-ai/dsh-tool-fs` deployments that only want `read` / `write` / `edit` do not need to load bash.
+
+### Package shape
+
+The v1 package stays small. Inside `@deepseek-ai/dsh-tool-fs-search`, the source layout is:
+
+```text
+src/index.ts
+src/glob.ts
+src/grep.ts
+src/search-core.ts
+src/shell-quote.ts
+```
+
+`glob.ts` and `grep.ts` own their parameter validation, command construction, result parsing, formatting, and registration. `shell-quote.ts` is one shared helper because shell quoting is the safety boundary both tools must use; `search-core.ts` is the other (an implementation-time amendment to the original four-file plan): the `SEARCH_*` error vocabulary, the bash-run + raw-output acquisition, the formatted-spill handoff, and workdir-relative display are byte-identical between the two tools, and duplicating that delicate plumbing per tool is exactly the missed extraction the symmetry convention flags. Command builders must not hand-roll quoting or concatenate unquoted model-controlled values into the shell command.
+
+### Schemas and config
+
+`glob` exposes the small discovery shape:
+
+```ts
+interface GlobArgs {
+  pattern: string
+  path?: string
+}
+```
+
+`grep` exposes the OpenCode-style minimal shape:
+
+```ts
+interface GrepArgs {
+  pattern: string
+  path?: string
+  include?: string
+}
+```
+
+Routine budgets stay out of the model-facing schema. `@deepseek-ai/dsh-tool-fs-search` owns these defaulted, validated config fields:
+
+| Field | Default | Role |
+|---|---:|---|
+| `globMaxResults` | `100` | Max paths retained inline; matches Claude Code's default `GlobTool` result limit. |
+| `grepMaxMatches` | `250` | Max flat matches retained inline; matches Claude Code's default `GrepTool` `head_limit`. |
+| `grepMaxLineBytes` | `2000` | Max bytes retained for one matched-line preview, applied with `TextRetainer({ kind: 'head', maxBytes: grepMaxLineBytes })`. |
+| `rawOutputMaxBytes` | `20000000` | Max complete raw `rg` stdout the tool will parse; matches Claude Code's ripgrep raw buffer. |
+| `timeoutMs` | `30000` | Tool-call timeout attached to both tool definitions and enforced by `@deepseek-ai/dsh-timeout-policy`. |
+
+`globMaxResults` and `grepMaxMatches` use `ItemRetainer({ kind: 'head' })`. `grepMaxLineBytes` uses `TextRetainer({ kind: 'head', maxBytes: grepMaxLineBytes })` for each matched line so preview cuts preserve UTF-8 boundaries. This follows the [tool result retention library](../architecture/2026-07-06-tool-result-retention-library.md) mapping for discovery items: collect the complete result, retain head items inline, and keep path mapping, grouping, and per-line preview outside the retainer. `grep` does not expose `case_insensitive`, `head_limit`, `offset`, `count`, multiline, context lines, output modes, or file type filters in v1. A model that needs surrounding context reads the matched file with `read`; a model that needs later results follows the returned spill locator's retrieval hint.
+
+The Claude Code values are reference points for the two-layer budget, not model-facing schema precedent. Its dedicated search tools buffer raw ripgrep output up to 20 MB for internal processing, use a 20-second ripgrep timeout on non-WSL platforms (60 seconds on WSL), then apply search-specific caps before the model sees a result: `GrepTool` defaults to `head_limit = 250` and persists formatted results above 20,000 characters, while `GlobTool` defaults to 100 paths and persists formatted results above 100,000 characters. This Agent Note mirrors the raw-buffer and inline-count defaults, chooses a 30-second default search timeout, and uses this harness's `ctx.spillStore.saveText()` path for formatted-result recovery.
+
+The `path` field follows the same split as Claude Code: `grep.path` is a file-or-directory ripgrep target, while `glob.path` is a directory search root. v1 does not expose a separate cwd/workdir argument on these tools.
+
+`include` is one positive glob filter, not a list and not an exclude syntax. Reject comma-separated or negated include patterns up front with a structured argument error. Every model-controlled value used in a shell command, including `pattern`, `path`, and `include`, must pass through the package-private shell quoting helper.
+
+### Execution
+
+`glob` builds a fixed `rg --files` command rooted at the resolved directory search root (`path` when supplied, else the bash workdir): `rg --files --glob <pattern> --sort=modified --no-ignore --hidden`, plus VCS metadata excludes for `.git`, `.svn`, `.hg`, `.bzr`, `.jj`, and `.sl`. This aligns with Claude Code on hidden/ignored-file discovery and modified-time ordering while keeping VCS internals out of broad searches. The tool parses one path per line, maps results back to paths relative to the bash workdir when possible, pushes each path into `ItemRetainer({ kind: 'head', maxItems: globMaxResults })`, and formats the full sorted path list for a spill artifact when the retained result is capped.
+
+`grep` builds a fixed line-oriented `rg --json` command against the supplied file/directory target (`path` when supplied, else the bash workdir) so file path, line number, and line text are parsed without colon-splitting ambiguity. It consumes `match` records, treats malformed JSON or malformed match records as `SEARCH_FAILED`, maps result paths relative to the bash workdir when possible, applies per-line preview retention with `grepMaxLineBytes`, pushes each match into `ItemRetainer({ kind: 'head', maxItems: grepMaxMatches })`, then groups only the retained preview matches by file for inline output. The spill artifact stores the full formatted match list, not only the omitted tail, so the retrieval hint points at the same logical result the model saw.
+
+Raw `rg` stdout is an internal transport detail. The tool requests `stdoutMaxBytes: rawOutputMaxBytes` through `ctx.bash.resolve()` and parses `stdout.text` only when the executor returns untruncated stdout within that cap. If stdout is larger than `rawOutputMaxBytes`, or the executor still returns `stdout.truncated`, the tool fails with a clear search error telling the model to narrow `pattern`, `path`, or `include`. The tool never exposes raw `rg` output or bash raw spill paths to the model.
+
+Only stdout is a parse source. Stderr is diagnostic text for invalid patterns, missing `rg`, and search failures; if bash truncates stderr, the tool uses the retained stderr tail with a truncation note and does not read `stderr.spillPath`.
+
+If `ctx.bash.run()` reports `aborted` because the tool timeout or caller cancellation fired, the tool returns a structured failure rather than pretending there were no matches. If bash reports its own timeout first, the tool likewise fails with a clear timeout message. Nonzero ripgrep exit semantics are tool-owned: exit 0 is success with matches, exit 1 is success with no matches, invalid pattern / missing `rg` / inaccessible search workdir are failures.
+
+Search failures use a package-owned `HarnessError` subclass with `SEARCH_*` codes, not `FsErrorCode`, because these tools are not `ctx.fs` provider operations. The v1 vocabulary is `SEARCH_INVALID_PATTERN`, `SEARCH_FAILED`, `SEARCH_RAW_OUTPUT_OVERFLOW`, and `SEARCH_ABORTED`. Model argument validation failures such as missing required fields, blank strings, or unsupported negated/list `include` values remain ordinary tool argument errors.
+
+### Formatted result spill
+
+`ctx.spillStore` is optional and used only for model-facing formatted results. This is the first tool-owned spill call pattern in the codebase, and it is intentional because search retention is item-level policy: `globMaxResults` caps paths and `grepMaxMatches` caps matches while the tool still holds the complete logical result. The generic `dsh-spill-policy` caps final text bytes on `tools/post-execute`; by then a search tool would already have omitted later paths or matches, so the policy cannot recover them.
+
+When a search produces more logical results than the inline cap and `ctx.spillStore` is present, the tool saves the complete formatted result with `saveText()`. The spill owner is the calling agent's session header id (`exec.agent?.session.header.id`); without that owner, the search keeps the inline result and reports that the complete result could not be saved. The spill source is the tool execution identity: `{ toolName: exec.name, callId: exec.callId, label: 'result' }`. The suggested filenames are `grep-results.txt` and `glob-results.txt`; the spill backend still treats them as hints, never paths.
+
+When spill storage is absent, the call has no session owner, or saving fails, the tool still returns the inline page and a footer explaining that the complete result could not be saved. Search success must not turn into an `isError` result solely because formatted-result spill storage is unavailable.
+
+The bash raw output stream and the formatted search spill artifact are different artifacts. Raw `rg` stdout is parsed only in memory within the requested bash stdout cap; the formatted spill artifact is the stable model-facing recovery locator produced by `ctx.spillStore.saveText()`.
+
+### Result shape
+
+A capped `glob` result with successful formatted spill returns the inline page and a spill notice:
+
+```text
+<first N paths>
+
+(Showing N of M paths. Full sorted result stored at: /.../session-abc123/9f8e7d-glob-results.txt. Use read with offset/limit, or grep this path to search within it.)
+```
+
+A capped `grep` result with successful formatted spill returns grouped preview matches and a spill notice:
+
+```text
+Found N of M matches
+
+<file>
+Line 12: ...
+
+(Full grep result stored at: /.../session-abc123/9f8e7d-grep-results.txt. Use read with offset/limit, or grep this path to search within it.)
+```
+
+If the complete logical result fits under the inline cap, no formatted spill artifact is created. If the complete logical result is too large but formatted spill is unavailable, the footer says that the result was capped and the complete result could not be saved. The `truncated` / omitted count is a budget fact, not an incomplete-search fact; timeout, invalid regex, missing `rg`, inaccessible workdirs, raw-output overflow, binary skips, and parse failures stay in tool-domain error or incomplete fields.
+
+## Alternatives considered
+
+**Put `glob` / `grep` on `ctx.fs`.** Rejected for v1: it forces every filesystem backend to grow a search API and makes local ripgrep behavior part of the provider seam. Search is useful product behavior, but it is not a universal text-storage primitive like `readText` or `writeText`.
+
+**Directly spawn ripgrep from `dsh-fs-local`.** Rejected for this Agent Note's v1: direct spawn gives the cleanest argv boundary, stdout/stderr control, and early-stop control, but it duplicates process execution concerns that the bash seam already owns: environment scrubbing, process-group kill, timeout propagation, sandbox/remote executor substitution, and bounded output capture. It remains a reasonable optimization if bash-backed search proves too shell-string-sensitive or if foreground streaming becomes necessary.
+
+**Use `ctx.bash.start()` for streaming early stop.** Rejected: `start()` creates model-visible background task semantics: task ids, owner tokens, `bash_output`, `bash_kill`, completion notifications, and no built-in timeout. `grep` needs a foreground tool result, not a background bash workflow. If streaming search becomes necessary, the right abstraction is a foreground streaming process handle on the bash/process seam, not borrowing the public background-task API.
+
+**Expose bash raw spill paths to the model.** Rejected: a bash raw spill path contains raw `rg` stdout (`rg --json` records for grep), not the stable formatted search result. Search parses raw stdout only as an internal transport; model recovery uses a formatted result saved through `ctx.spillStore.saveText()`.
+
+**Add `spillStore.saveFile()` for bash output normalization first.** Rejected for this Agent Note's v1: `saveFile()` would help a future bash normalization pass move existing executor spill files into session-scoped spill storage, but search only needs bounded in-memory raw `rg` stdout before producing the model-facing artifact. `saveText()` is sufficient for the formatted search result.
+
+**Rely on the generic `dsh-spill-policy`.** Rejected: generic post-execute spill sees only the final tool result. If `grep` / `glob` return the first page inline, the generic policy cannot recover omitted results. The search tools must save the complete formatted result themselves before returning the bounded model-facing text.
+
+**Expose Claude Code's full `GrepTool` schema.** Rejected for v1: `output_mode`, context flags, multiline, `head_limit`, `offset`, `case_insensitive`, and type filters make the model-facing surface into a ripgrep wrapper. This harness keeps routine budgets and continuation mechanics in deployment policy and spill artifacts.
+
+**Keep early-stop search and skip formatted spill artifacts.** Rejected for this proposal: early stop is more efficient but gives the model no path to inspect later results. The chosen v1 optimizes result recoverability and implementation simplicity, with `timeoutMs`, `rawOutputMaxBytes`, bash backend caps, and formatted spill artifacts as safety backstops.
+
+**Expand the bash seam with a raw-output reader first.** Rejected: a portable `readRawOutput(ref, maxBytes)` API would add reference lifetime, permission, and backend storage semantics. A per-run `stdoutMaxBytes` request is the narrower seam: search either receives complete stdout within `rawOutputMaxBytes` or fails clearly.
+
+## Testing
+
+- Tests prove an aborted `exec.signal` reaches the bash backend (same-reference spec assertion plus the `SEARCH_ABORTED` result), and cover command construction/quoting (malicious patterns, paths with spaces, leading-dash values, quotes, newlines, glob metacharacters — unit assertions plus a real `bash -c` round-trip for every hostile value), `grep.path` as file and directory targets, `glob.path` as a directory search root, invalid pattern handling, no matches, malformed `rg --json` output, matched-line preview truncation, raw-output overflow, timeout/abort, formatted spill success/failure, the package-owned `SEARCH_*` error codes, and the no-background-task invariant.
+- The first-party tool-owned spill precedent is covered directly: spill backend present, spill backend absent, `saveText()` failure, and missing spill owner.
+- The package has real Loader-path coverage for the namespace plugin export shape (`name`, `inject`, `Config`, and `apply`, with no default export).
+- A real-executor integration suite (`dsh-bash-local` + a real `rg`) verifies the world: hostile patterns stay inert, per-session cwd resolution, VCS-metadata exclusion, modification-time ordering, and real ripgrep stderr classification. It self-skips where `rg` is not on PATH (a CI accommodation mirroring the keyless e2e skip); the fake-executor suite alone carries the per-file 100% coverage gate.
+- Snapshot gap note for the transcript-visible spill notice: this landed with the gap note, not a snapshot. The snapshot tier replays the acp-agent tree, and adding the search plugin there changes the assembled system prompt — every expected output would need re-recording with a real key, which the implementing environment did not hold. The spill notice's exact transcript text is pinned by unit tests (`formatGlobOutput`/`formatGrepOutput` and the through-the-registry spill tests); wiring the plugin into the acp-agent tree plus a `test:snapshot:record` pass is the follow-up for the next key-holding session.
+
+## Consequences
+
+- `glob` and `grep` are model-facing tools in `@deepseek-ai/dsh-tool-fs-search`, not `ctx.fs` provider methods and not part of the existing `@deepseek-ai/dsh-tool-fs` root plugin. The package injects `tools`, `systemPrompt`, and `bash`; it does not inject `fs`, and `ctx.spillStore` stays optional via `ctx.get('spillStore')`.
+- The schemas are exactly `glob(pattern, path?)` and `grep(pattern, path?, include?)`; search caps and timeout are defaulted, validated Config fields (`globMaxResults`, `grepMaxMatches`, `grepMaxLineBytes`, `rawOutputMaxBytes`, `timeoutMs`).
+- The tools execute through `ctx.bash.resolve(request)` → `ctx.bash.run(spec)`, forward `exec.signal`, never call `ctx.bash.start()`, and never expose a bash task id. The bash request workdir comes from `exec.agent?.session.header.cwd` when available; the resolved `spec.workdir` drives execution and relative-path display.
+- The tools request `stdoutMaxBytes: rawOutputMaxBytes` from the bash seam, parse only untruncated stdout within that cap, and treat over-cap or still-truncated raw output as a clear search failure; raw `rg` output is never exposed to the model.
+- Oversized complete formatted results are saved through `ctx.spillStore.saveText()` when available while inline results stay bounded; spill failure, a missing backend, or a missing owner preserves the inline result and reports the unsaved remainder — never an `isError`.
+- The package README, the generated config catalog, and exported JSDoc document the Config fields and `SEARCH_*` codes; the repl-agent example ships the tools (the acp-agent tree waits on the snapshot re-record above); the fs group README records the co-located bash/filesystem deployment requirement.
+
+## Risks
+
+Full-run `grep` can be slower than an early-stop search on broad patterns. The v1 accepts that cost for simpler implementation and complete-result recovery, bounded by tool timeout, bash timeout, `rawOutputMaxBytes`, and output caps. If this proves too slow, the direct-ripgrep or foreground-streaming alternatives remain available.
+
+Shell command construction is the sharpest safety edge. Because `ctx.bash` accepts a command string rather than an argv vector, the implementation must centralize shell quoting and test malicious patterns, paths with spaces, leading-dash patterns, quotes, newlines, and glob metacharacters.
+
+The v1 assumes a co-located bash/filesystem deployment. If bash searches one workspace and the `read` tool resolves paths against another, returned paths may not be follow-up-readable. The package documents this requirement but does not verify it at runtime.
+
+Spill locators are backend-owned. The current local backend returns local filesystem paths and works in deployments where `read`/`grep` can open those files; remote or workspace-confined deployments can use a backend whose locator and retrieval hint point at a supported retrieval mechanism.

.agents/notes/implemented/feature/2026-07-10-agent-session-identity-and-log-location.md

@@ -0,0 +1,85 @@
+# Agent Note: Expose agent session identity and JSONL location to tools and hooks
+
+Status: implemented
+
+## Problem
+
+An agent can identify its workspace through `session.header.cwd`, but a model using bash cannot reliably identify the session that owns the call or the durable transcript that records it. Searching `./.sessions` guesses deployment config and JSONL layout; custom roots, alternate persistence backends, resume, forks, and concurrent parent/child agents make that guess unreliable. Hooks have the same need for transcript location, while future plugins may need to expose other harness-owned environment facts to shell commands.
+
+The boundary must preserve two properties: the owner of a fact decides how to resolve it, and every child receives a per-execution snapshot rather than process-global mutable state. In particular, a nested harness must not leak its ambient `DSH_*` values into a child whose current agent, persistence backend, or configuration differs.
+
+## Decision
+
+Extend the [`SessionPersistence`](../architecture/2026-06-14-session-persistence.md) seam with a synchronous, side-effect-free location query:
+
+```ts
+import type { SessionHeader } from '@deepseek-ai/dsh-session'
+
+interface SessionLocation {
+  readonly kind: string
+  readonly path: string
+}
+
+interface SessionPersistence {
+  locate(meta: SessionHeader): SessionLocation | undefined
+}
+```
+
+`path` is an absolute local path to the backend's dedicated log for `meta`; `kind` identifies the representation. JSONL returns `{ kind: 'jsonl', path }` using its resolved root and path helpers. SQLite and any backend without an honest local per-session artifact return `undefined`. The query creates and flushes nothing, so it can report a lazy target path before that file exists.
+
+The model-facing bash package owns a `ctx.bashEnv` registry. A contributor declares its stable name, every `DSH_*` key it may return, a description for each key, and `resolve(execution: ToolExecution)`. Duplicate contributor names, duplicate key ownership, reserved keys, malformed declarations, undeclared runtime output, and non-string output fail loudly. Registration is a Cordis effect and is removed with the contributing plugin fiber. `list()` exposes declarations without running resolvers, keeping the environment surface enumerable for diagnostics and future prompt/UI consumers.
+
+The registry rebuilds a trusted overlay for every foreground and background bash `ToolExecution`:
+
+- `DSH_HOME` is always the absolute configured Harness home. The standalone [`@deepseek-ai/dsh-home`](../../../../packages/util/home/README.md) utility owns its precedence: explicit `dshHome`, then ambient `$DSH_HOME`, then `~/.dsh`.
+- `DSH_SHELL=1` is always present and identifies a model bash child managed by DeepSeek Harness.
+- `DSH_SESSION_ID` is present when the execution has an agent and equals `agent.session.header.id`.
+- The built-in persistence translator contributes `DSH_SESSION_JSONL` only when `ctx.sessionPersistence.locate(header)` returns `kind: 'jsonl'`.
+
+Session persistence remains the fact owner: JSONL does not depend on tool-bash or register shell variables itself, and hooks continue to consume `locate()` directly. Tool-bash is the translation layer from the persistence fact into a shell convention. Other plugins that need shell-visible facts depend on the registry and register their own keys; they do not modify `process.env`.
+
+The bash seam exports `DSH_ENV_PREFIX` as the single namespace source and derives `DshEnvironmentKey` from its `typeof`. Tool-bash derives built-in names and model guidance from that constant, while executors use it for filtering and channel validation. The seam carries the managed overlay separately as `BashExecRequest.dshEnv` / `BashExecSpec.dshEnv`. Ordinary `env` remains the general in-process plugin surface used by hooks, but cannot contain managed keys; symmetrically, `dshEnv` cannot contain ordinary keys. The local executor rejects either wrong channel before spawn, removes every inherited ambient managed key, applies its ordinary scrub/terminal environment/explicit `env`, and finally merges the trusted `dshEnv` snapshot. This guarantees that a missing value means absent now rather than inherited from an outer or previous harness. The model-facing tool still ignores model-supplied `env`/`stdin` arguments.
+
+The bash tool description teaches only the durable convention: current harness environment facts are available through managed `$DSH_*` variables and may be inspected when needed. It does not enumerate persistence-specific keys or add a permanent system-prompt section. Tool schemas are already logged in request headers and tool output is logged as `tool/result`, so no new session event is required.
+
+The [Claude Code and Codex hook bridges](2026-06-30-hook-bridges.md) resolve transcript location from the same persistence seam when constructing payloads. Codex uses `transcript_path: string | null`; Claude Code preserves its string field and falls back to `''`. Hook lookup neither materializes nor flushes a session.
+
+## Peer product findings
+
+Peer products separate stable identity from physical storage. Codex injects stable `CODEX_THREAD_ID` into spawned shells while recorder and hook surfaces own transcript paths. Claude Code supplies `session_id` and `transcript_path` as structured hook/status input. OpenCode carries identity in structured tool context; Kimi Code expands a session placeholder; Reasonix keeps the active session path on its controller. The portable rule is to inject identity at the invocation boundary, let storage resolve location, and never use a process-global current-session variable in a concurrent harness.
+
+## Lifecycle and persistence semantics
+
+A fresh session receives its id before the first turn, so its first bash call can read `DSH_SESSION_ID` and a JSONL target. The JSONL file may still be absent until the first successful turn-end checkpoint, and during an open turn it contains only the last flushed prefix. `DSH_SESSION_JSONL` is a location hint, not an authorization credential or freshness guarantee.
+
+Resume reuses the loaded header and therefore the same id and location. Fork and spawn create new session ids and locations. Parent and child calls resolve from their own `ToolExecution.agent`; each command receives an immutable snapshot even when calls overlap. A persistence service replacement affects later collections because the translator queries `ctx.get('sessionPersistence')` at execution time; the registry itself is effect-scoped and HMR-safe.
+
+`dshHome` is session-independent deployment context. Agent-core resolves one value through `@deepseek-ai/dsh-home` and routes it to both tool-bash and local skill discovery; standalone consumers call the same resolver. If top-level `dshHome` and `skills.local.dshHome` are both supplied and resolve differently, composition fails instead of exposing contradictory homes. Persistence may change independently without freezing its facts into the session prefix.
+
+## Testing
+
+Unit coverage pins registry declaration validation, effect disposal, per-execution collection, the `dshHome` precedence, and the local executor's `DSH_*` scrub/rebuild order. Request-recording tests cover foreground/background snapshots, no-agent calls, absent/JSONL persistence, ignored model `env`, and parent/child isolation. JSONL/SQLite locator contract tests and both hook bridge suites pin available and unavailable transcript dialects.
+
+A keyless full-loop integration drives the real agent loop, JSONL persistence, tool-bash, and bash-local on the first turn. The child prints `DSH_HOME`, `DSH_SHELL`, session id, JSONL target, and an inherited stale sentinel; the test verifies current values, absence of the stale variable, pre-flush file absence, and the eventual persisted header. Snapshot coverage pins the generic bash description in the recorded request header. No with-key test is required because the contract is deterministic local execution rather than model choice.
+
+## Alternatives considered
+
+**Only an id plus `find`.** Search cannot know a custom root or backend layout and races under multiple sessions.
+
+**Only an absolute path.** A path can be unavailable, lazy, or representation-specific and is not stable session identity.
+
+**Global `process.env`.** Concurrent agents would overwrite one another and nested harnesses would inherit stale current-session values.
+
+**Put persistence instructions in the session prefix.** A session prefix is frozen while the active service can change across HMR or future backend switching; persistence-specific guidance would become stale.
+
+**A typed waterfall event.** Listeners cannot declare ownership without running, and later listeners can silently overwrite keys. A registry detects key conflicts at registration and remains enumerable.
+
+**Have each persistence backend register bash env directly.** That reverses the dependency from storage into one consumer and forces bash into deployments that do not use it. `locate()` is also still required by hooks.
+
+**A model-facing `session_info` tool.** It adds schema and another call while bash already supplies the query surface; the registry generalizes to future environment facts without one tool per fact.
+
+## Consequences
+
+Every model bash child receives current Harness home and shell identity, and agent calls additionally receive stable session identity. JSONL-backed calls get an optional target path; non-file persistence omits it honestly. The complete `DSH_*` namespace inside these children is managed by the harness: ambient values are removed, current trusted values are re-added, and ordinary callers cannot use `env` to bypass ownership checks.
+
+The namespace is discoverable but not secret. Paths can reveal configured roots, lazy targets can be absent or stale, and a command can override variables inside its own shell syntax. Consumers treat them as correlation and environment facts, verify transcript metadata when attribution matters, and rely on sandbox/filesystem policy rather than variable secrecy for authorization.

.agents/notes/implemented/feature/2026-07-10-parallel-tool-call-execution.md

@@ -0,0 +1,101 @@
+# Agent Note: Parallel tool-call execution by per-call safety
+
+Status: implemented
+
+## Problem
+
+An assistant message may contain several sibling `tool-call` blocks. Running them serially adds the latency of independent reads and web requests even though the model has already requested them together.
+
+Concurrency is a host scheduling concern, not model-facing tool metadata. The loop needs to decide which calls may overlap without hardcoding tool names or exposing scheduler policy in the JSON schema.
+
+The session log remains authoritative: every started call has an audit event, every started call receives a result, and model history observes results in the original call order regardless of completion order.
+
+## Decision
+
+Each tool may provide an optional `isConcurrencySafe(args)` classifier. It is synchronous and pure: it examines only the current call's parsed arguments and performs no I/O or mutation. Only an explicit `true` opts in; a missing classifier, invalid arguments, a thrown classifier, or any other return value makes the call exclusive. The canonical type contract lives in the [tool data structures](../../../../docs/core-data-structures/tools.md).
+
+The classifier is deliberately unary. Returning `true` is the tool's promise that this call may overlap with any sibling call that also returns `true`; the scheduler does not compare calls or prove that their resource accesses are compatible.
+
+The unary classifier remains input-sensitive. A tool may classify a read-only operation as parallel and a mutating operation as exclusive. The interface cannot express relational rules such as "these writes are safe only when their paths differ," so a call whose safety depends on a sibling remains exclusive.
+
+`defineTool()` validates arguments before invoking a typed classifier. Invalid arguments classify as exclusive and produce the ordinary argument error only if the call executes. `ctx.tools.executionMode(exec)` resolves the live tool definition and returns the tagged `parallel` or `exclusive` mode; unknown tools fail closed to exclusive.
+
+A tagged mode, rather than a public boolean scheduler API, keeps resource-aware variants representable without changing the classifier contract.
+
+## Scheduling and ordering
+
+The loop waits for the complete assistant message, parses every call once, creates a distinct `ToolExecution` for each call, and scans them in model order. Consecutive parallel calls form one group; every exclusive call forms a singleton group and an ordering barrier. Groups execute sequentially. Classification is lazy: the scheduler resolves the next call after each barrier and reclassifies every later call before replenishing a parallel pool. If a registry mutation makes that call exclusive, the current pool drains before the call starts as the next barrier.
+
+For example:
+
+```text
+[parallel read(A), parallel read(B), exclusive write(A), parallel read(C)]
+
+→ [read(A), read(B)]
+→ [write(A)]
+→ [read(C)]
+```
+
+`read(A)` and `read(B)` may overlap. `write(A)` starts after both finish, and `read(C)` starts after the write finishes.
+
+Every group uses a rolling pool bounded by `maxParallelToolCalls`: the loop starts calls in model order up to the cap and starts another whenever one settles. An exclusive group is a pool of one. A cap of `1` preserves serial execution.
+
+Only dispatch and the tool body overlap. `tools/pre-execute` and `tools/post-execute` run in model order because middleware may maintain ordering-sensitive state. `tools/execute` wrappers run around concurrent dispatches and therefore must be reentrant across distinct executions.
+
+Each started call appends `tool/call` immediately before its pre-execute gate. Completed dispatches occupy model-order slots, and a commit cursor appends `tool/result` and collects `additionalContexts` only when the next slot is ready. Live surfaces may show several pending calls, but results and post-tool context remain model-ordered.
+
+An abort before a group starts records no calls from that group. An abort during a group stops replenishment, waits for already-started calls, commits their results in order, drains accepted batch context after those results, and then ends the step through the existing abort path. Calls that never start have no audit event.
+
+Code Mode remains outside this scheduler because the model emits one native `run_code` call. `run_code` and its internal dispatch queue remain serial; native sibling calls in `mode: 'both'` use the normal scheduler.
+
+## Safety contract
+
+A tool that returns `true` promises that its body is safe to run at the same time as other parallel calls. It must not directly mutate the parent session or other parent-owned state; it returns its outputs to the loop, which commits them in model order.
+
+Any shared state touched during execution must be concurrency-safe. This includes tool wrappers and providers: they may serialize internally or enforce their own capacity, but they must support concurrent dispatch without corrupting state.
+
+## Configuration and declarations
+
+`maxParallelToolCalls` is a positive AgentLoop deployment cap shared by every agent the factory creates. It defaults to `10`; `1` preserves serial execution. Exact fields and defaults live in the generated [configuration catalog](../../../../docs/config-catalog.md).
+
+The shipped declarations are conservative. Web search, web fetch, and filesystem read opt in. Filesystem writes and edits, bash tools, subagent delegation, workflow, user interaction, todo mutation, Code Mode, and Cordis mutation tools remain exclusive. A subagent may share its parent's workspace or external resources, and the unary classifier cannot prove that sibling delegations have disjoint effects. Bash has no proven input-sensitive classifier and remains exclusive.
+
+Filesystem read relies on a narrow recorder exception: its synchronous observation updates may settle out of order, but write and edit re-check the observed version before mutation, so stale state only produces `FS_STALE_VERSION`.
+
+## Verification
+
+Unit coverage pins fail-closed classification, typed argument validation, grouping, barriers, live reclassification after registry replacement, the rolling cap, distinct execution objects, middleware order, ordered results and context, and abort draining. First-party tests pin each parallel declaration.
+
+Snapshot coverage pins the visible multi-call transcript: pending calls may overlap while completed results remain model-ordered. Code Mode coverage pins its serial boundary. No provider-backed e2e is required because scheduling is deterministic loop behavior.
+
+## Alternatives considered
+
+**Keep serial execution.** This avoids new ordering and abort cases but retains unnecessary latency for independent sibling calls.
+
+**Use one tool-level boolean.** A fixed `supportsParallelToolCalls` flag is smaller but cannot distinguish a tool's read-only and mutating operations. The argument-sensitive classifier preserves that distinction.
+
+**Use stateful classification.** Giving the classifier a live agent, registry, or I/O access makes the decision depend on when it runs and creates a gap between classification and dispatch. Mutable authorization and stale-state checks remain execution-time responsibilities.
+
+**Use sibling-aware or resource-aware classification.** The scheduler could compare calls pairwise or let each call declare resource read/write claims. This can parallelize non-conflicting writes, but it requires shared resource identity and conflict semantics across unrelated tools. The unary contract instead gives up that concurrency and fails closed when safety is relational.
+
+**Parallelize the complete tool pipeline.** This keeps the loop on the public one-call API but runs pre- and post-execute middleware concurrently. Existing guards and hook bridges may carry ordered state, so only dispatch overlaps.
+
+**Expose staged methods or a scheduling waterfall.** Public `prepare` / `dispatch` / `finalize` methods or a `tools/execution-mode` event add extension surface before another consumer needs it. The loop uses an internal scheduler view, while `executionMode(exec)` leaves an insertion point for a policy seam.
+
+**Start calls while the model streams.** This may reduce latency further but changes assistant-message authority, replay, and call/result pairing. The scheduler starts only after the assistant message is complete.
+
+**Use fixed-size windows.** Waiting for every call in one window before starting the next leaves capacity idle behind a slow call. The rolling pool preserves the cap without that delay.
+
+**Expose concurrency metadata to the model.** The model can already emit sibling calls. Host scheduling metadata would enlarge requests without improving tool choice.
+
+## Consequences
+
+The design is fail-closed and simple for tool authors, but it cannot exploit concurrency whose safety depends on comparing siblings. A tool that opts in too broadly can expose latent shared-state races.
+
+Parallel calls may begin in cases where serial execution would have aborted before reaching them. The scheduler therefore records only started calls, drains them on abort, and never starts replacements after cancellation.
+
+Ordered commits may hold a fast result behind a slow earlier sibling. This preserves replay and model-history order while live surfaces still show pending progress.
+
+Concurrent external calls can compete for quota or process capacity. Providers own their capacity controls; the loop cap only limits calls from one agent step.
+
+Tool registration is a scheduling boundary. Registry mutations affect not-yet-started calls because the scheduler reclassifies after each barrier and before every pool replenishment. Already-started calls retain the scheduling decision under which they entered the pool.

.agents/notes/implemented/feature/2026-07-10-session-query-service.md

@@ -0,0 +1,40 @@
+# Agent Note: Exact session query service
+
+Status: implemented
+
+## Problem
+
+Session history exists in two places: current `SessionStore` objects and an optional persistence backend. Consumers that need exact inspection would otherwise duplicate live-versus-persisted precedence, persistence lifecycle handling, raw-event surface classification, relationship tracing, and defensive cloning. Durable state can lag the live log between checkpoints, so persistence alone is not a truthful current source.
+
+Full-text search is related but materially larger. Designing provider registration, extraction, synchronization, invalidation, ranking, and cursor contracts before a real backend exists creates two speculative state machines: one in the interface service and another in the eventual database package.
+
+## Decision
+
+`@deepseek-ai/dsh-session-query` owns `ctx.sessionQuery`, a small trusted exact-inspection service over one logical corpus. It exposes `listSessions()`, `listEvents(sessionId)`, bounded `readEvent(request)`, `traceSession(sessionId)`, and `traceEvent(request)`. It does not expose filters, text extractors, search requests, provider registration, or derived-index synchronization. The separate [tracing decision](2026-07-13-session-query-tracing.md) owns lineage and event-relationship semantics.
+
+The service observes the optional `ctx.sessionPersistence` binding dynamically but retains no persisted cache or invalidation listener. Each cross-corpus list asks the active backend for authoritative metadata, then overlays a fresh live-store list. Matching ids become one `SessionRecord`: the live header wins and `live`/`persisted` independently report source availability. Immutable header disagreement is `SESSION_QUERY_SOURCE_CONFLICT`.
+
+An exact target read first checks the live store and snapshots the live header and event log. This path never consults persistence, so a failing durable backend cannot make known live history unreadable. With no live target, the service lists current persistence metadata, proves the id exists, loads it, and rejects a list/load header mismatch. All returned headers and events cross one structured-clone boundary.
+
+## Surface semantics
+
+`dsh-session` exports `foldSurface(events)`, and `SurfaceManager` uses the same transition functions for its incremental cache. The fold returns detached current event sequences and each replacement's actual removed seqs. `listEvents()` and `traceEvent()` use that result to classify every raw event, so inspection cannot disagree with model-history derivation about positional replacement semantics.
+
+`readEvent()` returns the complete target plus raw neighbors by contiguous seq. `before` and `after` default to zero and are independently bounded by `readWindowMax`, default 50. The result carries a cloned `SessionHeader`, not a source-availability record, because determining a live target's persisted flag would violate the guarantee that live exact reads do not depend on persistence health.
+
+## Security boundary
+
+The service is context-wide trusted infrastructure, not an authorization layer. A future model-facing history tool or human UI applies explicit caller/session scope. The service adds no model-facing tool and changes no transcript or snapshot surface.
+
+## Alternatives considered
+
+- **Put logical-corpus resolution directly in every consumer** — rejected because source precedence, conflicts, optional-service lifecycle, cloning, and surface classification are shared correctness rules.
+- **Query only persistence** — rejected because checkpoints can lag the current live log.
+- **Cache persisted metadata and listen for writes/removals** — rejected because exact reads can ask the authoritative sources directly, while cache invalidation adds lifecycle and concurrency state before scale requires it.
+- **Define a provider-neutral search protocol now** — rejected because no provider consumes it. The first SQLite FTS package should own one reconciliation/transaction state machine; a smaller shared seam can be extracted later only when a second implementation proves the boundary.
+
+## Consequences
+
+The service has one source-resolution state variable: the currently mounted persistence service. There are no provider queues, fingerprints, extractor registries, observation generations, or derived index updates. Exact reads and event traces remain usable in live-only deployments and deterministic when persistence is present.
+
+Cross-corpus listing, lineage tracing, and persisted event operations perform backend I/O on each call. That is deliberate: correctness comes from current authoritative state, and scale-oriented search belongs to the proposed database package. Full-text search is unavailable until that package defines and implements its complete contract.

.agents/notes/implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md

@@ -0,0 +1,92 @@
+# Agent Note: Configure subagent persona, tool visibility, and depth
+
+Status: implemented
+
+## Problem
+
+A reusable subagent provider answers how to run a child, but different delegation tools need different child behavior. One deployment may want a reviewer persona, a research-only tool set, or a hard recursion bound without creating a new provider for every combination.
+
+These controls affect the child's first model request and therefore cannot be installed after the child is visible. They also need honest provider support: an ACP backend cannot silently accept an in-process-only tool filter, and a filter must not be described as a security boundary when every plugin runs in the same trusted process.
+
+## Decision
+
+Subagent starts have three independent composition controls: `persona`, `toolFilter`, and `maxDepth`. A provider advertises support for each control, the service rejects unsupported requests before starting a run, and an in-process provider installs the requested composition while the child is still unpublished.
+
+The controls answer different questions:
+
+| Control | Question | Result |
+|---|---|---|
+| `persona` | What role instructions replace the deployment persona for this child? | A child-local prompt section shadows `deployment:persona` |
+| `toolFilter` | Which deployment-global tools enter this child's visible tool view? | A scoped restriction filters globals before child-local tools are added |
+| `maxDepth` | How deep may this delegation tree grow? | A start whose child depth exceeds the absolute cap is rejected |
+
+`dsh-tool-subagent` exposes the controls as plugin configuration and copies them into each request it creates. Direct `SubagentService` callers may choose them per request. The provider capability descriptor remains the source of truth for whether a backend can honor each field.
+
+### Persona is a scoped shadow
+
+The persona control changes one child without changing deployment-wide prompt assembly. During unpublished setup, an in-process provider registers a child-scoped section named `deployment:persona`; ordinary most-specific-wins resolution replaces the global section only in that child's assemblies.
+
+The value has the same strict template semantics as the deployment persona. Omitting it inherits the deployment section through the global layer; an explicit empty string shadows the global persona with an empty section. Parent and sibling personas never enter the child's flat scope.
+
+This uses the normal system-prompt registration mechanism rather than a second persona channel. The first prompt therefore sees the same named contribution that later prompts and prompt-inspection tools see.
+
+### Tool filtering is one live global-view rule
+
+The tool filter controls capability visibility and executable lookup together. An in-process provider installs `ToolRegistry.restrict()` in the child's scope before publication, and the registry's single resolver applies the same result to wire tool schemas, lookup, execution, and Code Mode SDK generation. Independently registered system-prompt sections are outside `ToolRegistry`, so filtering a tool does not remove that plugin's standalone guidance.
+
+Resolution follows these rules:
+
+1. Each restriction applies `allow` before `deny` to the live deployment-global tool registry.
+2. Multiple restrictions intersect, so every installed restriction must admit a global tool.
+3. Child-scoped tools are added after global filtering and may shadow an admitted global tool.
+4. Reserved `run_code` presentation and other scope-local protocol contributions are outside the global filter.
+
+Configuration fails loudly when a filter supplies neither `allow` nor `deny`, or names something outside the current global restrictable set, including a scope-local-only or reserved name. `allow: []` is valid and deliberately hides every global tool. These checks catch misspellings and prevent configuration from appearing effective when it cannot affect the named entry.
+
+The global registry remains live. A deny-only filter admits a later global name unless it explicitly denies that name; an allow-list excludes a later global name unless it explicitly allows that name. Removing a global tool removes it from every resolved view. These semantics preserve hot registration while making the difference between allow and deny explicit.
+
+### Depth is an absolute tree cap
+
+The depth limit bounds recursive delegation independently of tool visibility. A top-level agent has depth zero; an in-process child has its parent's validated depth plus one. `maxDepth` is an absolute non-negative safe integer, and a start rejects before child ownership begins when the derived child depth is greater than the cap.
+
+Every public entry validates the domain rather than relying on one model-facing configuration path. Negative values, fractions, negative zero, non-finite values, unsafe integers, malformed stored parent depth, and derived overflow all reject. Omitting the cap leaves depth unbounded by this mechanism.
+
+A deployment can combine depth and filtering. For example, it may keep the delegation tool visible at depth one but set `maxDepth: 1`, or deny the delegation tool entirely in children. Neither choice changes the provider's conversation-history behavior.
+
+### Capability gating keeps providers honest
+
+Capabilities separate a requested feature from a provider implementation. `SubagentCapabilities` advertises `persona`, `toolFilter`, and `depthLimit`; `SubagentService.start()` checks every present request field against those flags before calling the provider.
+
+This lets spawn and fork providers share the in-process implementation while external providers advertise only what they can enforce. A request never degrades silently: selecting an unsupported control produces `UNSUPPORTED_CAPABILITY`, and no run or lifecycle event exists.
+
+### Unpublished setup makes the first request correct
+
+All child-local composition is complete before the child becomes observable. The in-process provider supplies one setup callback to agent creation; that callback installs persona, tool restriction, and structured-output contributions in the child's scope. Only after setup succeeds does creation publish the session and agent and allow the driver to start.
+
+A setup failure rolls back the private child. No observer can acquire a child whose first prompt used the deployment persona or unfiltered tool set and whose later prompts use the requested configuration.
+
+## Visibility is not authority
+
+These controls compose trusted same-process behavior; they do not authorize it. `toolFilter` changes the child view resolved by the tool registry, but it does not create a parent-to-child grant lattice, require a child to be a subset of its parent, sandbox plugins, or prevent code with another Cordis context from calling services directly.
+
+In particular, a child-local tool is added after the global filter and may be absent from the parent's view. A deny-only child also sees later global tools not named by the deny-list. Those are deliberate live-composition semantics, not non-escalation guarantees.
+
+A security design would need a separate authority representation, propagation rule, and execution-time enforcement point. Creation-time grant snapshots, parent-subset grants, explicit future-grant APIs, and generic capability/output/termination tags are outside this feature.
+
+## Alternatives considered
+
+**Create one provider per persona or tool set.** This multiplies providers that share the same transport and lifecycle implementation, makes dynamic deployment configuration awkward, and still needs a recursion mechanism. Providers remain about execution transport; requests carry per-child composition.
+
+**Copy the parent's complete tool view.** Registration scope is flat by design, and lifetime ownership does not imply visibility inheritance. Copying a resolved view would also freeze dynamic global registrations and conflate composition with authority without defining either contract fully.
+
+**Snapshot allowed global tools at child creation.** A frozen allow-set makes future registration uniformly unavailable, but it changes hot-registration semantics and starts an authorization design. The implemented filter stays a live registry predicate and documents allow-versus-deny behavior directly.
+
+**Hide only tool schemas.** Presentation-only filtering lets the model execute a tool that the prompt says does not exist through Code Mode or a forged call. One resolver governs both presentation and execution instead.
+
+**Use only tool filtering to stop recursion.** Removing the delegation tool is useful but provider-specific and does not protect direct service callers or alternate delegation tools. Absolute depth is an independent structural bound.
+
+## Consequences
+
+Contributors can configure child role, visible global tools, and recursion without defining new providers. Capability checks fail before ownership starts, unpublished setup makes the first request consistent, and one tool resolver prevents presentation/execution drift.
+
+The cost is that deployments must understand live allow/deny behavior and the distinction between visibility and authority. Provider authors must advertise each supported control accurately, and in-process providers must install every requested contribution before publication. The controls deliberately do not solve security confinement or parent-to-child non-escalation.

.agents/notes/implemented/feature/2026-07-13-session-query-tracing.md

@@ -0,0 +1,34 @@
+# Agent Note: Session query relationship tracing
+
+Status: implemented
+
+## Problem
+
+Session relationships are encoded across immutable headers, positional surface operations, and logged provenance arrays. A consumer reconstructing those relationships directly would need to duplicate corpus precedence, surface folding, malformed-log handling, deterministic lineage ordering, and cloning. Positional replacement and provenance are different graphs, so collapsing them into one generic edge type would also lose meaning.
+
+## Decision
+
+`ctx.sessionQuery` exposes `traceSession(sessionId)` and `traceEvent({ sessionId, seq })` alongside its exact reads. Both are one-shot views over the existing live-preferred corpus: session tracing consumes one complete corpus listing, while event tracing consumes one loaded logical log and one canonical surface fold. The service retains no lineage, reverse-index, or replacement state after a call.
+
+`SessionLineageTrace` returns the target, known parents in immediate-to-outward order, and recursive descendant trees whose siblings sort by creation time and then session id. `complete: true` carries the known root; `complete: false` carries the first unresolved parent id. A cycle connected to the target fails with `SESSION_QUERY_INVALID_LINEAGE`.
+
+`SessionEventTrace` keeps positional and provenance relationships separate. `replacedBy` is the immediate positional replacer, `replacementChain` follows replacers to the final node, and `replacedEventSeqs` lists the actual surface nodes directly removed by the target. `sourceEventSeqs` preserves direct logged source order, while `derivedEventSeqs` lists later direct reverse references in log order. Provenance is not expanded transitively.
+
+## Validation boundary
+
+Event tracing checks target existence before surface analysis. Both event listing and tracing then use `dsh-session`'s one-pass surface fold, which accepts or rejects the loaded log as a whole: event seqs are zero-based and contiguous, surface markers obey event-type eligibility, provenance belongs only to surface event types, present arrays are nonempty and duplicate-free, every source is an earlier seq, and every positional replacement names and cites all surface nodes it removes. Every contract failure uses `SESSION_QUERY_INVALID_SURFACE`; there is no weaker classification-only surface standard.
+
+All returned records and arrays are detached. A known live event trace never consults persistence; persisted event traces preserve the exact-read list/load consistency check. Session lineage is necessarily a cross-corpus operation and therefore preserves cross-corpus persistence failure semantics.
+
+## Alternatives considered
+
+- **Expose standalone tracing helpers** — rejected because the source-precedence and detachment boundary belongs to `ctx.sessionQuery`; public helpers would invite callers to bypass it.
+- **Combine replacement and provenance edges** — rejected because a positional replacement can shadow surface nodes while also citing non-surface construction inputs, and consumers need to distinguish those meanings.
+- **Return transitive provenance closure** — rejected because it obscures logged direct evidence, increases result size, and lets one malformed distant edge alter otherwise local output.
+- **Best-effort traces over malformed provenance** — rejected because a structurally plausible partial graph would look authoritative. Exact inspection fails loudly when the canonical relationship contract is broken.
+
+## Consequences
+
+Consumers receive deterministic relationship views without a cache or second corpus. Event tracing performs whole-log validation and allocation on each call, while lineage tracing lists the complete logical corpus on each call. Those costs keep the source of truth explicit and are separate from the content-bearing full-text-search and filtering API.
+
+The feature has unit and service-level coverage but no snapshot or end-to-end fixture because it introduces no model-facing consumer, transcript change, or cross-process protocol.

.agents/notes/implemented/feature/2026-07-14-time-context-plugin.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-14-time-context-plugin.md: 189f75fc12fe12e9dec56fc71ea901ec2eaa8b19
+2026-07-14-time-context-plugin.zh.md: 12671cb891531627fffabb7bd91a1532bc3de6b9

.agents/notes/implemented/feature/2026-07-14-time-context-plugin.md

@@ -0,0 +1,59 @@
+# Agent Note: Optional time-context plugin
+
+Status: implemented
+
+English | [中文](2026-07-14-time-context-plugin.zh.md)
+
+## Problem
+
+The dynamic system-prompt storage and refresh decision in this record is superseded by [Durable per-step time context](2026-07-16-durable-per-step-time-context.md). The opt-in package, zoned formatting, and validation remain; the follow-up owns the current model-visible and durability contract.
+
+An agent request has no live clock unless a deployment puts one in prompt text or gives the model a query tool. Static text becomes stale, while a tool call adds overhead to ordinary reasoning about dates, deadlines, or idle time. Without elapsed time, the model cannot distinguish an immediate follow-up from one sent hours after the preceding message.
+
+Prompt assembly can derive both facts per step from durable session timestamps, and request-header logging can record the exact rendered value. Accumulating stale readings in conversation history or waking idle agents would violate the existing request lifecycle.
+
+## Decision
+
+`@deepseek-ai/dsh-time-context` is an opt-in function plugin at `packages/context/time-context/`. The `context/` product group holds bounded request-context enrichments that define neither a tool nor a service. `dsh-agent-spine-demo` and shipped examples do not load the package; deployments mount it explicitly when its token and disclosure costs are acceptable.
+
+The plugin registers the global `context:time` system-prompt section at order 10, after the deployment persona and before tool guidance. For an active turn it emits an ISO-shaped timestamp with numeric UTC offset and IANA zone, plus a compact whole-second duration since the last model-visible message before the turn opened. Bare and idle assemblies receive an empty section.
+
+### Previous-message baseline
+
+At a turn's first assembly, the provider scans before `turn/start` for the latest `user/message`, `assistant/message`, `tool/result`, `context/message`, or `steering/message`. It excludes the current prompt so the duration expresses the inter-turn gap instead of approximately zero. Every refresh in that turn keeps the same baseline, and the first turn reports `unavailable (no earlier message in this session)`.
+
+The baseline is the session event's append time, not an unlogged client timestamp. Resume and fork behavior are therefore deterministic from the durable log, and the model-visible value remains reconstructable without a new event. A backward wall-clock adjustment clamps the duration to zero.
+
+### Refresh policy
+
+`refreshIntervalMs` defaults to 60,000 and must be a non-negative safe integer. Every turn's first request refreshes. Later assemblies in that turn reuse the block until its age reaches the interval; `0` refreshes every step. No timer creates work during model calls, tools, or idle time because refresh is request-bound.
+
+When `timeZone` is omitted, `Intl.DateTimeFormat` resolves the Node process's system zone once at plugin load. Node honors `TZ`; without that override, the host or container supplies the zone. An explicit value must be an IANA identifier and is validated at load. The captured zone remains stable until plugin reload, and the ISO-shaped local timestamp includes its current numeric offset so daylight-saving changes stay explicit. This is the deployment process's zone, not a remote user's zone.
+
+### Logging and token shape
+
+The loop records the temporal block in full `request/header` snapshots before transmission, satisfying the [reconstructable-requests contract](../architecture/2026-07-05-reconstructable-requests.md). Each request carries one current block; earlier readings do not remain in conversation history. The plugin owns the fact and contributes it through the prompt registry, following the [prompt-variables Agent Note](../architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md) without a loop special case.
+
+## Testing
+
+Unit tests pin formatting, baselines, refresh policy, validation, per-agent state, disposal, and load-time system-zone capture. A real agent-loop test pins the transmitted prompt and full `request/header` snapshots. A keyless subprocess e2e boots a test-only `cordis.yml` through the real Loader and stdio app, omits `timeZone` under a controlled `TZ`, drives two turns, and verifies the persisted request headers externally. Default snapshot compositions omit the plugin, so their transcript fixtures contain no temporal block.
+
+## Alternatives considered
+
+- **Append a `context/message` on every turn or refresh** — rejected because readings and token cost would accumulate in history. Replacing a prior surface node would preserve its old position, while replacing the tail would hide intervening conversation.
+- **Use `agent/session-prefix`** — rejected because the session-stable prefix cannot represent a per-turn or per-step clock.
+- **Mutate requests in `agent/request`** — rejected because that seam shapes call config after the message boundary; inserted model content would bypass prompt-pressure accounting and request-header logging.
+- **Register separate `{{current_time}}` and `{{elapsed}}` variables** — rejected because independent providers can sample different instants and require shared caching. One section records the pair atomically without a deployment-authored template.
+- **Refresh from a background timer** — rejected because a new value has no consumer outside request assembly. Timer-driven `agent.inject()` would create turns and wake idle sessions merely to report time passing.
+- **Keep UTC as the omitted default** — rejected because an explicitly enabled clock should follow its deployment environment unless the operator chooses UTC. `timeZone: UTC` remains available when a deployment requires it.
+- **Add a time-zone detection library** — rejected because Node's `Intl` runtime already exposes the process's IANA zone. Another dependency cannot infer a remote user's zone either.
+- **Mount the plugin in `dsh-agent-spine-demo`** — rejected because time zone, disclosure, token budget, and freshness are deployment policy. Opt-in keeps default context stable.
+- **Place the package in `core/`** — rejected because `core/` owns the product API spine, while this plugin is an optional leaf with no service key.
+
+## Consequences
+
+- Opted-in models receive a zoned clock and inter-turn duration without a tool call. The system-prompt cost is fixed per request instead of growing with the session.
+- An omitted `timeZone` follows the process's `TZ`, host, or container zone as observed at plugin load. Operators must configure an explicit zone when the deployment environment does not represent the intended user.
+- A refresh changes the request header and can add a full `request/header` snapshot with reason `change`. `refreshIntervalMs` trades freshness against the number and size of durable full snapshots; `0` records a new value on every step whose whole-second rendering changes.
+- No request exists solely to refresh time. A long-running tool leaves the prior reading until the next step assembles.
+- Duration reflects harness processing time at durable append boundaries, not client-network latency before logging. Preserving a client-origin timestamp requires a separate durable input contract.

.agents/notes/implemented/feature/2026-07-14-time-context-plugin.zh.md

@@ -0,0 +1,59 @@
+# Agent Note：可选时间上下文插件
+
+Status: implemented
+
+[English](2026-07-14-time-context-plugin.md) | 中文
+
+## 问题
+
+本记录中的动态系统提示词存储和刷新决策已由[持久的逐步骤时间上下文](2026-07-16-durable-per-step-time-context.md)取代。需要显式启用的包（package）、分区时间格式和校验仍然保留；后续 Agent Note 负责当前的模型可见与持久性契约。
+
+如果部署方既未在提示词中提供时钟，也未给模型提供查询工具，agent（智能体）请求就无法获得实时准确的时间。静态文本会变得陈旧，而对于日期、截止时间或闲置时长等常规推理，调用工具会增加开销。缺少已经过去的时长时，模型无法区分紧接着发送的消息与上一条消息几小时后才发送的消息。
+
+提示词组装流程可以在每个步骤中根据持久会话时间戳派生这两项信息，请求头日志则可以记录实际渲染的确切值。在会话历史中累积陈旧读数或唤醒空闲 agent 都会违反现有请求生命周期。
+
+## 决策
+
+`@deepseek-ai/dsh-time-context` 是位于 `packages/context/time-context/`、需要显式启用的函数插件。`context/` 产品分组用于容纳既不定义工具、也不定义服务的有界请求上下文增强。`dsh-agent-spine-demo` 和仓库提供的示例都不会加载该包；只有当 token 与信息披露成本可接受时，部署方才显式挂载它。
+
+该插件注册顺序值为 10 的全局系统提示词区段 `context:time`，位置在部署方角色设定之后、工具指导之前。对于活跃轮次，它会输出带数字 UTC 偏移和 IANA 时区、形似 ISO 的时间戳，以及从轮次开始前最后一条模型可见消息起算的紧凑整秒时长。未绑定 agent 或 agent 处于空闲状态时，该区段为空。
+
+### 上一条消息基线
+
+在轮次首次组装时，提供方会在 `turn/start` 之前查找最近的 `user/message`、`assistant/message`、`tool/result`、`context/message` 或 `steering/message`。它会排除当前提示词，使时长表达轮次间隔，而不是接近零。同一轮次中的每次刷新都保留这条基线；首个轮次报告 `unavailable (no earlier message in this session)`。
+
+基线采用会话事件的追加时间，而不是日志中不存在的客户端时间戳。因此，恢复和 fork 行为可以从持久日志中确定性重现，模型可见值也无需新增事件即可重建。系统挂钟向后调整时，插件会将时长钳制为零。
+
+### 刷新策略
+
+`refreshIntervalMs` 默认值为 60,000，并且必须是非负安全整数。每个轮次的首次请求都会刷新。同一轮次中的后续组装会复用该区块，直至其存在时间达到该间隔；设为 `0` 时每个步骤都刷新。刷新仅由请求驱动，因此在模型调用、工具运行或空闲期间，计时器不会创建任务。
+
+省略 `timeZone` 时，`Intl.DateTimeFormat` 会在插件加载时解析一次 Node 进程的系统时区。Node 会遵循 `TZ`；没有该覆盖值时，时区由主机或容器提供。显式值必须是 IANA 标识符，并在加载时接受校验。捕获的时区在插件重新加载前保持稳定，形似 ISO 的本地时间戳包含其当前数字偏移，使夏令时变化保持显式可见。该默认值代表部署进程的时区，而不是远程用户的时区。
+
+### 日志与 token 形态
+
+agent loop（智能体循环）会在发送前通过完整的 `request/header` 快照记录时间区块，从而满足[可重建请求契约](../architecture/2026-07-05-reconstructable-requests.md)。每个请求只携带一个当前区块；先前的读数不会保留在会话历史中。该插件拥有时间信息，并按照[提示词变量 Agent Note](../architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md)通过提示词注册表贡献该信息，无需为循环添加特殊分支。
+
+## 测试
+
+单元测试固定格式化、基线、刷新策略、校验、逐 agent 状态、资源释放行为，以及系统时区在加载时的捕获行为。使用真实 agent loop 的测试固定实际发送的提示词和完整的 `request/header` 快照。无密钥子进程端到端测试通过真实 Loader 和 stdio 应用启动测试专用 `cordis.yml`，在受控 `TZ` 下省略 `timeZone`，驱动两个轮次，并从外部校验持久请求头。默认快照组合不包含该插件，因此其中的 transcript（文本记录）fixture（测试前置数据）不包含时间区块。
+
+## 考虑过的替代方案
+
+- **每个轮次或每次刷新都追加一条 `context/message`**——不予采纳，因为读数和 token 成本会在历史中累积。替换先前的表层节点会保留其旧位置，而替换尾部节点会隐藏中间的会话内容。
+- **使用 `agent/session-prefix`**——不予采纳，因为会话期间保持稳定的前缀无法表示逐轮次或逐步骤变化的时钟。
+- **在 `agent/request` 中修改请求**——不予采纳，因为该边界在消息边界之后塑造调用配置；插入模型可见内容会绕过提示词压力核算和请求头日志。
+- **注册独立的 `{{current_time}}` 和 `{{elapsed}}` 变量**——不予采纳，因为独立提供方可能在不同时间点采样，并且需要共享缓存。单个区段会以原子方式记录两项信息，也不需要部署方编写时间模板。
+- **通过后台计时器刷新**——不予采纳，因为请求组装之外没有消费新值的对象。由计时器驱动 `agent.inject()` 会创建轮次，并且只为报告时间流逝就唤醒空闲会话。
+- **省略配置时仍默认使用 UTC**——不予采纳，因为显式启用的时钟应跟随部署环境，除非运维方选择 UTC。需要 UTC 的部署仍可配置 `timeZone: UTC`。
+- **引入时区探测库**——不予采纳，因为 Node 的 `Intl` 运行时已经能够提供进程的 IANA 时区，而且额外依赖同样无法推断远程用户的时区。
+- **在 `dsh-agent-spine-demo` 中挂载插件**——不予采纳，因为时区、信息披露、token 预算和新鲜度都属于部署策略。选择加入能保持默认上下文稳定。
+- **将包放入 `core/`**——不予采纳，因为 `core/` 负责产品 API 主干，而该插件是没有服务键的可选叶节点。
+
+## 后果
+
+- 选择加入的模型无需调用工具，即可获得分区时钟和轮次间隔时长。每个请求的系统提示词成本固定，不会随会话增长。
+- 省略 `timeZone` 时，插件采用加载时观察到的进程 `TZ`、主机或容器时区。当部署环境不能代表目标用户时，运维方必须显式配置时区。
+- 刷新会改变请求头，并可能新增一份 reason 为 `change` 的完整 `request/header` 快照。`refreshIntervalMs` 用新鲜度换取完整持久快照的数量与大小；设为 `0` 时，每个整秒渲染结果发生变化的步骤都会记录新值。
+- 系统不会仅为刷新时间而创建请求。长时间运行的工具会保留先前读数，直至下一步骤开始组装。
+- 时长反映持久追加边界处的 harness 处理时间，不包含消息进入日志之前的客户端网络延迟。若要保留客户端来源时间戳，需要单独的持久输入契约。

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-16-durable-per-step-time-context.md: 2d7076d51dbe1a64e5042230bddc6844141ff265
+2026-07-16-durable-per-step-time-context.zh.md: 432e0305cf44dcce1053c6580c9f0039309a7af4

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.md

@@ -0,0 +1,70 @@
+# Agent Note: Durable per-step time context
+
+Status: implemented
+
+English | [中文](2026-07-16-durable-per-step-time-context.zh.md)
+
+## Problem
+
+A request-only clock can tell the model the current time, but replacing that value in the system prompt removes the evidence behind earlier time-sensitive reasoning. Multi-step turns need requests to retain the readings that shaped preceding steps. The request must remain reconstructable after restart, and automatic compaction must account for the same timing context the model receives.
+
+A process-local refresh cache makes displayed time depend on state that cannot survive resume or be reconstructed from the durable session. Durable interval scheduling can reduce append frequency without introducing that hidden state.
+
+## Decision
+
+`@deepseek-ai/dsh-time-context` is an opt-in function plugin in `packages/context/time-context/`. It registers a prepended `agent/pre-step` listener and, when an injection is due, calls `agent.inject()` for a pre-step attempt whose signal is not already aborted. The injected `context/message` carries source `{ kind: 'plugin', plugin: 'time-context' }` and append surface metadata; a suppressed attempt appends nothing.
+
+The listener records preparation context before a possible `step/start`. Its prepended registration runs before ordinary automatic compaction listeners, so pressure estimation and any resulting surface rewrite observe a newly appended reading. A later pre-step listener can cancel or fail the attempt before the step opens; the reading remains because the durable log is append-only and this plugin performs no rollback.
+
+The optional `timeZone` config resolves the Node process's IANA zone once at plugin load when omitted; an explicit value is validated by `Intl.DateTimeFormat`. The timestamp includes the numeric UTC offset and resolved IANA zone.
+
+The optional `refreshIntervalMs` config is manually validated at plugin load as a non-negative safe integer. Omission or `0` injects on every eligible preparation attempt. A positive value scans the raw session events for the most recent `context/message` with this plugin's source and injects when none exists, wall time moved backward, or the event is at least the configured age. The raw event timestamp governs even after compaction shadows the message, so scheduling persists across turns and process resume without a timer or process-local cache.
+
+### Text and elapsed baselines
+
+An injected first-step reading is:
+
+```text
+Time sampled while preparing turn <turn>, step 1: <timestamp>
+Elapsed since the preceding model-visible message: <duration-or-unavailable>.
+```
+
+The baseline is the latest preceding user, assistant, tool-result, context, or steering message. This includes the accepted prompt that opened an ordinary message turn. If no model-visible message exists, the duration is `unavailable`.
+
+An injected later-step reading is:
+
+```text
+Time sampled while preparing turn <turn>, step <step>: <timestamp>
+Elapsed since the preceding step context: <duration-or-unavailable>.
+```
+
+Their baseline is the durable event timestamp of the preceding time-context message in the same turn. If interval suppression leaves no earlier same-turn reading, the duration is `unavailable`. Duration formatting uses compact whole-second units and clamps backward wall-clock movement to zero. The explicit turn and step make every retained reading attributable to its historical preparation attempt after later turns append more context.
+
+### Durability and request reconstruction
+
+Each reading remains a normal surface node until compaction shadows it; positive interval scheduling never removes existing readings. A later request therefore sees the cumulative unshadowed readings that affected earlier preparation and steps, rather than a system-prompt value rewritten in place.
+
+The plugin contributes nothing to system-prompt assembly. `request/header` contains no time-context text; request reconstruction obtains the complete durable surface prefix at each `step/start`. Readings and requests need not map one-to-one because a failed preparation can leave a reading while interval suppression can prepare a request without appending one. The plugin depends on the agent registry for its lifecycle listener and does not require the system-prompt service at runtime.
+
+## Testing
+
+Unit and real-loop tests pin formatting, both elapsed baselines, interval omission and zero, threshold boundaries, cross-turn and per-session scheduling, backward-clock behavior, invalid config, resumed raw-event lookup after compaction, aborted-signal behavior, later-listener cancellation and failure, listener disposal, source and surface metadata, cumulative multi-step visibility, and absence from request headers. A keyless subprocess e2e boots the real Loader and stdio app, drives two turns, and verifies the persisted context events externally.
+
+## Supersedes
+
+This decision supersedes the dynamic system-prompt storage and refresh policy in [Optional time-context plugin](2026-07-14-time-context-plugin.md). It keeps the package location, opt-in deployment stance, timestamp formatting, process-zone default, and load-time validation. Durable history replaces the `context:time` prompt section, process-local refresh cache, and request-header deltas; `refreshIntervalMs` controls durable append frequency instead of prompt replacement.
+
+## Alternatives considered
+
+- **Keep the dynamic system-prompt section and process-local refresh cache** — rejected because replacement erases earlier readings, cache state is not replayable, and a frozen request envelope would make the value stale for an entire loop instance.
+- **Replace the preceding context surface node** — rejected because replacement preserves the old node's position or shadows intervening conversation; neither represents when the new reading became visible.
+- **Inject from a background timer** — rejected because idle time has no pending request to consume the value, and timer-driven injection would create durable turns solely to report time passing.
+- **Expose time only through a tool** — rejected because ordinary temporal reasoning would require an avoidable tool round trip and would not guarantee a reading before every step.
+- **Use `agent/session-prefix`** — rejected because one loop-instance prefix cannot represent distinct step timestamps and does not accumulate historically attributable readings.
+
+## Consequences
+
+- Omission or `0` records every eligible preparation attempt; a positive interval reduces append frequency and history growth while preserving durable scheduling across resume.
+- Timing context remains append-only until compaction shadows older surface nodes, including a preparation reading left by a later cancellation or failure.
+- The first-step duration normally measures from the prompt that opened the turn, while later-step durations measure model and tool processing since the preceding step context.
+- An omitted `timeZone` still reflects the deployment process rather than a remote user, and elapsed time still uses durable harness append boundaries rather than client-origin timestamps.

.agents/notes/implemented/feature/2026-07-16-durable-per-step-time-context.zh.md

@@ -0,0 +1,70 @@
+# Agent Note: 持久的逐步骤时间上下文
+
+Status: implemented
+
+[English](2026-07-16-durable-per-step-time-context.md) | 中文
+
+## 问题
+
+仅存在于请求中的时钟可以告诉模型当前时间，但在系统提示词中替换这个值会移除先前时效性推理所依据的证据。在包含多个步骤的轮次中，请求需要保留影响先前步骤的读数。系统必须能在重启后重建请求，自动压缩（compaction）也必须核算模型实际收到的同一份时间上下文。
+
+进程本地刷新缓存使显示的时间依赖无法在恢复后保留、也无法从持久会话重建的状态。持久的间隔调度可以减少追加频率，而不引入这种隐藏状态。
+
+## 决策
+
+`@deepseek-ai/dsh-time-context` 是位于 `packages/context/time-context/`、需要显式启用的函数插件。它注册一个前置的 `agent/pre-step` 监听器，并在需要注入时，为信号尚未取消的预步骤尝试调用 `agent.inject()`。注入的 `context/message` 携带来源 `{ kind: 'plugin', plugin: 'time-context' }` 和追加表层元数据；受间隔抑制的尝试不会追加任何内容。
+
+监听器在可能出现的 `step/start` 之前记录准备上下文。它采用前置注册，因此先于普通自动压缩监听器运行，使压力估算和由此产生的表层重写都能观察到新追加的读数。后续预步骤监听器可能在步骤开启前取消尝试或使其失败；持久日志仅追加，且本插件不执行回滚，因此该读数会保留下来。
+
+省略可选配置 `timeZone` 时，插件在加载时解析一次 Node 进程的 IANA 时区；显式值由 `Intl.DateTimeFormat` 校验。时间戳包含数字 UTC 偏移和解析后的 IANA 时区。
+
+插件在加载时手动校验可选配置 `refreshIntervalMs`，其值必须为非负安全整数。省略或设为 `0` 时，每次符合条件的准备尝试都会注入。设为正数时，插件扫描原始会话事件，查找来源属于本插件的最新 `context/message`；不存在此类事件、系统挂钟向后移动，或该事件已达到配置时长时，插件执行注入。即使压缩已隐藏消息，调度仍以原始事件时间戳为准，因此该机制无需计时器或进程本地缓存，也能跨轮次和进程恢复持续生效。
+
+### 文本与时长基线
+
+第一个步骤的注入读数为：
+
+```text
+Time sampled while preparing turn <turn>, step 1: <timestamp>
+Elapsed since the preceding model-visible message: <duration-or-unavailable>.
+```
+
+基线是前一条用户消息、助手消息、工具结果、上下文消息或 steering（中途引导）消息。对于普通消息轮次，这包括开启轮次的已接受提示词。如果不存在模型可见消息，时长为 `unavailable`。
+
+后续步骤的注入读数为：
+
+```text
+Time sampled while preparing turn <turn>, step <step>: <timestamp>
+Elapsed since the preceding step context: <duration-or-unavailable>.
+```
+
+其基线是同一轮次中上一条时间上下文消息的持久事件时间戳。如果间隔抑制导致同一轮次中没有更早的读数，时长为 `unavailable`。时长采用紧凑的整秒单位，并在系统挂钟向后移动时钳制为零。显式的轮次号和步骤号使每个保留的读数在后续轮次追加更多上下文后，仍可归属于对应的历史准备尝试。
+
+### 持久性与请求重建
+
+每个读数都作为普通表层节点保留，直至压缩将其隐藏；正数间隔调度绝不会移除已有读数。因此，后续请求会看到影响先前准备过程和步骤且尚未被隐藏的累计读数，而不是一个被原地改写的系统提示词值。
+
+插件不向系统提示词组装贡献任何内容。`request/header` 不包含时间上下文文本；请求重建从每个 `step/start` 取得完整的持久表层前缀。读数与请求无需一一对应，因为失败的准备过程可能留下读数，而间隔抑制也可能使请求准备过程不追加读数。插件通过 agent 注册表使用生命周期监听器，运行时不需要系统提示词服务。
+
+## 测试
+
+单元测试和真实 agent loop（智能体循环）测试固定格式化、两种时长基线、间隔省略和零值、阈值边界、跨轮次和各会话独立调度、挂钟后退行为、无效配置、压缩后基于恢复会话的原始事件查找、已取消信号行为、后续监听器取消和失败、监听器 dispose（资源释放）、来源与表层元数据、多步骤累计可见性，以及请求头中不存在时间上下文。无密钥子进程 e2e 测试通过真实 Loader 和 stdio 应用启动，驱动两个轮次，并从外部校验持久化的上下文事件。
+
+## 取代的决策
+
+本决策取代[可选时间上下文插件](2026-07-14-time-context-plugin.md)中的动态系统提示词存储和刷新策略。它保留包位置、选择加入式部署、时间戳格式、进程时区默认值和加载时校验。持久历史取代 `context:time` 提示词区段、进程本地刷新缓存和请求头增量；`refreshIntervalMs` 用于控制持久追加频率，而非提示词替换。
+
+## 考虑过的替代方案
+
+- **保留动态系统提示词区段和进程本地刷新缓存**——不予采纳，因为替换会抹去先前读数，缓存状态无法回放，而且冻结的请求内容集合会使该值在整个 agent loop 实例期间保持陈旧。
+- **替换前一条上下文表层节点**——不予采纳，因为替换会保留旧节点的位置或隐藏中间的会话内容；两者都不能表达新读数何时开始可见。
+- **通过后台计时器注入**——不予采纳，因为空闲期间没有待处理请求消费该值，而且计时器驱动的注入会仅为报告时间流逝而创建持久轮次。
+- **只通过工具提供时间**——不予采纳，因为普通时间推理会产生本可避免的工具往返，也不能保证每个步骤之前都有读数。
+- **使用 `agent/session-prefix`**——不予采纳，因为一个 loop 实例前缀无法表示不同的步骤时间戳，也不会累计具有历史归属的读数。
+
+## 后果
+
+- 省略 `refreshIntervalMs` 或设为 `0` 时，每次符合条件的准备尝试都会留下记录；正数间隔会减少追加频率和历史增长，同时使持久调度在恢复后继续生效。
+- 时间上下文仅追加并保留到压缩隐藏旧表层节点为止，其中也包括后续取消或失败所留下的准备读数。
+- 第一个步骤的时长通常从开启轮次的提示词起算，后续步骤的时长则反映自上一条步骤上下文以来的模型与工具处理时间。
+- 省略 `timeZone` 时仍采用部署进程而非远程用户的时区，时长仍采用 harness 的持久追加边界而非客户端来源时间戳。

.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write
+2026-07-17-dedicated-full-screen-tui-front-door.md: 178b5ea44be67f820a8ea7fed8acb987dffb3f80
+2026-07-17-dedicated-full-screen-tui-front-door.zh.md: ac055bad1b7a692c7a980430fdbd1e34737a9994

.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.md

@@ -0,0 +1,48 @@
+# Agent Note: Dedicated full-screen TUI front door
+
+Status: implemented
+
+English | [中文](2026-07-17-dedicated-full-screen-tui-front-door.zh.md)
+
+## Problem
+
+The line-oriented `@deepseek-ai/dsh-stdio` front door works in pipes and ordinary terminals, but a full-screen coding interface must own raw input, differential screen drawing, cursor state, overlays, and terminal restoration. Combining those contracts in one UI plugin couples the pipe-safe path to a TTY-only lifecycle and makes it unclear which terminal behavior a composition selects.
+
+The interactive channel must remain a Cordis plugin over the same agent, session, tool, and user-interaction services as every other front door. It needs to resume durable history, follow compaction replacements, display tool-owned presentation, and restore the terminal on startup failure and disposal. A standalone chat application or a second agent composition would duplicate behavior outside the plugin graph.
+
+## Decision
+
+DeepSeek Harness ships [`@deepseek-ai/dsh-tui`](../../../../packages/ui/tui/README.md) as a dedicated Cordis plugin. It owns terminal input and presentation only; agent lifecycle, session persistence, tool execution, and the model-facing question tool remain separate composition entries. The plugin requires both stdin and stdout to be TTYs and fails instead of silently changing to line-oriented behavior.
+
+The app layer selects a concrete terminal front door before mounting it. `@deepseek-ai/dsh-stdio-demo` can resolve `auto` from the two process streams, while the `repl-agent` and `tui-agent` leaves explicitly select readline and TUI respectively. The TUI leaf reuses the repl-agent backend and tool composition through an asserted include patch, so the three runnable agent leaves remain symmetric without duplicating deployment choices.
+
+The selected front door receives the exact generated or resumed `SessionId` used by the pre-created agent. It mounts before the agent composition, waits for the matching root agent, and enters full-screen mode only after that agent exists. A matching `agent-loop/config-start-failed` event is therefore reported before screen takeover and exits with status 1.
+
+### Session projection and interaction
+
+The TUI rebuilds the transcript from the active `session.surface` and reprojects it whenever an event carries a `surfaceOp`, so resumed and compacted history matches the model-visible conversation. It renders Markdown text and reasoning, token totals, the latest `todo/write` plan, and tool cards produced through each tool definition's `presentCall` and `presentResult` methods. Pending chunks and tool calls update the same components that completed events settle.
+
+Editor input calls `agent.send()` while idle and `agent.steer()` while a turn is running. Cancellation, reasoning visibility, tool-card expansion, redraw, transcript clearing, and exit are terminal-only controls. The plugin registers the shared `userInteraction` provider and presents questions as queued keyboard overlays; agent behavior and answer logging remain owned by their existing services.
+
+### Terminal ownership
+
+Before model output, session data, tool presentation, questions, configuration, or diagnostics reach pi-tui or the terminal title, `displayText()` renders C0 and C1 controls other than line feeds as visible hexadecimal escapes. Only the TUI and pi-tui create ANSI control sequences.
+
+The built-in palette uses standard 16-color ANSI foregrounds and SGR attributes, keeps body text and backgrounds at terminal defaults, and uses reverse video for selection. Host terminals therefore remap the interface for light and dark themes without a TUI-specific theme setting; `color: false` removes styling.
+
+## Verification
+
+The implemented [TUI terminal-state snapshot Agent Note](../testing/2026-07-18-tui-terminal-state-snapshots.md) owns the four-layer verification contract: direct behavior tests, transient semantic terminal snapshots, recorded JSONL journeys through production tools, and Loader/PTY smoke tests. The package README owns configuration, commands, model-visible effects, and current limitations.
+
+## Alternatives considered
+
+- **Keep readline and full-screen modes inside `@deepseek-ai/dsh-stdio`** — rejected because line-oriented output and differential TTY rendering have different dependencies, input rules, logging ownership, and teardown obligations. Separate packages keep the pipe-safe contract small and explicit.
+- **Let the TUI plugin silently downgrade when either stream is not a TTY** — rejected because a fallback hides deployment mistakes and changes interaction semantics. The app bundle may select a front door with `auto`; an explicitly mounted TUI fails loud.
+- **Keep TUI wiring and tests under the readline `repl-agent` leaf** — rejected because one leaf would represent two distinct front doors and break symmetry with `acp-agent`. A dedicated `tui-agent` leaf owns TUI overlays and tests while reusing the repl-agent backend composition.
+
+## Consequences
+
+- Interactive terminal work gains a stateful Markdown, card, plan, and question interface without changing the line-oriented protocol used by pipes and automation.
+- The TUI carries a pi-tui dependency and a strict TTY requirement; non-TTY deployments select `@deepseek-ai/dsh-stdio` at composition time.
+- Session projection makes resume and compaction consistent with the durable conversation, but one configured session owns the transcript and editor.
+- Tool packages extend terminal cards through their existing presentation methods without adding tool-specific branches to the TUI.