The Codex simplification concern plus the duplication comment on the spawn
apply, resolved by deletion: the backend-lifetime holds are gone, so the
runtime registers at the first structured run and disposes when the last
settles — a deployment that never passes outputSchema carries no always-on
global state, and there is no per-backend acquisition block left to extract.
The driver spec now drives an INLINE spawn-shaped provider over
startInProcessRun, which removes the spawn/fork devDependencies (the
test-only workspace cycle); plugin-level structured coverage moves to the
backends' own specs (capture through the shipped plugin, mid-run backend
unload, seeded fork capture). tools.md, the driver README, and both backend
READMEs describe the run-scoped lifetime; the module-graph regenerates
without the cycle edges.
The cross-seam blocker: structured_output recorded its value in the tool
BODY, before tools/post-execute could block the call — a PostToolUse hook's
block turned the logged result into isError while readResult still returned
structured success and the continuation veto ended the turn.
Two-phase commit: the body validates and STAGES (RunState.pending); a fourth
runtime listener on tools/post-execute — prepend, so await next() returns the
composed final decision — promotes the stage to captured only on an accepted
call, and clears it on every path. A block now yields a consistent pair: the
model and log see the isError feedback, the run settles error with no
structured value, and the turn continues so the model can react. Regressions:
block denies the capture end-to-end; accept-with-replacement still commits.
Three Codex findings on json-schema.ts, one discipline:
- required-declared and every value check now use Object.hasOwn — 'in' let
inherited names (toString) satisfy required, dodge additionalProperties:
false, and validate a declared property against the value's prototype
member instead of a carried one
- isObjectLike now means PLAIN JSON object (proto chain of at most one link,
realm-agnostic): a Date annotation or a Map-as-properties no longer passes
structurally and serializes lossily — they fail loud as subset violations
- startInProcessRun asserts BEFORE the defensive structuredClone, so a
hostile schema fails as OutputSchemaError, never a raw DataCloneError
Also the type-equiv catalog gap: tools.md gains the structured-output subset
vocabulary (4 blocks) with matching manifest entries. The driver index also
drops the runtime internals from its public re-export (runs acquire it
internally; no external consumer remains — see the following commit).
Two human review directives:
- No re-prompt. A structured child that finishes a turn cleanly without
calling structured_output settles error to the parent immediately —
readResult already carried that mapping; the nudge loop only delayed it.
Deletes the loop, its cancellation-window guard, STRUCTURED_OUTPUT_NUDGE,
and the structuredNudgeRetries Config on both backends.
- FIXME in the structured module doc: per-agent/per-session tool registry and
prompt assembly would dissolve the final-assembly enforcement dance (the
placeholder tool, the swap, the strip, the global-registration lifetime).
Two bot findings on the structured runtime:
- Terminal means terminal WITHIN the step: a model response listing
structured_output before further tool calls executed those calls after the
final answer was accepted (the turn-continuation veto only fires at step
end). A third runtime listener now denies every call for a captured agent
at the tools/pre-execute gate — dispatch skipped, isError result naming the
contract. Calls preceding the capture in the same response are untouched.
- The output schema is structuredClone'd before the subset assertion: the
caller keeps its reference, so asserting and attaching the original let a
post-start() mutation drift the enforced schema away from the asserted one.
The clone pins assertion, model-visible parameters, and validation to one
value.
Carved out of #170 per review feedback — the foundation the workflow tool
builds on, now standing alone on master:
- dsh-tools: the structured-output JSON Schema subset (StructuredOutputSchema,
assertSupportedOutputSchema, validateStructuredValue) — rejects loud outside
the enforced subset, listing every violation
- dsh-subagent: SubagentStartRequest.outputSchema / SubagentResult.structured
become a real capability; the service rejects a schema'd request whose
provider lacks it
- dsh-subagent-inprocess: the shared structured runtime — one global
structured_output capture tool, a prepend final-assembly listener that
strips the placeholder for plain agents and swaps in the run's own schema
(plus the calling instruction as a trailing section) for structured
children, an agent/turn-continuation veto once captured, and the
capture/nudge loop in the run driver (structuredNudgeRetries, cancellation
honored mid-nudge); lifetime refcounted by backends and live runs
- subagent-spawn / subagent-fork flip outputSchema: true
One deliberate divergence from the #170 revision: the backends do NOT add
'tools' to their plugin inject. Doing so deferred their apply past the todo
plugin, and the delegation tool mirrors provider lifecycle — so the
model-visible tool order of every existing prompt changed, invalidating every
recorded snapshot fixture. The runtime now gates its capture-tool registration
on tools availability itself (sync when live, a scoped inject fiber when the
Loader starts the backend first), keeping this PR byte-invisible to existing
transcripts: all 35 snapshot scenarios pass against master's fixtures
unchanged.
New doc-sync gate verify-export-jsdoc walks every module-level exported
name under packages/*/*/src and requires description prose everywhere,
plus @param per parameter and @returns on non-void annotated returns for
function-like exports, public class methods, properties, and accessors.
The parsing + check helpers move out of gen-cordis-catalog.ts into a
shared scripts/jsdoc.ts so 'documented' means one thing on both gated
surfaces.
Deliberate exemptions (documented in the RFC): heritage-declared class
members (the seam declaration is the doc's one home — the one checker
query in an otherwise pure-AST walk), cordis plugin-protocol slots
(name/inject/reusable/Config/apply, top-level and static), constructors,
overload implementations, declare-module augmentation bodies, and
re-export statements (checked at the defining module).
The 203 under-documented exports the gate found at adoption are filled
in this change, so the gate lands green; generated catalogs/graphs are
regenerated for the shifted line pointers.
RFC: docs/rfc/implemented/process/2026-07-06-export-surface-jsdoc-gate.md
ds-review-bot round 2, both warnings:
- subagent/provider-removed now routes through emitLifecycle (per-listener
containment, the subagent/start|end precedent) instead of raw ctx.emit,
whose dispatch halts on the first throw: a throwing subscriber can no
longer starve a later mirror into keeping a stale tool, nor disrupt the
backend fiber's teardown mid-disposer. provider-added deliberately keeps
propagation (register-time rollback semantics, like the system-prompt
registries); the asymmetry is documented on emitLifecycle, the event
JSDoc, and the provider-lifecycle RFC.
- The documented model-via-agent/request fallback composes with a
{{model}} persona via the ownership rule itself: the plugin supplying
the model late states it early on the system-prompt/assemble waterfall.
Declined re-ordering render after agent/request — it would break the
agent/pre-step contract (compaction must measure the prompt the model
sees). New loop test pins the supply path end-to-end; the RFC's
{{model}} consequence bullet now covers supply as well as switch.
Review round 2 (tianyicui inline comments):
- dsh-system-prompt itself registers the harness:identity (-100) and
deployment:persona (0) sections — they must survive a swapped loop
plugin, so they leave dsh-agent-loop; the persona text is the plugin's
own validated 'persona' config. The model/cwd variables STAY on the
loop: runtime facts of the agents it drives.
- AgentOptions.systemPrompt is deleted with all its forwarding plumbing:
the app configs' systemPrompt keys become 'persona' routed through
dsh-agent-core (schema = z.intersect of the owners'), the ACP bridge
and tool-subagent stop carrying persona configuration, and subagent
children now render the deployment persona like every other agent.
- Example personas drop transport/interface trivia (ACP, CLI) — facts
irrelevant to the model.
- Root CONTEXT.md removed (not idiomatic); its persona definition was
wrong under the new ownership anyway.
- Docs, READMEs, the prompt-variables RFC, and generated catalogs
updated; new loop test pins the assemble-waterfall escape valve
(an emptied assembly sends NO system field).
Codex round-2 findings:
- A committed test now proves the REAL plugin fiber (not a direct apply)
owns the provider-lifecycle listeners: disposing a mounted tool's fiber
unmounts the tool and leaves the provider intact, and a fiber disposed
while WAITING never zombie-mounts when its provider arrives later.
- TODO(subagent-dup-toolname) records the invalid-config blast radius of
two waiting fibers sharing a toolName (the duplicate throw propagates
through subagent/provider-added and rolls back the provider).
- CONTEXT.md drops its creation-history sentence; the RFC's acceptance
checklist becomes present-tense shipped invariants (docs/AGENTS.md
writing rules).
Codex round-1 findings, both confirmed:
- renderPrompt: variable lookup now uses Object.hasOwn (an unregistered
{{constructor}} previously resolved through Object.prototype and spliced
function source into the prompt), and a {{ that opens no complete group
while a }} still follows ({{{model}}}, {{a{b}}) now throws instead of
passing or partially interpolating. A lone {{ with no }} after it stays
verbatim; substituted values are never re-scanned.
- tool-subagent: the apply-time provider lookup assumed a load order the
cordis Loader does not guarantee (siblings start concurrently). The seam
now announces subagent/provider-added/-removed and the tool mirrors the
provider's lifecycle: registers when the provider is (or becomes)
available, unregisters when it goes away, re-derives wording on reload.
No load-order requirement remains.
- loop.spec containment test now proves live continuation: after the
contained render failure, a waterfall listener rescues {{cwd}} and the
same agent completes a real model turn.
RFC/READMEs updated to the shipped contract; cordis catalog regenerated.
One principle: every fact in the assembled prompt has exactly one owner.
- dsh-system-prompt: merge-extensible AssembleContext on assemble();
a variable(name, provider) registry; {{name}} interpolation in
renderPrompt, strict (unknown/valueless/malformed references throw);
duplicate section and variable names rejected; assembly carries
resolved section text + variables through the assemble waterfall.
- dsh-agent declares AssembleContext.agent; dsh-agent-loop registers
the agent:persona section (order 0 - identity renders before tool
guidance) and the model/cwd variables, and drops its string join:
renderPrompt(assembly) IS the full prompt.
- Tool guidance moves to its owners: descriptions carry per-tool
semantics; sections only cross-call habits (tool:bash exit-code
habit at order 105; read's not-shell nudge). todo/subagent need no
section - their descriptions already carry the contract.
- SubagentProvider.inheritsParentContext (spawn/acp false, fork true);
dsh-tool-subagent derives truthful per-provider wording and resolves
the provider at load (backend must be listed first).
- Example personas shrink to identity + behavior with {{model}} (and
{{cwd}} in the ACP tree); the welcome banner stops enumerating tools.
RFC: docs/rfc/implemented/architecture/2026-07-05-prompt-variables-and-tool-guidance-ownership.md
gen-cordis-catalog now hard-errors (aggregated, not fail-fast) when an
event lacks description prose or a payload @param, or a public service
method lacks JSDoc, a @param per parameter, a @returns on a non-void
result, or an explicit return type annotation. The this receiver and the
trailing waterfall next are exempt on events (mode machinery owned by
@mode); a stale @param naming no real parameter errors, mirroring the
@mode contradiction check. parseJsDoc now ends prose at the first block
tag (standard JSDoc semantics), so the tags never change the rendered
catalog — only Source: line pointers moved.
Fills the ~139 gaps found across the 15 surface files, extends the spec
with negative-path fixtures for every new guard plus the exemptions,
records the decision as an implemented process RFC, and extends the
AGENTS.md typed-events bullet with the authoring rule. Runs inside
verify-cordis-catalog -> doc-sync, so CI and pre-push enforce it with
zero new wiring.
A Codex review pass on the draft caught four real gaps and two solid
suggestions; all addressed except one pushed back on the merits:
- hooks-claude/hooks-codex: stderrSummaryMaxChars was the one new knob
with NO range validation — a negative/NaN cap would silently
misbehave inside slice(). Both bridges now assert a positive integer
at the TOP of apply() (before the config-file parse's early return,
so a bad value fails the load loudly), with rejection tests.
- tool-fs: the read caps count lines/chars/bytes, so positive-FINITE
was too loose (a fractional readLimit would flow into windowing
arithmetic and the schema description). All four now require a
positive integer, matching tool-web's cap.
- Doc drift the gates cannot catch: tool-web's README tools table
still named WEB_SEARCH_MAX_RESULTS as the mechanism; compact-basic's
README/module doc and the compaction-capability-seam RFC still
described estimation as fixed char/4 rather than the charsPerToken
default.
- subagent-acp: the dispose graces were tested only at the
startAcpRun level, so a regression that stopped threading plugin
config into AcpRunSpec would have survived. A provider-path test now
drives the trap-escalation scenario through ctx.subagents.start with
small config graces and bounds dispose at 4s.
Pushed back on: converting compact-basic's charsPerToken to a
schemastery field. The package's whole config is deliberately
hand-rolled (resolveConfig, every threshold REQUIRED with no default —
a documented design posture); one schemastery field beside it would be
incoherent. The knob is cordis.yml-reachable, defaulted, and validated,
which is what the convention requires; migrating the package to
schemastery wholesale is pre-existing config-surface hygiene out of
this change's scope.
The audit swept every packages/*/* plugin for the new AGENTS.md
convention (no hardcoded tunables in plugins) and exposes each finding
as a defaulted, validated Config field. Defaults are the previously
hardcoded values throughout, so no deployment or golden changes.
- tool-fs (had NO Config): readLimit, readMaxLineLength, readMaxBytes,
readStreamMinSize. The caps thread through ReadToolCaps/ReadWindow —
read-render already documented that the consumer applies the caps, so
they become explicit per-request fields.
- tool-web: searchMaxResults (WEB_SEARCH_MAX_RESULTS stays as the
schemastery default). Also fixes the stale GREP_LIMIT references in
search.ts and the web-capability-seam RFC (no such constant exists).
- bash-local: graceMs (SIGTERM->SIGKILL escalation grace). The
RunInternals.graceMs test seam is gone: graceMs is now a required
SpawnSpec field filled from config, so tests exercise the real
config path and the defaults live in exactly one place.
- subagent-acp: disposeEofGraceMs / disposeGraceMs. The AcpRunSpec
fields become required for the same one-defaulting-layer reason.
- session-persistence-sqlite: journalMode ('wal' default; the
rollback-journal modes serve filesystems where WAL's shared-memory
files do not work, e.g. network mounts).
- hooks-claude + hooks-codex: stderrSummaryMaxChars for the persisted
hook/result stderr summary. The duplicated summarize() helpers merge
into hook-protocol's summarizeStderr(stderr, maxChars), beside the
HookResultRecord field it feeds, with the bound parameterized the
same way runHook's defaultTimeoutMs already is.
- compact-basic: charsPerToken for the token estimator (default 4, the
English-text heuristic; CJK-heavy deployments need ~1-2 or compaction
fires far too late). Also corrects the BasicCompactService class doc,
which claimed defaults the required-field config never had.
- fs-local: deletes the dead STREAM_MIN_SIZE constant and the dead
FsIoInternals.streamMinSize seam — the read-routing bound lives in
the consumer (tool-fs), where it is now config. This is item 1 of
the proposed prune-write-only-fs-surface RFC, annotated accordingly.
Every new field gets range validation (following the existing
assertPositiveFinite pattern), a README row, and tests covering the
configured behavior, the schema default, and load-time rejection.
Address review: `agentType` was a Claude-Code concept (`subagent_type`) that
does not fit our own subagent seam — nothing in the harness interprets it, and
its only consumer was the CC-dialect hook bridge. Rather than let a foreign
concept sit on the core seam, remove it:
- `SubagentStartRequest`, `SubagentRunInfo`, `SubagentRunEndInfo`: drop the
`agentType` field; the `subagent/start`/`subagent/end` payloads now carry
`provider`/`id` (+ end `stopReason`/`lastAssistantMessage`) only.
- `dsh-tool-subagent`: drop `Config.agentType` and its request plumbing.
- Tests: keep the lastAssistantMessage / clone-containment / reject-path
coverage (rewritten to not assert agentType); delete the two tool-subagent
tests that only exercised agentType forwarding (dead behavior).
- Docs: retitle + rewrite the subagent-observe-enrich RFC to the one shipped
enrichment (lastAssistantMessage), with a note on why agentType was dropped;
update rfc/README index title, both subagent READMEs, and the
core-data-structures/subagent.md type-equiv block + prose; regenerate catalog.
The CC bridge (PR-F) will feed Claude Code's own default matcher value
"general-purpose" for its SubagentStart/Stop agent_type matcher instead.
Review noted the deep-clone of the child output runs inside `onFulfilled`,
OUTSIDE emitLifecycle's per-listener containment, and the settle `.then` is
`void`ed — so an uncloneable output (a future non-serializable content-block
type, or a contract-violating result) would throw and become an UNHANDLED
rejection, contradicting the "any throw is contained" guarantee the comment
claims. Wrap the clone in try/catch: on failure, log via ctx.logger.warn and
emit subagent/end WITHOUT lastAssistantMessage (preserving stopReason/agentType)
rather than dropping the event or crashing. Regression proves the unfixed code
produces an unhandled rejection.
Codex review caught an observe-only violation: the subagent/end emit fires from a
detached `.then` registered BEFORE start() returns — so before the caller's own
`await run.result` continuation runs. Carrying `result.output` by reference let a
mutating subagent/end listener corrupt the SubagentResult.output the caller/tool
then consumes. structuredClone() makes the event a read-only snapshot. Added a
regression test that mutates the event's array and asserts the caller's result is
untouched; proven to fail red without the clone. Updated the RFC + READMEs to note
the clone is load-bearing for the observe-only guarantee.
A hooks bridge translating SubagentStart/SubagentStop needs to know WHICH kind of
subagent ran and WHAT it produced — Claude Code's hooks carry subagent_type and the
child's final message. Enrich the existing lifecycle emits to match, observe-only:
- agentType: an optional caller-supplied subagent-kind label (CC's subagent_type),
added to SubagentStartRequest and carried VERBATIM onto both subagent/start
(SubagentRunInfo) and subagent/end (SubagentRunEndInfo). The seam never interprets
it. dsh-tool-subagent threads it from a new optional Config.agentType, so a
deployment exposing multiple subagent kinds (one tool load per kind) labels each.
- lastAssistantMessage: the child's final output (SubagentResult.output), added to
SubagentRunEndInfo on the settle path so an observer sees what the subagent
produced without holding the run. Absent on the reject path (no result produced).
Strictly observe-only: both events stay plain emits (subagent/end fires from a
detached .then and awaits no listener). A control-flow subagent/end (awaited
waterfall returning a decision) would need the emit→waterfall reshape, awaiting
listeners before settling, and a provider resume capability — deferred to the
background/steering redesign (FIXME(subagent-continuation) anchors it). RFC:
implemented/feature/2026-06-30-subagent-observe-enrich.md.
Reshape the agent's interception surface so every seam returns a small, typed
Decision union, and the set covers the hook points a CC/Codex bridge (and a
native plugin) needs. "Native hooks" are not a package — a native hook is just a
cordis plugin on these canonical events; the bridges (a later PR) only translate
an external protocol onto the same surface.
dsh-agent:
- NEW agent/session-start(agent, source) emit (once before turn 1; SessionStartSource
startup|resume|clear|compact) — a pure notification, seeds context via inject().
- NEW agent/prompt-submit waterfall → PromptDecision (allow, optionally rewriting the
prompt or attaching additionalContext, or block).
- RESHAPE agent/turn-continuation boolean → ContinuationDecision ({action:'stop'} |
{action:'continue', reason?}; a continue reason is recorded as next-step steering).
- New HookContext envelope (required source — inject() would mislabel a missing one).
dsh-tools: split the single tools/execute waterfall into tools/pre-execute
(PreToolDecision allow/deny/ask gate) and tools/post-execute (PostToolDecision
accept/block, optionally replacing content or attaching additionalContext). Core
dispatch sits between as plain code; the tool body keeps its inner try/catch so a
thrown tool still reaches post-execute as an isError. ToolExecutionResult gains
additionalContext (ferried to the loop's per-step buffer). Input rewrite is
deliberately NOT offered (a proposed RFC designs it consistently).
dsh-session: new `rejected` TurnEndReason — a turn whose whole prompt batch was
blocked by prompt-submit.
agent-loop firing points: session-start emitted at create (source threaded —
startup for create/fork, resume for resume()); prompt-submit per drained message
with the always-open-turn rule (a fully-blocked batch is a zero-step rejected
turn); the continuation reshape; post-tool additionalContext buffered and appended
after all tool/results (adjacency). ACP codec maps rejected→cancelled.
A worked native-plugin example (interception.spec.ts) proves all four seams compose
end-to-end through the real loop with NO hook/* events (those belong to the bridge
lib). All existing tools/execute + turn-continuation tests migrated. The
tool-subagent abort test now aborts after a microtask so it still exercises the
live onAbort bridge (execute() awaits pre-execute before the body runs).
RFCs: implemented/feature/2026-06-30-interception-seams.md (the reshape) +
proposed/feature/2026-06-30-pre-tool-input-rewrite.md (the deferred rewrite design).
A fork subagent seeds its child session with a prefix of the parent's log, and
that seed becomes the child's persisted log — so a fork child's .jsonl begins
with the PARENT's events, including the parent's assistant/chunk events. The
snapshot replay harness derived a child's script from its whole log, which would
replay the parent's recorded responses as the child's model calls. Spawn-only
scenarios never hit it, but a fork snapshot would mis-route silently.
Record the seed boundary and skip the inherited prefix at replay:
- SessionHeader gains an optional `seedLength` (how many leading events were
inherited via a seed), threaded through CreateSessionOptions/CreateAgentOptions
meta and stamped by the fork backend (= seeded-prefix length; absent for spawn).
It is EXPLICIT, never inferred from seed.length: a resume seeds the whole stored
log, so the resume path passes the persisted boundary back.
- Both persistence backends round-trip it: JSONL header line, SQLite seed_length
column. The SQLite table change bumps SCHEMA_VERSION 2->3; per the pre-release
stance the backend rejects an older user_version on open with NO migration.
- llm-replay's parseSessionHeader reads seedLength and loadSessionScripts derives
a child script from events AFTER the boundary. seedLength is 0 for spawn, so
spawn replay is byte-for-byte unchanged.
Closes the routing-correctness gap the per-session snapshot replay RFC under-
stated; a recorded fork scenario remains a future addition but now derives
correctly. RFC: docs/rfc/implemented/testing/2026-06-22-fork-child-replay-seed-boundary.md.
Regression coverage: a fork child fixture whose seeded prefix carries a parent
chunk (derived script must exclude it, proven red without the slice); a seedLength
persistence round-trip through the shared coordinator contract (both backends);
the fork backend stamping it; resume preserving it from the persisted header.
Two findings on the ACP backend:
Blocking: cancel() only sent session/cancel, so a child that ignores the notify
or wedges the prompt left result hung forever — the model-facing tool awaits
result before its finally disposes, so the parent cancellation hung and the
child stayed alive, violating the SubagentRun.cancel() contract (result settles
aborted). The result path now races the ACP drive against a cancelSettled
promise that requestCancel resolves, so result settles aborted the instant a
cancel is requested, regardless of the child. dispose() still kills+reaps the
process. New MOCK_IGNORE_CANCEL mock mode (receives cancel, never resolves the
prompt, never exits) drives a regression proven to hang without the race.
Nit: the drive-path catch was an empty broad catch that discarded the error
(AGENTS.md forbids). Because cancellation is now handled by the race arm, a
rejection reaching the catch is always a genuine child-level error — bind it,
flatten to error, and surface the original via a new AcpRunSpec.onError sink
that the provider wires to ctx.logger.warn, so a real fault is preserved.
addEventListener('abort') does not fire for a signal already aborted before the
listener is added, so a parent step cancelled before the subagent tool ran
would never reach the child — the tool leaned on each provider re-checking
request.signal itself, leaving the bridge's own claim incomplete for any
provider that relies on run.cancel(). Re-check exec.signal.aborted right after
registering and cancel explicitly. Regression test uses a spy provider that
only reacts to cancel() (never inspects the signal); proven to hang without the
fix (result never settles) and settle aborted with it.
Master's #36 moved declaration output to lib/types (and types/exports/files
point there). The merge applied that to all pre-existing packages, but the
subagent backends introduced on this stack (subagent-inprocess, subagent-spawn,
subagent-fork) still used the old lib/ layout. Bring them onto the new
convention and add them to the single typecheck tsconfig.json references.
The shared run driver lived inside dsh-subagent-spawn, so the spawn package
carried fork-aware seeding logic and dsh-subagent-fork depended backward on
dsh-subagent-spawn — the two in-process backends were not independent.
Move the driver (startInProcessRun, depthOf, SubagentDepthError,
InProcessRunOptions) into a new pure-library package
@deepseek-ai/dsh-subagent-inprocess that registers nothing. spawn and fork now
both depend only on that driver and neither knows about the other; spawn no
longer re-exports it and fork no longer imports from spawn.
Also wire BOTH backends in examples/coding-agent/cordis.yml (config-only): load
dsh-subagent-spawn + dsh-subagent-fork + two dsh-tool-subagent instances with
distinct toolNames (subagent → spawn, subagent_fork → fork), demonstrating that
exposing multiple transports needs no code change.
Two round-3 findings:
(A) The EOF-quiesce window reused the 3000ms SIGTERM grace, the SAME value as
dsh-bash-local's own SIGTERM->SIGKILL grace. The child acp-agent's EOF teardown
disposes its loop, which stops child-owned bash -- and a SIGTERM-trapping bash
grandchild can hold that for up to ~3s before its own SIGKILL, then the child
still owes a final flush. With both graces equal, the parent's SIGTERM fired
exactly as the child reached its own SIGKILL+flush, cutting it off. Split the
EOF grace into its own knob (disposeEofGraceMs, default 6000ms) that exceeds a
single signal-grace of nested-teardown headroom. The child is an arbitrary ACP
agent, so the value is a standalone generous default, NOT derived from any
child's internals. Tier-1 test now uses a flush that outlasts the SIGTERM grace
but fits the EOF grace, so it lands only because the EOF tier honors its own
wider window (proven RED when tier 1 reuses the small SIGTERM grace).
(B) The middle-tier (SIGTERM) test only asserted dispose returned in time, so
an EOF->SIGKILL ladder with the rung removed would still pass. The mock's
MOCK_IGNORE_EOF mode now installs a SIGTERM handler that touches an observable
marker before exiting; SIGKILL is uncatchable, so removing the SIGTERM rung
leaves the marker absent (proven RED). The test asserts the marker exists.
dispose() ended stdin and sent SIGTERM in the same tick, so the child's
EOF-driven quiesce had no window to run. The real acp-agent has no SIGTERM
handler in a normal session — it flushes persistence and stops child-owned
work via the server bridge's connection-close path (conn.closed → per-agent
dispose → final session/flush), driven by stdin EOF, NOT by a signal. A prompt
response can resolve from a turn/end before that post-turn flush lands, so the
child still owes durable work when dispose runs; a same-tick default SIGTERM
terminated it mid-flush, orphaning child-owned bash and dropping the flush.
dispose now waits for the child's natural exit after stdin EOF first, then
escalates SIGTERM (grace), then SIGKILL — a three-tier ladder. Add an
`exitsWithin` helper for the bounded waits.
Regression coverage: a new mock mode (MOCK_FLUSH_ON_EOF) flushes a marker
asynchronously on EOF then self-exits; the tier-1 test asserts the marker
lands (proven RED on the same-tick-SIGTERM ordering — child killed mid-flush).
MOCK_IGNORE_EOF covers the middle tier (ignores EOF, dies on default SIGTERM);
the existing MOCK_TRAP_SIGTERM test covers the SIGKILL tier.
Two lifecycle findings from the review:
- A (blocker): dispose() could hang forever. It only sent SIGTERM and awaited
exit, with no escalation — a child that traps SIGTERM (or our acp-agent if it
doesn't quiesce on stdin EOF) would wedge dispose, stranding tool-subagent's
finally cleanup and orphaning child-owned work (e.g. bash subprocesses). dispose
now: ends stdin (graceful ACP close so the child can flush + exit), SIGTERM,
then escalates to SIGKILL if it doesn't exit within a grace period
(DEFAULT_DISPOSE_GRACE_MS, injectable via spec.disposeGraceMs), awaiting the
certain exit. Mirrors the bash executor's bounded teardown. Regression test
drives a SIGTERM-trapping mock subprocess and asserts dispose returns promptly
— proven to hang (red) without the escalation.
- B: an already-aborted request still spawned the configured binary. startAcpRun
now returns an inert already-aborted run BEFORE spawning, so a pre-cancelled
request launches nothing. Test points the command at `touch <sentinel>` and
asserts the sentinel never appears.
The dispose regression test exposed (via systematic-debugging) that the child
must signal trap-armed readiness before the test cancels — a bare timeout raced
the trap install and the default SIGTERM handler killed the child, making the
guard a no-op. The mock now touches its ready file once the trap is in place and
the test waits on that condition. The `cancelled` flag moved onto a holder object
so TS control-flow doesn't narrow the catch-time read to always-false.
The first OUT-OF-PROCESS subagent backend, proving the seam generalizes past the
in-process backends. @deepseek-ai/dsh-subagent-acp runs each child agent in a
spawned subprocess, driven over the Agent Client Protocol as the CLIENT — the
direction-inverted twin of the dsh-acp server bridge. Point the configured
command at the acp-agent example and the harness talks to its own process.
- Fresh process per run: start spawns, runs one ACP session (initialize →
newSession → prompt), dispose kills the subprocess and awaits its exit.
- Minimal client stub: advertises no fs/terminal; accumulates agent_message_chunk
text as the result output; auto-answers session/request_permission by a
configured policy (reject default / allow). No start-time capabilities (an
out-of-process child can't enforce the parent's depth/tool-filter); ignores
request.parent; injects only `subagents`.
- StopReason mapping (end_turn→completed, cancelled→aborted, …); result resolves
error/aborted on a child failure, never rejects (seam contract).
- Security: credential-shaped ambient env vars are scrubbed; the child's own key
is forwarded only via explicit config.env. A spawn-level error (ENOENT) is
captured and raced against the ACP drive so a bad command settles error rather
than crashing the parent.
Testing designed at every tier: keyless integration drives a scripted mock ACP
server subprocess (cancellation incl. the pre-newSession race and a
torn-pipe-after-cancel, permission auto-answer, non-message updates, spawn
failure, HMR, export shape) at 100% coverage; a with-key e2e drives the REAL
acp-agent example process (PONG + real file write, verified on disk) — the
harness driving itself. Snapshot coverage of an ACP child is deferred as
TODO(acp-subagent-replay) (each child is its own process with its own replay).
Stayed on @agentclientprotocol/sdk 0.25.1: the proposed 0.28.x bump only
deprecates the stable ClientSideConnection/AgentSideConnection API this layer
uses (33 sites incl. the server bridge), turning no-deprecated red across code
this PR shouldn't rewrite — that fluent-API migration is its own follow-up. The
backend needs nothing 0.28.x adds.
This completes the subagent seam stack (PR1 interface → PR2 in-process → PR2.5
snapshot infra → PR3 ACP); the seam RFC moves to implemented/, amended.
A request signal aborted BEFORE the run starts never fires an `abort` event
(`addEventListener` only fires on the transition), so the backend-level bridge
missed it and ran the child to `completed`. The driver now checks
`request.signal?.aborted` at the top of the result path and settles `aborted`
without running the child. Regression test proven red on the pre-fix code.
Also refresh two stale RFC prose blocks the round-1 fix left behind: the
subagent RFC's Problem statement (cited the removed `TODO(sub-agents)` markers
and claimed nothing existed yet) and the unify-id RFC's fork/spawn risk bullet
(described the seam as "explicitly deferred" via `AgentLoop.create`'s old TODO),
now pointing at the realized seam.
Two merge-blocking bugs in the shared in-process run driver, both rooted in
`readResult` scanning the whole child session and deriving the stop reason only
from `turn/end`:
- A pre-turn `cancel()` cleared the queued prompt before any `turn/end` was
logged, so the run settled `error` instead of `aborted`, violating the
`SubagentRun.cancel()` contract. The driver now tracks that a cancel was
requested and maps the no-turn case to `aborted`.
- A fork child whose own turn produced no `assistant/message` returned the
SEEDED parent's last message as a `completed` success. `readResult` now scopes
to the child's OWN events (after the seed prefix), so a message-less child
yields empty output.
Both fixes carry a regression test proven to go red on the pre-fix driver.
Also: correct the `SubagentRun.id` / event-payload docs (it is the child AGENT
id, not a session id — the backend mints distinct tokens); refresh the stale
`coding-agent` welcome string (subagent is now a tool); and replace the stale
`TODO(sub-agents)` "deferred" prose in the Agent interface, core.md, and
architecture.md with an accurate pointer to the realized seam.
The second PR of the subagent seam: the two in-process backends that run a
child agent on the same cordis context, reusing the agent factory's quiescent
AgentHandle teardown. Both register on ctx.subagents (PR1's named-provider
registry) and share one run driver.
- dsh-subagent-spawn: a FRESH child via ctx.agents.create — own session, the
parent's model by default (overridable), zero inherited conversation. Also
exports the shared in-process run driver (startInProcessRun): mint ids, stamp
cwd/parentSession-lineage/depth, drive the one-shot (send → whenIdle), read
the last assistant/message + turn/end reason, dispose to quiescence.
- dsh-subagent-fork: a child SEEDED with the parent's balanced completed-turn
prefix (the log up to and including its last turn/end), so the child inherits
context. The in-flight unbalanced turn is excluded — a raw seed would fail the
invariants replay. Proven: a regression test goes red if the boundary seeds
the open turn.
- Seam extension: CreateAgentOptions.seed, threaded through AgentLoop.createAgent
→ ctx.sessions.prepare({ seed }) (the primitive resume already used). This is
the fork-lineage path the TODO(sub-agents) markers anticipated.
- Depth: a merge-extensible AgentOptions.subagentDepth (0 top-level, parent+1 for
a child); the depthLimit capability refuses a spawn past request.maxDepth.
Tests: real-loop unit tests for both backends (mock MODEL only, real loop +
invariants), a multi-subagent test (one parent drives a fork AND a spawn child
then keeps working), and a with-key e2e (a real parent delegates via the
`subagent` tool to a real child that writes a file on disk — world-verified).
100% per-file coverage. The coding-agent demo wires the spawn backend + tool.
Snapshot coverage of nested agents is deferred to a stacked follow-up
(TODO(subagent-snapshots)): dsh-llm-replay is a single global positional cursor
that cannot route calls to a parent vs. a child on one context. Recorded in the
RFC's deferrals and a new AGENTS.md rule: designing a subsystem must design its
test infrastructure END TO END up front, verifying the snapshot/e2e harness can
express the new shape — a gap this plan hit.
A single try/catch around ctx.emit prevented a thrown subagent/start or
subagent/end listener from propagating, but cordis emit dispatches listeners in
a `.map(cb => cb())` that HALTS on the first throw — so a bad subscriber still
starved the listeners registered after it, violating the AGENTS.md
callback-boundary rule ("one bad subscriber must not starve the listeners after
it"). Resolve the listener callbacks via ctx.events.dispatch and contain each
call individually, the same per-listener guarantee BashExecutor.notifyTaskDone
gives its own listener set.
The two containment tests now register TWO listeners where the first throws and
assert the second still observes the event (start) and the settle (end) — a
regression that fails on the per-emit code (verified: reverted, watched both go
red, restored).
Address four findings from the first Codex review round:
- Contain subagent/start|end listener throws (emitContainedStart/End): a
thrown lifecycle listener could escape SubagentService.start() before the
caller received the live run to dispose it (a leaked child), and a thrown
subagent/end listener could surface as an unhandled rejection on the detached
result-settle hook. Both emits now log-and-contain, mirroring the agent
registry's agent/created|disposed containment.
- Make the model-facing tool name configurable (Config.toolName, default
subagent). The docs say to load dsh-tool-subagent once per provider to expose
multiple transports, but the hardcoded name made the second load throw a
duplicate-tool-name error; a distinct toolName per load is now required and
documented.
- Reach the per-file 100% coverage gate: tests for the subagent/end error
branch, lifecycle-listener containment, every stopReasonError arm + the
merge-extensible default, the multi-provider toolName path, agentOptions
forwarding, and the direct-apply schema-bypass fallbacks.
- Document the seam vocabulary in docs/core-data-structures/subagent.md with
verbatim type-equiv blocks + manifest entries, and link it from core.md (a
brand-new core/seam type the doc-sync gate cannot detect on its own).
Introduce the `packages/subagent/` group and the abstract subagent seam — an
agent delegating to a child agent — as a named-provider registry (`ctx.subagents`),
unlike the single-implementation bash seam, so multiple transports (in-process,
ACP, future A2A) coexist. This first PR lands the interface, a scripted test
backend, and the model-facing tool, validated through the real cordis load path.
- dsh-subagent: SubagentService registry + SubagentProvider/SubagentRun
vocabulary + subagent/start|end events. Start-time capabilities (outputSchema,
depthLimit, toolFilter) are checked pre-start and rejected loud; runtime
capabilities (sendMessage, resume) are optional methods on SubagentRun.
- dsh-subagent-mock (support): scripted provider for keyless, deterministic
tests through the real Loader/export path.
- dsh-tool-subagent: the model-facing `subagent` tool, config-bound to one
provider; synchronous collect with try/finally dispose, signal->cancel
bridging, and non-completed-stop-reason -> isError mapping.
- Proposed RFC documenting the seam, the fork-vs-spawn-as-separate-backends
decision, own-session isolation, synchronous-collect scope, and the deferral
of background/poll/spill to a future unification with bash.
- Wire the new group into tsconfigs, build refs, package hierarchy docs, the
module graph, and the cordis catalog.
RFC: docs/rfc/proposed/feature/2026-06-21-subagent-capability-seam.md
Tianyi Cui