The three-entry cordis.yml (dsh-sandbox-local + dsh-bash-sandbox at a
read-only default + dsh-approval) served over ACP: the first live approval
composition. Recorded snapshot scenarios pin the wire end to end —
config-options advertisement, the mode-switching arc as the suite pinned
header (both switches, the prompt-section delta, one changed-by-the-user
notice per knob, a confined write landing under the switched mode), and
both escalation branches over scripted permissionAnswers (a grant runs
confined under workspace-write; a rejection executes nothing and pins the
fail-closed text). The with-key escalation e2e drives a real model +
real runner + the real bridge answerer, world-verified; ci.yml snapshot
lane and e2e.yml install bubblewrap so the confined replays actually
execute.
Both RFCs move to implemented/ (Decision/Consequences form, deferred
phases tracked in their own sections), with every cross-reference flipped.
effective(session) = findLast(the session own knob events)?.value ?? the
composition-config default. One log-only event per knob, owned by its
domain (bash/sandbox-mode in dsh-bash, approval/policy in dsh-approval),
each exporting the same three-piece kit: the event declaration, a pure
fold, and THE write path — a switch IS its event; no owner service, no
facts map. Restart immunity and multi-session isolation fall out of the
log replay by construction.
Execution follows the fold on both sides: the bash tool stamps
escalation grant > session override > executor default, and the approval
seam prepends the never-gate that auto-rejects before any interactive
answerer. Visibility is two layers per knob: a per-agent prompt section
states the effective value on every request (logged through
request/header*, so what-the-model-was-told replays from the log), and an
agent/pre-step narrator injects at most one coalesced delta notice with
positional attribution (user switch vs operator/config drift). The ACP
bridge advertises one capability-gated select per composable knob with
currentValue folded per session, validates set_config_option against the
closed vocabularies, and anchors idle switches at the next turn
prompt-submit under the turn-enclosure contract.
The tool gate advertises sandbox_permissions (an enum of exactly the modes
STRICTLY WIDER than the mounted executor default — the schema makes a
non-widening request inexpressible) plus a required justification, exactly
when ctx.bash.sandboxMode reports a confining mode at registration:
composition truth, never a dead lever. An escalating call resolves
ctx.approval BEFORE anything executes with the audit-self-contained reason
"escalate sandbox to <mode>: <justification>"; allowed-once stamps the
granted mode onto that one bash request (the seam-level per-call override),
while rejected / cancelled / unavailable and the no-service / no-agent
paths each fail closed with their own error text and execute nothing. The
description teaches the flow only when the fields exist: retry the SAME
command once after a real denial, never preemptively; a rejected
escalation is final. No new session events: the attempt is an ordinary
tool/call, the decision is the approval audit pair, the outcome an
ordinary tool/result whose facts name the mode it ran under.
dsh-bash grows the per-call policy carrier: BashExecRequest.sandboxMode
(request-optional, spec required-but-nullable — the owner pattern; resolve()
is the one explicit defaulting step) and the BashExecutor.sandboxMode
capability fact (undefined in the base class — composition truth the tool
layer can read). dsh-bash-local carries the field verbatim and confines
nothing.
dsh-bash-sandbox extends LocalBashExecutor and hands ctx.sandbox the exact
argv it is about to spawn. A denial is a RESULT FACT (the command RAN;
result.sandbox.denied is orthogonal to exitCode/signal), classified
conservatively against the wrap own dialect; a RUNNER failure outranks
denial — foreground re-throws the structured SANDBOX_UNAVAILABLE, a settled
background task stamps sandbox.runnerFailed — so a broken sandbox never
reads as a failing command and the command never runs unconfined.
dsh-tool-bash renders the markers and teaches the model not to retry around
a policy denial; escalation and per-session switching are staged follow-ups.
ctx.sandbox (dsh-sandbox): confine(argv, policy) returns the argv to spawn
instead — wrapped so the process and its children run confined — plus the
enforcement completeness and the backend denial/runner-failure dialects;
no usable backend throws the fail-closed SANDBOX_UNAVAILABLE. Policy rides
per call. dsh-sandbox-local selects by platform and caches the verdict:
multi-candidate chains probe FUNCTIONALLY in preference order (Linux:
bwrap → the registry-installed node-addon-landlock-run launcher), a sole
candidate is selected unprobed (darwin: sandbox-exec/Seatbelt) and fails
closed at execution via runnerFailureSignatures; win32 is a reserved empty
chain. Profile parity is honest per backend (documented temp-area and ABI
differences; enforcement full|partial is a structured result fact).
CI: the sandbox-e2e matrix proves real-kernel confinement per rung (bwrap,
Landlock per architecture through the registry-installed launcher,
Seatbelt), failing on a silent all-skip; the packed-install rehearsal
installs the launcher family from the registry and asserts the binary
executable apart from kernel enforcement.
The ACP bridge registers the first real approval answerer: an ask for an
agent it owns becomes session/request_permission attached to the already-
streamed tool call (one-shot allow_once/reject_once only), outcomes map
conservatively (unknown optionId never grants, client cancel → cancelled),
and foreign or call-less requests delegate down the waterfall. The snapshot
harness accepts scripted permissionAnswers (FIFO; an unscripted prompt
answers cancelled, fail closed) so recorded scenarios can drive the wire
keylessly.
ctx.approval (dsh-approval): request() dispatches the approval/request
waterfall and always resolves a closed outcome — allowed-once / rejected /
cancelled / unavailable — never rejects; zero listeners fall through to
fail-closed unavailable; abort settles cancelled and discards late answers;
throwing or rogue answerers are contained as unavailable; every ask lands
the log-only approval/asked / approval/decided audit pair. dsh-tools routes
a pre-execute ask through the seam opportunistically (ctx.get) with three
distinct deny reasons, keeping the historical ask→deny degrade when the
seam is absent.
The per-session policy tier, the ACP bridge answerer, and the sandbox
escalation asker are staged follow-ups of the approval-seam RFC.
The default-relative ladder conflicts with per-session overrides: schemas
are registry-global while the effective mode is switchable, so a session
overridden NARROWER than the default loses exactly the lever it needs — a
workspace-write default with a read-only override could only request
danger-full-access (forced over-escalation), and a danger-full-access
default with a read-only override advertised no fields at all (confined,
no escalation path). The enum now pins the closed target vocabulary
(workspace-write / danger-full-access) whenever the executor confines;
strict widening is enforced per call against the session effective mode
(override ?? default), failing closed without prompting anyone. The
default-relative ladder and per-session dynamic schemas move to
Alternatives; the harmless Risks claim is corrected to name the runtime
check as the safety boundary.
The proposed forms claimed examples/sandbox-acp-agent, its recorded
scenarios, and the approval servicing of ask as already existing; a reader
of this docs-only change would look for a composition and goldens that are
not in the tree. Usage walkthroughs now name the arc the staged scenarios
are to record, both Testing sections open with the plan-time banner, and
the grounding bullet states todays degrade honestly.
agent-end was worker-authored only, so a start already forwarded to
observers lost its paired end whenever the worker could no longer speak
- the grace force-settle terminating a wedged script, or an unexpected
worker death - stranding progress consumers with agents that never
finish (ds-review-bot finding on #233).
The host now keeps a ledger of forwarded starts and funnels every
agent-end through one gate: worker-reported ends pair (and clear) their
entry, and both termination paths drain the remainder as synthesized
'cancelled' ends BEFORE the run settles, so ends always precede
workflow/end. A real settlement racing the force-settle loses to the
synthesized cancellation - the same first-wins override onResult applies
to the run's own result.
The sandbox execute wrapper JSON round-tripped the return and blindly cast it
to ToolExecuteReturn. A JSON-valid but wrong-shape return — a bare string,
{ content: 'ok' }, blocks without a type tag — sailed through: the registry
spreads result.content, so { content: 'ok' } became ['o','k'], passed the
session log's isJsonValue gate, and the DeepSeek serializer then flattened it
to '(no output)' — silent corruption of the next model request and every
replay, instead of a contained tool error.
The round-tripped value is now shape-checked against the two ToolExecuteReturn
forms (array of content blocks, or { content: blocks, meta? }); block checks
are structural only (plain object + string type tag) because the ContentBlock
union is merge-extensible. A wrong shape — and the formerly cryptic
forgot-return/bare-string cases — fails that one call with a teaching error
echoing a truncated preview of what was returned and the two valid forms.
New specs pin the object-form pass-through (meta included), six rejection
shapes, and the preview truncation; per-file 100% coverage holds.
P1 review finding: extractMeta timed only the literal's vm evaluation;
materializing the RESULT then read properties ordinarily on the HOST
stack, so a meta literal smuggling a getter (get name() { while(true){} })
could wedge the host outside any timeout — defeating the exact spin
isolation the worker thread exists for.
Rather than harden the evaluator (descriptor walks, AST validation),
delete the mechanism: the workflow's identity now reaches the seam as a
plain JSON field (WorkflowStartRequest.meta), carried by the tool as a
schema-validated `meta` object parameter the model fills directly. The
engine only shape-validates data (validateMeta, every violation named)
and pre-parses the body; the scanner, the vm evaluation, and the
host-side materialization are gone, and with them the hole. A body
still opening with a Claude Code-style `export const meta` statement
gets a pointed SCRIPT_PARSE message (the likeliest authoring slip; a
CC script's body stays drop-in, only its meta header moves into the
parameter). syncTimeoutMs now governs exactly one thing: the initial
synchronous slice inside the worker.
The RFC's decision section is rewritten in place (implemented-RFC
rule); the embedded-meta format moves to alternatives-considered with
the hole as the reason. Tool description, presentation (title now reads
meta.name directly — the textual sniff is gone), seam vocabulary docs,
and catalogs follow.
Pure mechanical rename now that the package's internals are the
worker-thread engine: directory, package name, spec/e2e filenames,
module tags and logger prefixes, tsconfig/knip/run-gates/AGENTS.md
references, example cordis.yml plugin ids, doc links; catalogs
regenerated and the lockfile refreshed.
The outer ring catches up with the engine swap (the package's own
README/JSDoc rode the port commit):
- Seam module doc and README name the worker-thread engine as THE
implementation, with isolated-vm/separate-process sandboxing as the
deferred hardening; the seam service doc states the holder-owned-runs
contract (engine-fiber disposal deliberately leaves live runs to
their holders).
- Seam contract precision: agentsStarted documents the termination-path
degradation to the host-observed count; the events section scopes the
agent-start/agent-end pair to calls that STARTED a child run;
WorkflowRun wording drops the vm-era abandonment language.
- The dynamic-workflows RFC is rewritten in place to the shipped
mechanism (implemented-RFC rule): why worker threads, the thread's
concrete buys, the in-process node:vm first cut recorded under
alternatives considered; the tool section describes the usage policy
as the tool's own prompt section.
- gen-doc-graphs: six workflow/* DYNAMIC_EVENT_DISPATCHERS entries (the
catalog no longer claims nothing dispatches them) and the seam-note
wording; core-data-structures gains its workflow.md index row;
packages/README + AGENTS.md layout line + example cordis.yml comments
say worker-thread; catalogs regenerated.
The sandbox docs overclaimed a containment contract the design never makes:
"capability access is routed through cordis services, never Node built-ins,
so everything a mounted plugin does stays inspectable and disposable". The
host-realm helpers on the sandbox global (harness, console, btoa) are
reachable functions, so mount code that goes looking can reach the host realm
through one of them — accepted under the trust stance, because the ctx a
mount ultimately receives is fully privileged anyway. Reword the sandbox
module doc, the README trust stance, and the RFC sandbox-semantics section to
say exactly that: the traps and small global surface STEER honest code onto
the cordis services; they are not a security boundary.
The master merge (5309ea54) resolved the module-graph conflict by hand,
placing the subagent-subprocess dependency-table row ahead of util/timeout's;
the generator's deterministic order (group order, util first) wants them
swapped, so the freshness gate (gen-module-graph --check) failed CI's static
job. Regenerated on the merged tree — a two-line swap; every other generated
catalog was already resolution-fresh (regen-all changed nothing else).
Two review findings (#220) on the sandbox context façade:
- Undeclared services were reachable: the façade resolved any live global via
ctx.get(name), so ctx.bash worked without inject: ['bash']. A cross-mount
consumer could then depend on a provider cordis never saw — unmounting the
provider would neither park the consumer nor unwind its registered tools,
leaving a model-visible tool that fails only at execution. The façade now
reads ctx.fiber.inject and refuses any service the mount did not declare
(with a teaching error naming the inject fix), so the dependency is always
visible to cordis and its activation/unload semantics bind.
- ctx.tools.get returned the live ToolDefinition, including execute — mount
code could call another tool directly and bypass ToolRegistry.execute and
its pre/post-execute hooks and accounting. get now returns the same
read-only name/description/parameters view as schemas(), never an invocable.
Adds inject-gate and schema-view regression cases to sandbox-context.spec.ts
(undeclared property/get denied, declared allowed, the cross-mount zombie-tool
scenario refused at call time, get exposes no execute). Package stays at
per-file 100% coverage. RFC, mount description, and tool-catalog updated.
Review finding (#220): the guarded proxy only special-cased ctx.tools, so
mount code could reach an UNGUARDED context through ctx.root, ctx.extend(), or
a service instance's .ctx, then ctx.root.tools.register({…}) to bypass the
marker check and host-realm normalization — a raw vm-realm result would later
error a real agent turn at the session-log plainness check.
The sandbox ctx is now a whitelist façade, not a pass-through proxy: it exposes
only what a mount needs — tools.register (marker-guarded), on/once, provide, the
timer helpers, and injected services resolved through a guarded get — and denies
every framework-plumbing member (root, parent, fiber, reflect, registry, extend,
isolate, intercept, plugin, set, mixin, …) with a teaching error. Injected
services are wrapped so a method returning a Context is rejected on the way back
(the .ctx escape), closing the one indirect leak. There is no context-valued
member left to reach; cross-mount provide/inject is untouched (the plugin's own
inject and the fiber's pending/active gating are unchanged). ctx.plugin (child
plugins) and ctx.set are denied by design; ctx.effect is deferred (FIXME).
Adds tests/sandbox-context.spec.ts covering the escape class (root/extend/fiber/
plugin/set/… denied, the classic root.tools.register bypass, the .ctx escape,
read-only writes) plus the async-service and symbol/in-operator paths for 100%
coverage. RFC/README/tool-catalog/config-catalog updated; api-catalog.ts
regenerated (also picks up the codeRuntime service that entered on the master
merge and was left stale).
The tree SHAPE was the wrong surface for the model: what it needs from
cordis_inspect is what services, plugins, and capabilities are loaded, not the
fiber hierarchy. The plugins section is now a flat name + lifecycle-state list
from ctx.registry (deterministically sorted, one line per instance); the ASCII
tree renderer, the parent→child rebuild, and the dyn-id tree annotation are
deleted — dynamic mounts keep their own richer dynamic section (id, state,
provides, waits). Net -49 lines; RFC and READMEs state the flat-list contract.
Field sessions showed models writing tool schemas in the JSON-Schema dialect
by strong prior — type: 'integer', required: false, then the full
{ type:'object', properties, required: [...] } wrapper — and the rejection
text itself pushed a nearly-correct DSL attempt BACK to raw JSON Schema: one
stats tool cost three consecutive schema errors before mounting. The boundary
now normalizes wherever the input has exactly one meaning (wrapper unwrapped
with the required array becoming per-property flags at any nesting level,
integer → number, required: false → optional, all rebuilt as fresh host-realm
objects) and rejects only genuinely meaningless input, enumerating the valid
vocabulary in the error. Re-running the failing session mounts first-try.
The mount description documents both accepted forms.
The sandbox deliberately provides no Node API, and now says so instead of
letting a bare ReferenceError teach nothing: require, the timers, and fetch
are callable traps whose error redirects to the cordis alternative (inject:
['fs'] + ctx.fs, ['web'] + ctx.web, ['bash'] + ctx.bash, ['timer'] +
ctx.setTimeout — a fiber effect, unwound on unmount). Only function-shaped
globals are trapped; process/Buffer stay undefined so typeof feature probes
stay inert. The mount description and the demo persona state the routing rule,
and the demo mounts ctx.fs (local provider) and ctx.web (seam + keyless local
fetch provider) so agent-built plugins have real capabilities to build on.
Live-validated: a model that reached for Node setTimeout self-corrected to
inject: ['timer'] in one step and built a working ctx.web fetch tool.
Everything outside the package and example that a new top-level group and a
new demo touch: GROUP_ORDER in gen-module-graph and gen-doc-graphs (plus the
tools-service consumers list, the APP_EXAMPLES entry, and the graph-atlas
label/mode rows), the knip e2e entries, the packages/README group row, the
AGENTS.md layout and demo lines, and the regenerated module-graph /
config-catalog / graph-atlas / capability-seams / composition artifacts.
AGENTS.md and examples/AGENTS.md word-budget ceilings rise to current+5%
(1802 / 653): the new group and demo rows are genuine additions to both docs,
not condensable restatements.
New top-level packages/cordis/ group with the self-referential toolset:
cordis_inspect (services / plugin tree / tools / dynamic mounts / api / events,
the api section intersecting the generated catalog with the live service store),
cordis_mount (model-written code evaluated in a node:vm sandbox, mounted under
one cordis-dynamic group fiber as dyn-<n>), cordis_unmount (awaited disposal to
quiescence). Boundary mechanisms: dual-realm instanceof, JSON realm
normalization of dynamic tool results, marker-guarded registration, SchemaSpec
teaching errors, parse failures surfaced with the offending line + caret and a
line-scoped TypeScript hint, and the unmount-first recipe on tool-name
collisions. Config: vmTimeoutMs (schemastery, default 5000). Design record:
docs/rfc/implemented/feature/2026-07-08-self-referential-cordis-toolset.md.
The tool-catalog boot manifest, its regenerated output, and the pinned
tool-name list land here rather than with the other repo registration: the
completeness guard globs packages/*/tool-* and fails the generator (and the
core/tools spec) the moment the package directory exists.
The design record for tool-cordis: the three-tool contract, the vm sandbox
trust stance and boundary mechanisms, the dynamic-group lifecycle, cross-mount
provide/inject composition, the generated runtime API catalog, and the
alternatives weighed (per-capability registration tools, hand-maintained API
tables, a mount provenance event, a hardened sandbox).
The extracted library's name sat one edit away from @deepseek-ai/dsh-subagent-inprocess
(process/inprocess), inviting a typo'd import to silently resolve to the wrong
package. subagent-subprocess also reads as the deliberate counterpart to
subagent-inprocess (in-process vs. subprocess), matching how the two shared
drivers actually differ.
Package directory, npm name, module doc, JSDoc module tags, test-file name and
its temp-dir prefixes, the subagent-acp import and its Config/tsconfig/package.json
references, root tsconfig.json/tsconfig.build.json/knip.json entries, and the
packages/subagent group README all renamed together; regenerated
docs/module-graph.md and docs/config-catalog.md. Pure rename — no behavior,
export, or Config shape changed.
