Adds the snapshot-test harness and the keyless replay pipeline end-to-end.
- snapshot-harness.ts: boots the real acp-agent subprocess via the cordis
Loader (preserving TSX_TSCONFIG_PATH so unbuilt dsh-* imports resolve from a
temp cwd), tees raw stdout into an SDK ClientSideConnection, interprets a
per-scenario input.json DSL (initialize / newSession capturing the random
sessionId / prompt / cancel), closes stdin to trigger graceful shutdown, and
harvests the persisted session.jsonl. Failure-safe: a finally block SIGKILLs
a live child, awaits its exit, and removes both temp dirs even on a thrown
step or harvest. Raw bytes are buffered and decoded once (no multibyte split).
- snapshot-normalize.ts (+ spec): two pure normalizers (stdout frames + session
JSONL) scrub cwd, session ids / UUIDs, and JSON-RPC ids, and zero time /
createdAt — but keep `seq` (deterministic by contract). normalizeStdout throws
on a non-JSON line (the stdout-purity check).
- start.ts: selects cordis.snapshot.yml (replay, providerless) or
cordis.snapshot-record.yml (record, real adapter) from DSH_SNAPSHOT, skips
.env in replay, and disposes the ctx on stdin end so persistence flushes
before exit (harvest-after-flush, not on the prompt response).
- acp.snapshot.ts: asserts the normalized stdout golden (and, for model
scenarios, the re-persisted JSONL golden) via toMatchFileSnapshot; record mode
writes the harvested log back to the scenario fixture; an orphan-fixture guard
fails on an unregistered scenario dir.
- handshake scenario: initialize + session/new (no model call; a header-only
session.jsonl, since session/new persists no events).
- vitest.snapshot.config.ts, test:snapshot / test:snapshot:record scripts, a
pre-push snapshot job, and the knip entry.
Incorporates Codex review: record-fixture writeback, failure-safe teardown,
seq-not-scrubbed, harvest-after-flush. Per docs/rfc/implemented/2026-06-19.
Per a design revision, the per-scenario snapshot fixture becomes EXACTLY the
persisted session JSONL (<scenario>/session.jsonl) rather than a hand-authored
llm.json. The log already holds all LLM behavior (assistant/chunk carries every
StreamChunk) AND all harness behavior (tool/call, tool/result, turn/*, usage),
so one artifact drives replay and doubles as a behavioral golden.
llm-replay becomes replay-only (the record-tee is removed; recording is now
"run the real agent once and harvest the .jsonl", done by the harness in a
later commit). deriveReplayScript(events) groups assistant/chunk by (turn,step)
in log order — exact because the loop makes one ctx.llm.stream() call per step
and tags each chunk with the current (turn,step). The two failure modes the log
can't express (a thrown stream — no terminal finish; cancel/hang — timing) use
an optional replay.override.json sidecar.
Hardens against a Codex review finding: a derived group is only valid if it
ends in a `finish` chunk. A group without one is the fingerprint of a thrown
stream() and is NOT silently replayed as a clean stop — deriveReplayScript
throws, naming the (turn,step), so a missing sidecar override fails loud.
Updates the unit tests (parse/derive/load helpers, sidecar override, finish-
terminated grouping, HMR), the example README, and the RFC prose to the JSONL
format. Two goldens (stdout transcript + re-persisted JSONL) and the harness
wiring land in the next commit.
Introduces examples/acp-agent/src/llm-replay.ts, a function/namespace plugin
that installs a single llm/stream waterfall listener. In record mode it tees
the real model's StreamChunks into a per-scenario llm.json (flushed atomically
after EACH stream, since the snapshot subprocess is SIGKILLed and start.ts has
no disposal path). In replay mode it short-circuits the waterfall and serves
recorded streams back positionally — the Nth stream() call gets the Nth entry —
so a snapshot test can drive the real agent with no API key.
Each fixture entry is a discriminated record {chunks|throw|hang} so it can
replay BOTH branches of the LLM failure contract (throw from stream() vs a
finish-error chunk) plus cancellation. A throw entry carries the prefix chunks
emitted before the throw, replayed before the error, so a mid-stream failure
(partial output then STREAM_CLOSED) reproduces what the loop saw live.
Fail-loud on a missing or exhausted fixture (never a silent skip). Unit tests
drive the real LlmService waterfall (record tee, ordered replay, the three
entry kinds, partial-then-throw, fail-loud, event-driven abort, HMR-safety).
Broadens the unit vitest include to examples/*/tests and registers the plugin
+ snapshot tests as knip entries. Per docs/rfc/implemented/2026-06-19.
Match claude-agent-acp / codex-acp: the bash tool_call title IS the command
(an execute card hides rawInput), the model description rides as a content
text block above the card, and the completed card carries an exit-status pill
via _meta.terminal_exit.
Bridge fixes found in review of the prior terminal-card commit:
- tool_call_update.content is OMITTED in terminal mode (an ACP update.content
REPLACES the call's content collection in Zed, so the fenced ```console block
would clobber the terminal content block).
- terminal.output preserves RAW newlines (terminal renderers rely on exact
bytes); only the fenced fallback trims trailing blank lines.
- a relative workdir is resolved against the session cwd for the card header,
matching where the command actually ran.
- result-side terminal output is gated on the pending call having registered a
terminal (no orphan _meta.terminal_output for a terminal Zed never made).
The exit pill is recovered by parsing renderResult's status markers (the pure
presentResult seam sees only content blocks); a round-trip test pins the parse
to the marker emission. Neutral ToolTerminal gains exitCode/signal; widened
ToolCallPresentation with a content block. Docs (RFC + 3 READMEs) updated;
with-key e2e verifies the card + exit pill against the real model.
When the client advertises clientCapabilities._meta.terminal_output (Zed), a
bash tool call now renders as a real TERMINAL card — a cwd header + the command
+ its output — instead of the plain ```console text block. Keeps agent-side
dsh-bash execution; rejects the spec's client-side terminal/create (which would
bypass sandbox/env-scrub/ownership/cwd). Matches what claude-agent-acp and
codex-acp do; wire contract verified against Zed's source.
- dsh-tools: a provider-neutral ToolTerminal shape ({ cwd?, output? }) on
ToolCallPresentation/ToolResultPresentation — a tool asks "render me as a
terminal"; no ACP types leak in.
- dsh-tool-bash: bash presentCall marks terminal (cwd from an explicit absolute
workdir, else left for the bridge to fill from the session cwd); presentResult
carries the output alongside the ```console fallback.
- dsh-acp: initialize reads/remembers the _meta.terminal_output capability;
streamSessionEventUpdate maps a terminal presentation to
content:[{type:'terminal',terminalId}] + _meta.terminal_info on the call and
_meta.terminal_output on the update WHEN capable — else the unchanged text
path. terminalId is the callId; cwd defaults to the session header. The pure
translator gained a TerminalRendering {enabled,cwd} param (off by default).
Tests via the REAL tool-bash + bash-local: capability ON -> terminal content +
_meta; OFF -> no _meta (text path). The with-key e2e adds a real-model terminal
card case (echo over ACP with the capability on). 773 tests, 100% coverage.
The exit-status pill (_meta.terminal_exit), live streaming
(_meta.terminal_output_delta), and command classification are RFC follow-ups.
In Zed the tool-call card showed only "bash" — the bare tool name — instead
of what the command does. Fix it by letting each TOOL own how its calls render,
rather than the bridge special-casing names.
dsh-tools: add an optional two-state presentation seam to ToolDefinition /
defineTool — `presentCall(args)` (pending: title, kind, rawInput) and
`presentResult(args, result)` (completed: title?, content?). Provider-neutral
`ToolCallKind`/`ToolCallPresentation`/`ToolResultPresentation` vocabulary so
tools never depend on ACP. defineTool soft-validates args (display runs on log
replay, so a malformed/old shape returns undefined instead of throwing).
dsh-tool-bash: bash declares presentCall (model `description` → title, exact
`command` → rawInput, kind execute) and presentResult (wrap output in a fenced
```console block — a UI-only affordance kept out of the model-facing result);
bash_output/bash_kill present task-scoped titles.
dsh-acp: inject `tools`; a per-session `ToolPresenter` looks the tool up by name
and maps its neutral presentation to the ACP tool_call/tool_call_update wire
shape, with a generic fallback (title = name) for tools that declare nothing.
Because the `tool/result` event carries only {callId, content, isError}, the
presenter keeps a small bridge-local map of ONLY in-flight calls' (name, args),
keyed by callId and removed as each result is presented — no event-schema or
core change. Replay uses a throwaway presenter so loaded sessions render
identically to live ones.
Tests: dsh-tools defineTool presenters (typed args, soft-validate), tool-bash
bash/bash_output/bash_kill presenters, acp ToolPresenter (tool-owned mapping,
unknown-callId fallback, in-flight-only map), and an end-to-end turn through the
bridge. The key-gated e2e now asserts a real bash call's title is the model
description (not "bash") and rawInput is the command — verified against the real
DeepSeek model. The test harness derives its inject from the bridge's exported
`inject` so it can't drift again.
Two independent bugs made the ACP server crash the moment an editor (Zed)
connected, despite 178 green unit tests at 100% coverage:
1. `session/new` threw `cannot get property "agents" without inject`. Root
cause: a stray `export default apply` made the cordis Loader's
`unwrapExports` (`exports.default ?? exports`) collapse the module to the
bare `apply` function, discarding the sibling `inject`/`name`/`Config`
named exports. The plugin fiber was built with empty `inject`, so every
`ctx.<service>` read in `apply` threw at load. Fix: remove the default
export so the Loader uses the namespace.
2. `session/load` threw `cannot get property "sessionPersistence" without
inject`. `AgentLoop.resume` read `this.ctx.sessionPersistence` (a service
it deliberately does NOT inject); the property proxy's ancestor-only fiber
walk fails through the bridge's traceable shadow. Fix: read it via
`this.ctx.get('sessionPersistence', false)`, the topology-independent
global-store lookup.
Why the suite missed both: every test mounted the plugin by hand
(`ctx.plugin({name,inject,apply})`), bypassing `unwrapExports` entirely, and
the only test driving these RPCs was key-gated (skipped in CI). Added a no-key
`session/new` e2e that boots the real example through the real Loader — it
fails loudly on bug #1 without an API key. Set `TSX_TSCONFIG_PATH` in the e2e
spawn so the subprocess resolves workspace `paths` from a temp cwd (it was
silently falling back to a stale built `lib/`).
Docs: post-mortem 0001; AGENTS.md "line coverage is not behavior coverage" +
with-key/smoke-test philosophy; packages/AGENTS.md plugin-export-shape and
ctx.get rules; dsh-code-review SKILL checks.
Lifts the RFC 010 § Deferred restriction that the server had to launch in the
workspace ("cwd must equal the launch directory"). An editor can now open any
project folder, and N concurrent sessions over one connection can each target a
different directory.
- packages/acp: drop the `cwd === process.cwd()` guard in validateWorkspaceParams
(keep "must be absolute" — the cwd becomes the session header / bash workdir),
and drop the persisted-cwd-vs-launch-dir check in session/load (a resumed
session keeps its original header.cwd, so its bash tools run in its workspace).
- packages/tool-bash: the missing link — default the bash workdir to the calling
agent's session cwd (`exec.agent.session.header.cwd`) via a new resolveWorkdir
helper. An explicit model `workdir` still wins; a relative one resolves against
the session cwd. This is the only correct spot for multi-session: N sessions
share one ctx.bash executor, so the workdir must come per-call from exec.agent,
not executor config. Falls back to the executor default when no session cwd is
available (preserves non-ACP behavior).
- Trust: the cwd originates from the ACP client (the user's editor) at
session/new — same trust level as the old launch dir; no new untrusted-input
path. `additionalDirectories` (scope widening / sandbox) stays rejected.
- Tests: bridge accepts any absolute cwd + records it on the header; session/load
honors the persisted cwd; bash defaults to / resolves relative against the
session cwd; two sessions with different cwds each run bash in their own dir;
non-absolute cwd still rejected. 100% per-file coverage maintained.
- Docs: RFC 010 status + § Deferred cwd bullet marked RESOLVED; acp README adds a
Per-session cwd section; tool-bash + example READMEs and e2e comments updated.
Lifts the RFC 010 single-session-per-connection cap: the bridge now runs N
concurrent sessions over one connection, each mapped to its own LoopAgent.
- packages/acp: live sessions held in a Map<sessionId, SessionRecord> with an
agent→sessionId reverse WeakMap so agent/* events (which carry only the
Agent) demux in O(1). Every session/event and agent/status is routed strictly
to its owning record — concurrent sessions never cross-settle or interleave
their session/update notifications. Per-session state: one in-flight prompt
each, session/cancel aborts+settles only its own agent/prompt, session/load
reserves a per-id load slot (distinct ids load concurrently; re-loading a live
id is rejected), and disposal drains every live session in parallel to
quiescence.
- packages/tool-bash: record each background task's owning agent at spawn and
keep it for the executor's lifetime (NOT cleared on completion).
bash_output/bash_kill reject a task owned by a different agent (a task with no
owner is open; a no-agent caller can't access an owned task). Task ids are
global and predictable, so this is the fence that stops one session's agent
from reading/killing another session's background task.
- Per-session permission ownership and a per-agent disposer seam stay deferred
(depend on the deferred permission gate); the reverse map the gate will route
through is in place. RFC 011 stays `proposed`.
- Tests: two sessions stream concurrently without interleave; cross-session
cancel isolation; per-session in-flight enforcement; dispose-all-to-quiescence;
bash cross-session read/kill rejected (+ no-agent and unowned-task cases).
- Docs: RFC 011 implementation-status note; acp + tool-bash READMEs; example
MVP-limitations updated. 100% per-file coverage maintained.
Implements the RFC 010 MVP: a new `@deepseek-ai/dsh-acp` package bridges the
harness agent to the Agent Client Protocol (JSON-RPC 2.0 over newline-delimited
stdio), so Zed and other ACP editors can drive the coding agent — streaming
render, tool-call display, and resumable sessions via `session/load`.
- packages/acp: AgentSideConnection wiring; initialize/newSession/loadSession/
prompt/cancel; a total TurnEndReason→StopReason codec; settle-once with a
fallback chain (agent/turn-end → logged turn/end → idle); single-session
guard; cwd-must-equal-launch-dir validation; load replays from the persisted
event log (assistant/chunk→agent_message_chunk, tool/call/result→tool_call*).
- agent: add Agent.whenIdle() quiescence signal to the interface; LoopAgent
implements it (resolves on the first running→idle/disposed transition). The
bridge awaits it on disposal so teardown reaches quiescence, not just abort.
- examples: extract the shared provider/tool core into examples/base.yml;
coding-agent nest-includes it; new examples/acp-agent serves the agent over
ACP with JSONL persistence and no stdout logger (stdout is the protocol).
- Permission gate deferred (TODO(rfc010-permission-gate)): tools run with the
executor's full authority; only the Agent→sessionId ownership seam is laid
down. Cancel is best-effort for a not-yet-started queued turn
(TODO(rfc010-cancel-prestep)). RFC 010 stays `proposed`.
- Docs: package README + Zed snippet; client-driver cookbook section; root and
packages layout/commands; RFC 010 implementation-status note.
48 bridge tests + whenIdle coverage; 100% per-file coverage; e2e boots the
example as a subprocess and verifies a written file on disk (key-gated, with a
no-key stdout-purity check).
