A subagent child's recursion depth lived only in runtime AgentOptions,
so a persisted child came back from resume counted as top-level and
maxDepth stopped binding after every restart. Add
SessionHeader.delegationDepth, round-trip it through the JSONL and
SQLite backends (SQLite schema v5), restore it on agent-loop resume,
and write it when the in-process backends create a child. The seam now
owns the shared depth vocabulary (delegationDepthOf): the persisted
header is authoritative and monotone — runtime options may deepen it
but never lower it.
Rewrite the agent-scope RFC with executable examples and an explicit security non-goal. Harden subagent scalar and depth validation, and pin live tool-filter semantics across code, tests, and generated docs.
Drain idle injection flushes before agent teardown, snapshot approval and subagent provider inputs, and gate subagent lifecycle events on real child readiness. Align the RFCs and generated contracts with the hardened behavior.
Reviewer-agent C1 on the audit delta: "byte-identical" overstated the
untampered-path guarantee — a 190-order section registered AFTER the
structured runtime sorts before the instruction in the registry's stable
sort but after it in the re-assert's band insertion. Intra-band section
order carries no contract, so the behavior is right and unchanged; the
module doc and both in-code comments now say exactly that instead of
claiming byte identity.
Supersedes the single-slot staging the execution-identity fix (06c5f17e)
kept: the one pending slot needed a mismatch-drop branch plus a defensive
coverage-ignored finally to manage orphans, and it carried a latent trap —
under the loop's documented parallel-execution TODO, two in-flight capture
trips would overwrite the slot and BOTH be dropped.
Staging in a WeakMap<ToolExecution, {value}> makes the stale-stage class
structurally impossible instead of managed: an entry orphaned by an outer
short-circuiting listener can never match a different execution's lookup
(whatever call id that execution carries), needs no drop bookkeeping (the
map reclaims it with the execution object), and staging cannot cross-clobber
under parallel execution. Staging is the only layer this future-proofs — a
parallel cut would still owe its own single-accept rule for the captured
value, which is documented rather than claimed. Behavior is pinned by the
existing orphan/call-id-reuse regression tests, which pass unchanged; the
commit listener loses two branches and the v8-ignore.
Re-auditing the review-fix commits surfaced a regression the REPLACE
re-assert (825cbab3) introduced: unconditionally rebuilding both arrays as
filter(...)+append moved structured_output to the END of the model-visible
tool list on every untampered assembly (overriding the registry's
toolOrder/lexicographic contract) and moved the instruction section to the
absolute array end — renderPrompt reads ARRAY order, so any section above
order 190 would render before the trailing instruction, violating the
sections-sorted-ascending contract. The presence-check version it replaced
touched neither array when the entries were intact.
The re-assert keeps its REPLACE content semantics but is now
placement-preserving: the tool is replaced IN PLACE (duplicates collapse,
append only when stripped); the section is re-inserted at its
ascending-order position (the first entry above 190 — exactly where the
registry's stable sort put it, so the untampered path reaches the model
byte-identical). Pinned by two regression tests that fail against the
filter+append form: untampered placement (tool before a lexicographically
later tool, instruction before an order-200 section) and tamper recovery
(stripped section re-enters its band; an added duplicate collapses to one
right-schema entry).
ds-review-bot round-2 findings: (1) dsh-subagent's runtime import of
@deepseek-ai/dsh-scope was undeclared in its manifest and tsconfig
references (the root paths map masked it; the emitted package would import
an undeclared dependency) — wired as peer+dev with the project reference,
module graph regenerated. (2) The structured re-assert only ensured
PRESENCE, so a downstream listener injecting a same-named entry with the
wrong schema kept it model-visible while validateStructuredValue enforced
the real one; it now REPLACES any same-named tool/section with the run's
own. Pinned by a wrong-schema-injection test asserting exactly one entry
carrying the run's schema.
Codex confirmation-round finding: an OUTERMOST prepend pre-execute deny
skips the runtime's own pre-execute clear, and the denied call still
reaches post-execute — so a reused adapter-minted call id could promote an
orphaned stage on the default accept path. The stage is now keyed by the
ToolExecution OBJECT identity, the one token that provably ties a stage to
one pipeline trip: only the execution whose own body staged can commit,
whatever any call id says. The pre-execute clear is gone (one mechanism);
the commit's mismatch drop is now the reachable primary guard. Repro test:
orphaned stage + outer pre-execute deny with the same call id never
promotes; a fresh valid call still captures.
Adversarial-review findings (own reviewer agent), each verified and pinned:
B1: agents.register() returned a wrapper lambda, so the factory composite's
yield could not identity-nest it — on OWNER unload the unregistration (and
agent/disposed) disposed as a concurrent sibling, firing mid-drain while
the final turn was still closing (pre-existing on master; this branch's
docs re-assert the order, so it must be true). register() now returns the
EXACT cordis effect disposer (the Scope.rawDispose move); the composite
nests it and owner unload runs stop/drain -> unregister -> detach -> scope
like every other path. Regression test pins turn-end before disposed
before detach on owner unload.
B2: the structured two-phase commit could promote a stale stage when a
later capture call REUSED the orphaned stage's call id with a body that
never staged (denied downstream, or invalid args throwing pre-stage). The
runtime's pre-execute listener now clears any stale stage unconditionally
when a new capture call enters the pipeline — only a call's own body can
stage for its commit; the call-id mismatch guard becomes a defensive
second layer. Repro test: blocked capture then same-id invalid call.
C1: an explicit empty toolFilter config now fails at plugin LOAD (the
check is self-contained) instead of killing every delegation at child
setup. C2: Scope.dispose/ScopeHost.dispose @returns state the single-shot
repeat-call semantics honestly.
Every subject-extractor row of the invariants carrier table is exercised
with a matching and a foreign-keyed carrier; the HMR re-apply seed path
(sessions of agents that predate the plugin are marked started) is pinned;
the scoped tool-provider disposal, plural restrict() validation, singular
scopeHost absentee, tool-subagent passthrough, stale-stage drop, and
disposing-parent spawn (INACTIVE_EFFECT, no orphan) each gain their test.
Two genuinely defensive branches carry justified v8-ignore markers.
SubagentStartRequest gains persona (capability-gated like toolFilter); the
in-process driver composes the child's scoped world in the factory's setup
window — persona as a scoped shadowing deployment:persona section,
toolFilter as a scoped tools.restrict() (loud unknown-name validation),
outputSchema as the scoped structured runtime. spawn/fork now advertise
every start-time capability; ACP stays all-false. A parent-scope teardown
effect links each child to its parent through the memoized handle, so a
disposed parent reaches its whole subtree even if the delegating tool's
finally never runs; subagent/start|end dispatch in the delegating parent's
scope.
structured.ts loses the placeholder schema, the final-assembly swap/strip,
the refcounted root runtime, and the WeakMap state: each child registers
its OWN capture tool (real schema), instruction section, and enforcement
listeners on child.ctx, riding the child's fiber. The commit listener is
call-keyed (a stale stage from a short-circuited post-execute chain is
dropped, never promoted on a later call), and one scoped prepend re-assert
listener preserves the final-assembly guarantee against a stripping global
listener.
tool-subagent gains persona/toolFilter/maxDepth passthrough config —
deny-listing the delegation tool (or maxDepth) is how a deployment bounds
recursion; the omitted-toolFilter schema key is forced absent (a
materialized {} would mean an empty allow-list, i.e. deny-everything).
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
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).
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.
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.
Yichen Jiang
Tianyi Cui