The subagent seam — an agent delegating work to a child agent. Like [bash](bash.md) it is **one optional capability**, not part of the agent-loop spine, so its vocabulary lives here rather than in [core.md](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](llm-streaming.md), not the single-service bash executor.
Interface: [dsh-subagent](../../packages/subagent/subagent) (`ctx.subagents` + the vocabulary below). Implementations are sibling packages (`dsh-subagent-spawn`, `-fork`, `-acp`); the model-facing consumers are [dsh-tool-subagent](../../packages/subagent/tool-subagent) (per-provider delegation) and [dsh-tool-subagent-control](../../packages/subagent/tool-subagent-control) (the optional global `send_message`). The same `ctx.subagents` service owns continuable-child orchestration through an internal activation manager. The rationale lives in [the subagent Agent Note](../../.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.md), [the continuable subagents Agent Note](../../.agents/notes/implemented/feature/2026-07-28-continuable-subagent-conversations.md), and [the merged-service Agent Note](../../.agents/notes/implemented/simplification/2026-07-26-merge-subagent-control-service.md).
Sources: [`packages/subagent/subagent/src/types.ts`](../../packages/subagent/subagent/src/types.ts), [`packages/subagent/subagent/src/index.ts`](../../packages/subagent/subagent/src/index.ts), and [`packages/subagent/subagent/src/continuation.ts`](../../packages/subagent/subagent/src/continuation.ts)
A provider advertises its **start-time** features on a static descriptor the service checks BEFORE a one-shot run exists; a request that needs one the provider lacks is rejected loud (`SubagentError('UNSUPPORTED_CAPABILITY')`), never accepted-then-ignored. Those flags describe only the one-shot [`start()`](#the-provider-seam-subagentprovider) path, where the provider composes the child. **Continuable** children are composed by the continuation manager itself, so they are gated by one optional method whose presence IS the capability, with TS narrowing as the discovery mechanism: [`SubagentProvider.prepareContinuable`](#the-provider-seam-subagentprovider).
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.
`signal` is the single cancellation channel before and after readiness. The [subagent composition-controls Agent Note](../../.agents/notes/implemented/feature/2026-07-12-subagent-persona-tool-filter-and-depth.md) owns the persona, live global-tool filter, absolute-depth, and visibility-not-authority rationale.
Providers receive exactly this request: one-shot delegation has no service-resolved continuation state, because a continuable child never reaches `SubagentProvider.start()`.
A **continuable background subagent** is one durable child Session with at most one process-local **Activation** — a residency epoch for a reconstructed child Agent. An Activation is not a request, result, cancellation, or Task boundary: it may execute many FIFO turns and stays resident while descendants it created are still running. The continuation manager owns activation admission, direct-parent authorization, the live ownership graph, cold resume, and child-first disposal; the Agent loop owns all turn ordering and execution. No continuable path creates a Task or an intermediate result-bearing wrapper.
`SubagentService.startContinuable()` reserves the stable child id, snapshots the versioned `subagent/descriptor` payload, asks the named provider for its detached `ContinuableCreateSpec`, creates the child Agent through a private activation-owner scope, establishes any continuable-parent ownership, and submits the initial prompt. It resolves with `{ childId, messageId }` when inbox acceptance yields the message id — without waiting for the turn to start or for the message to enter the Session log. Every failure before that acceptance rejects with neither id, disposing any created handle and rolling back the Activation and parent ownership.
`SubagentService.followup()` is the sole continuation-message operation, and routing depends only on Activation residency:
`running` means the Agent has an active admission or turn, or waking inbox work; `waiting` means it is quiescent but still owns at least one child Activation that has not completed disposal; `settled` means quiescent with every owned child disposed, at which point the manager disposes the `AgentHandle` and removes the Activation. The manager derives these internal conditions from Agent quiescence and the owned-child set rather than maintaining a second execution state machine.
The Agent inbox is the only queue. Every continuation message becomes one `Agent.followup()` FIFO turn, so accepted messages have one observable order and a follow-up cannot redirect a turn already underway. Successful delivery returns the accepted `MessageId`; the existing `agent/inbox/enqueue`, `agent/inbox/dequeue`, and `agent/inbox/discard` events remain the message-lifecycle observations, and the continuation layer defines no subagent-specific delivery route.
Follow-up authority comes from an exact live Agent tool context. The authenticated Agent must be the durable child's direct parent recorded in `SessionHeader.parentSession`. `MessageSource` and `senderSessionId` are durable provenance after admission and grant no authority; the optional model-facing tool uses `CoordinatorMessageSource`.
For both operations the caller signal owns lookup, materialization, and admission only until inbox acceptance. Afterwards the manager owns the Activation independently: later caller cancellation neither cancels the accepted turn nor disposes the child, and the seam exposes no public subagent cancellation or steering operation.
Every Activation owns its `AgentHandle` and an `ownedChildren: Set<SessionId>`; because one Session has at most one live Activation, the child Session id identifies the live child without another runtime-incarnation reference. Starting a child or submitting parent-originated work registers the child in a continuation-managed parent's set before the child can run, and that parent cannot settle while the set is non-empty. A top-level or other non-continuation Agent has no Activation and stays outside the waiting graph. Child release happens only after the child Agent is quiescent, every child of that child is disposed, the final durability checkpoint settles, and the child's `AgentHandle` completes disposal.
Only `ctx.sessions.flush(session) === true` confirms durability; `false` or rejection reports `DURABILITY_FAILED`. Either way the manager still disposes the handle and releases ownership, because retaining a failed child would permanently pin its ancestors in `waiting` — the persisted child state may then be missing or stale on a later resume. Manager unload invokes an internal manager-wide drain that closes admission and disposes every live forest; `drainContinuableDescendants(parents)` closes admission only below exact live host-owned Agents and disposes their continuable descendants while unrelated forests remain live. Both await already-admitted materializations in their scope, propagate cancellation top-down, release handles child-first, and await every selected branch despite individual failures. Durable child Sessions survive that process-local teardown.
/** Identities returned once a continuable child accepted its initial prompt. */
interface ContinuableStart {
/** The durable child session id, stable across activations. */
readonly childId: SessionId
/** The accepted initial prompt's inbox message id. */
readonly messageId: MessageId
}
```
The provider participates only in preparing the initial creation spec, where `spawn` and `fork` differ. Its returned spec carries only detached provider-specific creation inputs — today the optional parent-history seed — and no Agent, `AgentHandle`, prompt delivery, result, disposal, or resume operation. Cold resume does not dispatch through a provider at all: the manager folds the generic descriptor, calls `ctx.agents.resume()` through the same activation-owner scope, and submits the waiting turn.
The descriptor (`SubagentDescriptorData` in [descriptor.ts](../../packages/subagent/subagent/src/descriptor.ts)) snapshots explicit fields — provider name, resolved child `agentOptions.provider`/`model`, optional `persona`/`toolFilter` — never the merge-extensible `AgentOptions` object, so an unrelated extension value cannot break continuation and a later composition input is a deliberate version change. It omits `subagentDepth` (cold resume trusts the persisted header's `delegationDepth` as the monotone floor) and `outputSchema` (a one-shot result contract, not durable composition). The continuation manager appends the model-hidden `subagent/descriptor` event after any provider-supplied lineage and before the initial prompt is admitted; `header.seedLength` remains the fork-lineage boundary, so descriptor lookup reads the child's own suffix. The event is log-only: no `surfaceOp`, never in model history, and retained across compaction by the append-only log.
The outcome of a one-shot 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.
`SubagentStopReason` is a [merge-extensible derived union](core.md#the-map--derived-union-pattern) — a backend may add variants, so consumers branch on the known cases and treat an unknown terminal reason as a failure:
`SubagentRun` is the consumer-owned handle for a ready one-shot child — one disposable foreground delegation with one result, never a durable child handle. 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. A run has no steering and no resume: continuable conversations have no run at all, because the continuation manager holds their `AgentHandle` directly and orders every turn through the child's own inbox.
A local one-shot 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`.
Each provider is a named child-agent transport, and multiple providers may coexist. The service validates requested start-time capabilities before `start()`, and rejects a continuable start on a provider without `prepareContinuable`. `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.
Provider `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. Each continuable Activation emits the same observe-only pair for its residency epoch, so a cold resume is a new epoch with its own `runId`. The paired `subagent/end` carries the same identity and the final output or infrastructure failure. Both events are observe-only and contain listener exceptions. Their `provider` field is provenance for the run or Activation epoch, not a claim that the provider remains registered when the edge is emitted.
The spawn and fork backends create an ordinary one-shot agent through `parent.ctx`, pass cancellation into core creation, and dispose through `AgentHandle`; a continuable child is instead created by the continuation manager through its own activation-owner scope. 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, cold 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 `ctx.agents.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](../../packages/support/invariants) replay accepts it (the in-flight, unbalanced turn is excluded).
