deepseek-harness/docs/core-data-structures/subagent.md

Subagent

English | 中文

The subagent seam — an agent delegating work to a child agent. Like bash it is one optional capability, not part of the agent-loop spine, so its vocabulary lives here rather than in core.md. But it differs from every other seam on one axis: multiple provider implementations coexist in one context, registered by name (ctx.subagents), where bash allows only one executor. The registry shape mirrors the LLM adapter registry, not the single-service bash executor.

Interface: dsh-subagent (ctx.subagents + the vocabulary below). Implementations are sibling packages (dsh-subagent-spawn, -fork, -acp); the model-facing consumer is dsh-tool-subagent. The proposal and rationale: the subagent Agent Note.

Source: packages/subagent/subagent/src/types.ts

Two kinds of capability, discovered two ways

A provider advertises its start-time features on a static descriptor the service checks BEFORE a run exists; a request that needs one the provider lacks is rejected loud (SubagentError('UNSUPPORTED_CAPABILITY')), never accepted-then-ignored. Runtime features (steering, resume) are instead optional methods on SubagentRun — the method's presence IS the capability, and TS narrowing is the discovery mechanism.

/**
 * Which START-TIME features a provider supports. Checked by the service before delegating to
 * {@link SubagentProvider.start}: a request that needs a capability the chosen provider lacks
 * is rejected with a typed error rather than accepted-then-ignored (the "fail loud, no silent
 * degradation" rule). These static flags cover features needed before a run exists; runtime
 * capabilities such as steering and resume are optional {@link SubagentRun} methods whose presence
 * is the capability.
 */
interface SubagentCapabilities {
  /** Honor {@link SubagentStartRequest.outputSchema} (structured final output). */
  readonly outputSchema: boolean
  /** Enforce {@link SubagentStartRequest.maxDepth} (recursion cap). */
  readonly depthLimit: boolean
  /** Enforce {@link SubagentStartRequest.toolFilter} (child tool scoping). */
  readonly toolFilter: boolean
  /** Honor {@link SubagentStartRequest.persona} (a per-child persona). */
  readonly persona: boolean
}

The start request

The tool layer builds this request from the model input and its own config; the service validates it against the named provider before start. Required parent supplies the session cwd, lineage, and delegation depth. Optional output schema, depth, tool filter, and persona require matching capability flags. Unsupported schemas fail at start; in-process backends scope filters and personas to child creation and implement the supported object-rooted schema with a forced capture tool.

/**
 * What a caller asks for when starting a subagent. The tool layer builds this
 * from the model's `{ description, prompt }` plus its own config; the service
 * validates {@link SubagentCapabilities} against the named provider, then
 * passes it to {@link SubagentProvider.start}.
 */
interface SubagentStartRequest {
  /** The task/prompt for the child agent (a user message in the child session). */
  readonly prompt: ContentBlock[]
  /**
   * The spawning ("parent") agent — the one whose tool call started this
   * subagent. REQUIRED: in-process backends read `parent.session.header` for
   * the working directory, the `parentSession` lineage to stamp on the child,
   * and the parent's delegation depth. The out-of-process backend (ACP) reads
   * exactly one field — the session header's cwd, the child's workspace when
   * no deployment `cwd` override is configured; nothing else crosses the
   * process boundary.
   */
  readonly parent: Agent
  /**
   * Cancellation signal from the spawning context (the tool's `exec.signal`).
   * This is the canonical cancellation channel both before and after startup:
   * a provider rejects `start()` after cleaning partial resources when it
   * fires before publication, and cancels a published child when it fires
   * afterward.
   */
  readonly signal: AbortSignal
  /** Per-child agent options (model and plugin-defined extension fields). */
  readonly agentOptions?: AgentOptions
  /**
   * Object-rooted JSON Schema within `assertObjectJsonSchema`'s enforced subset. Start rejects
   * unsupported schemas or providers without the capability. Data must be plain host-realm JSON;
   * a successful child returns the matching value as {@link SubagentResult.structured}.
   */
  readonly outputSchema?: ObjectJsonSchema
  /**
   * Optional absolute delegation-depth cap for the child being started: its
   * computed depth must be less than or equal to this non-negative safe
   * integer. Requires {@link SubagentCapabilities.depthLimit}; rejected at
   * start otherwise.
   */
  readonly maxDepth?: number
  /**
   * Optional child tool scoping. Requires {@link SubagentCapabilities.toolFilter};
   * rejected at start otherwise. In-process backends apply it as a scoped
   * `tools.restrict()` in the child's creation window: the named tools vanish
   * from the child's prompt AND refuse to execute (one visibility), with loud
   * unknown-name validation.
   */
  readonly toolFilter?: ToolRestriction
  /**
   * Optional per-child persona. Requires {@link SubagentCapabilities.persona};
   * rejected at start otherwise. In-process backends register it as a scoped
   * `deployment:persona` section on the child, SHADOWING the deployment's
   * persona for this child alone — same template semantics as the deployment
   * persona (strict `{{…}}` interpolation against the registered variables).
   */
  readonly persona?: string
}

signal is the single cancellation channel before and after readiness. The subagent composition-controls Agent Note owns the persona, live global-tool filter, absolute-depth, and visibility-not-authority rationale.

The terminal result: SubagentResult

The outcome of a run, resolved by SubagentRun.result. structured is present only after a requested outputSchema was successfully satisfied; requesting a schema does not guarantee it, and a provider may return stopReason: 'error' when the child fails or finishes without a valid capture. A non-completed stopReason means output may be partial — the consumer maps it to an isError tool result rather than reporting partial output as success.

/**
 * The terminal outcome of a subagent run, resolved by {@link SubagentRun.result}.
 */
interface SubagentResult {
  /** The child's final assistant output (the last assistant message's content). */
  readonly output: ContentBlock[]
  /**
   * The structured result after a requested `outputSchema` was successfully
   * satisfied. Requesting a schema does not guarantee presence: a provider can
   * end with `stopReason: 'error'` when the child fails or finishes without a
   * valid capture. Shape is validated against the request schema by the
   * provider; `unknown` here because the seam is schema-agnostic.
   */
  readonly structured?: unknown
  /** Why the run ended. A non-`completed` reason means `output` may be partial. */
  readonly stopReason: SubagentStopReason
}

SubagentStopReason is a merge-extensible derived union — a backend may add variants, so consumers branch on the known cases and treat an unknown terminal reason as a failure:

/**
 * Why a subagent run ended. Merge-extensible (a backend may add variants);
 * consumers branch on the known cases and fall through `default`. The known
 * cases mirror the harness turn-end vocabulary so the tool layer can map a
 * non-`completed` result to an `isError` tool result.
 */
interface SubagentStopReasonMap {
  /** The child finished its turn normally. */
  completed: 'completed'
  /** The run was cancelled by its request signal or by disposal. */
  aborted: 'aborted'
  /** The child failed (model error, transport error). */
  error: 'error'
  /** The child hit its token ceiling before finishing. */
  'max-tokens': 'max-tokens'
  /** The child declined the task. */
  refusal: 'refusal'
}

A live run: SubagentRun

SubagentRun is the consumer-owned handle for a ready child. Consumers await result and always dispose the run to reach quiescence. Child failures resolve with a non-completed stop reason; only unrepresentable infrastructure faults reject. Optional sendMessage and resume methods advertise their runtime capabilities by presence.

/**
 * Child handle returned only after readiness. Consumers await {@link result} and must always
 * {@link dispose} to cancel remaining work and reach quiescence. Optional methods are runtime
 * capability discovery; narrow their presence before calling.
 */
interface SubagentRun {
  /**
   * Parent-scoped run id. For a local run, this MUST equal the published child
   * session id, whose `parentSession` records `request.parent.session.id`; a
   * remote provider mints an id unique in the parent namespace.
   */
  readonly id: SessionId
  /**
   * The exact published in-process child, or `undefined` for a remote run.
   * When present, its id is {@link id}; the provider retains no ownership
   * implication beyond the run's ordinary {@link dispose} contract.
   */
  readonly localAgent: Agent | undefined
  /**
   * Resolves with the child's terminal {@link SubagentResult} when the run
   * settles. Does NOT reject on a child-level failure — a model/transport
   * failure resolves with `stopReason: 'error'` so the consumer maps it to an
   * `isError` tool result. Rejects only on an infrastructure fault the seam
   * cannot represent as a stop reason.
   */
  readonly result: Promise<SubagentResult>
  /**
   * Cancel remaining work, reach child quiescence, and release the run's
   * resources (in-process: dispose the owned agent and remove its session;
   * ACP: kill and reap the subprocess). Idempotent.
   */
  dispose(): Promise<void>
  /**
   * OPTIONAL (steering capability): send additional content to the running
   * child between steps. Present only on providers that support live steering.
   */
  sendMessage?(content: ContentBlock[]): void
  /**
   * OPTIONAL (resume capability): send a follow-up task to a settled child,
   * continuing its session, and return a fresh run for the continuation.
   */
  resume?(content: ContentBlock[]): Promise<SubagentRun>
}

A local run MUST publish an ordinary child agent/session before start() fulfills, return that child session id as SubagentRun.id, expose the exact child as localAgent, and record request.parent.session.id in the child's parentSession header. Runtime ownership may place the child under the parent, provider, or root scope. A remote provider instead returns a parent-scoped lifecycle id and localAgent: undefined.

The provider seam: SubagentProvider

Each provider is a named child-agent transport, and multiple providers may coexist. The service validates requested start-time capabilities before start(). inheritsParentContext describes only conversation seeding (fork: true; spawn and acp: false), allowing consumers to generate accurate model-facing wording without implying inherited tools, services, or authority.

/**
 * A subagent backend: one transport for running a child agent (in-process
 * spawn/fork, ACP to another process, …). Implementations register under a
 * unique name via {@link SubagentService.registerProvider}; multiple providers
 * coexist in one context (unlike the single-implementation bash seam). The
 * Providers are trusted same-process implementations; callers treat their
 * descriptors and returned values as borrowed immutable data.
 */
interface SubagentProvider {
  /** Unique registry name (e.g. `spawn`, `fork`, `acp`). */
  readonly name: string
  /** The start-time features this provider supports (see {@link SubagentCapabilities}). */
  readonly capabilities: SubagentCapabilities
  /**
   * Whether the child sees the parent's completed-turn prefix. This is descriptive, not a
   * service-validated start capability: the model-facing tool derives truthful wording from it.
   * It says nothing about tool registration, injected services, or authority inheritance.
   */
  readonly inheritsParentContext: boolean
  /**
   * Establish a child and return its handle only after publication. The
   * service has already validated that every requested start-time capability
   * is supported, so an implementation may assume e.g. `request.maxDepth` is
   * honorable when present. If setup fails or `request.signal` aborts before
   * fulfillment, the provider owns and cleans all partial resources before this
   * promise rejects. Ownership transfers to the caller only on fulfillment.
   */
  start(request: SubagentStartRequest): Promise<SubagentRun>
}

start() fulfills only with a ready run. The service mints a unique runId, snapshots local from the provider's exact localAgent, observes the result, emits subagent/start, and returns the same run; rejection implies provider cleanup and emits no lifecycle pair. The paired subagent/end carries the same identity and the final output or infrastructure failure. Both events are observe-only and contain listener exceptions.

In-process backends: depth and seed

The spawn and fork backends create an ordinary agent through parent.ctx, pass cancellation into core creation, and dispose through AgentHandle. Provider removal blocks new starts without revoking accepted runs. Each child gets a new flat scope rather than inheriting parent registrations. Depth and fork seeding reuse existing agent and session vocabulary:

• Delegation depth is durable SessionHeader.delegationDepth plus the merge-extensible runtime field AgentOptions.subagentDepth; absence means top-level depth zero, and the greater present value is authoritative. The seam owns both fields — the loop neither sets nor reads them — so an in-process child persists parent depth + 1, resume cannot lower it, and every start rejects a derived depth outside the safe-integer domain or above a defined absolute request.maxDepth cap.
• Fork seeding uses CreateAgentOptions.seed (a SessionEvent[] prefix threaded through AgentLoop.createAgent → ctx.sessions.prepare({ seed }), the same primitive resume uses). The fork backend passes a balanced completed-turn prefix of the parent's log — the parent's events up to and including its last turn/end — so the seed is contiguous-from-0 and the invariants replay accepts it (the in-flight, unbalanced turn is excluded).