Files
deepseek-harness/packages/subagent/subagent-inprocess
Tianyi Cui 74502fa8c2 Structured output on the subagent seam: schema subset, capture runtime, spawn/fork support
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.
2026-07-06 23:29:08 +08:00
..

@deepseek-ai/dsh-subagent-inprocess

The shared in-process subagent run driver. A pure library (no provider, no registration) that the in-process backends — spawn (a fresh child) and fork (a child seeded with a prefix of the parent's log) — both build on. The backends are thin shells that differ ONLY in the session seed they pass; everything downstream lives here, so neither backend depends on the other.

What it exports

startInProcessRun(ctx, request, options): SubagentRun

Runs a child as a child Agent on the same cordis context (ctx.agents):

  1. computes child depth = depthOf(parent) + 1; if request.maxDepth is set and exceeded, throws SubagentDepthError (the depthLimit capability); a request.outputSchema is asserted against the supported subset (assertSupportedOutputSchema from dsh-tools) before any child exists;
  2. creates a child via ctx.agents.create with a fresh AgentId/SessionId, the parent's cwd + parentSession lineage, the optional options.seed (fork's completed-turn prefix; omitted for a fresh child), and agentOptions (the child inherits the parent's model by default — a child with no model can't run — overridable via request.agentOptions.model; the deployment persona needs no inheritance — it is a context-wide prompt section);
  3. drives the one-shot: child.send(prompt) then await child.whenIdle() (ordering matters — send enqueues synchronously, so whenIdle observes the queued work and resolves on the child's running → idle transition, never before the turn starts); a structured child that finished a turn CLEANLY without calling structured_output is re-prompted (a nudge — a fresh turn) up to options.structuredNudgeRetries times;
  4. reads the result, scoped to the child's OWN events (everything at or after seedLength, so a seeded child that produced no message of its own never returns the seeded parent's last message): the last assistant/message content (deep-cloned — the log is frozen) and the last turn/end.reason mapped to a SubagentStopReason. A structured run surfaces the captured value as result.structured; a structured child that finished cleanly WITHOUT ever capturing settles error (a clean finish without the demanded result is a failure, not a success with a missing field).

dispose() delegates to AgentHandle.dispose() (stop loop → await quiescence → remove session); cancel() cancels the child's in-flight turn. A cancel landing before any turn/end (the pre-turn window) still settles aborted, honoring the cancel contract rather than the generic no-turn error.

InProcessRunOptions

{ providerName: string; seed?: SessionEvent[]; structuredNudgeRetries: number } — the per-backend inputs: the provider name (for error context), the optional child-session seed, and the structured-run nudge budget (REQUIRED, resolved from the backend's validated Config — the driver never fills it with a hidden default).

Structured output: acquireStructuredRuntime(ctx): StructuredAcquisition

The mechanism behind outputSchema for in-process children. One globally registered structured_output capture tool (its registered parameters are a placeholder) plus two listeners, registered once per root context and shared by every holder:

  • an agent/request waterfall listener registered prepend: true that post-processes await next()final-request enforcement: the request that hits the wire never carries structured_output for an agent without a structured run, and for one that has it always carries the run's OWN schema (as the tool's parameters) plus the calling instruction appended to its system text (the demand travels with the tool — AgentOptions has no per-agent prompt field to carry it). Per-agent shaping lives here because the tool registry and prompt assembly are context-global while schemas differ per concurrent child; cooperative mutate-then-next() would not survive a downstream listener returning a replacement request.
  • an agent/turn-continuation listener (also prepend: true — an earlier-registered force-continue listener returning without next() must not decide the turn before the veto runs) that stops a child's turn once its output is captured, so a successful capture doesn't buy a wasted extra model step.

The capture tool validates each call against the run's schema (validateStructuredValue) — violations become an INVALID_ARGS isError result the model retries in-turn; a valid call records the value.

Lifetime is refcounted with two kinds of holder: each backend acquires for its plugin lifetime (apply), and each structured RUN holds its own acquisition from start to settle — so unregistration can never precede a live run's settle, and the runtime disposes only when the last backend AND the last run are gone. release() is idempotent per acquisition.

depthOf(agent): number

Delegation depth rides on a merge-extensible AgentOptions.subagentDepth field (0 for a top-level agent, parent + 1 for a child), so a nested spawn reads its parent's depth from parent.options.subagentDepth. depthOf reads it (absent ⇒ 0).

SubagentDepthError

Thrown by startInProcessRun when a spawn would exceed the request's maxDepth cap; carries attemptedDepth and maxDepth.