Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(subagent): continuable background subagents

Browse Source 
Implement the continuable background subagents RFC: a durable child
session with a series of Task-backed activations, each disposing its
run before the Task settles.

- dsh-subagent: rename SubagentRun.sendMessage to strict steer, drop
  run-level resume, add SubagentProvider.resume dispatch via
  SubagentService.resume, the continuation start field, and the
  versioned model-hidden subagent/descriptor session event.
- dsh-subagent-inprocess/-spawn/-fork: publish the control-allocated
  child id, append the descriptor inside the initial turn, implement
  cold resume from the child's own transcript under the live parent
  scope, and strict running-only steer.
- dsh-subagent-control (new): SubagentControlService owning stable
  child ids, descriptor snapshot/fold/authorization, Task-backed
  activation with settle-then-dispose ordering, the process-local
  active-run association, and steer-or-resume sendMessage routing.
- dsh-tool-subagent: background route branches on the provider's
  resume capability (continuable via the control service; one-shot
  task for ACP), returning both child and task ids.
- dsh-tool-subagent-control (new): the globally named send_message
  tool rendering steered/started routes.

Keyless coverage spans Task ownership and disposal ordering, running
delivery, cold follow-up, descriptor rejection and rollback, known-id
reconstruction, kill during lookup, admission races, and a new
subagent-continuable ACP snapshot scenario.
This commit is contained in:
Dudu-0223
2026-07-23 17:07:38 +08:00
committed by imccyu
parent 76ff841279
commit 99a778d63f
83 changed files with 3167 additions and 627 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
packages/subagent/subagent/README.md
View File

@@ -10,17 +10,19 @@ The family separates the stable interface from implementations and model-facing
| Package | Role |
|---|---|
| `@deepseek-ai/dsh-subagent` | Provider registry, request/result types, and lifecycle events. |
| `@deepseek-ai/dsh-subagent-spawn` | Fresh in-process child. |
| `@deepseek-ai/dsh-subagent-fork` | In-process child seeded with completed parent turns. |
| `@deepseek-ai/dsh-subagent-acp` | Fresh out-of-process ACP child. |
| `@deepseek-ai/dsh-tool-subagent` | Model-facing tool over one configured provider. |
| `@deepseek-ai/dsh-subagent` | Provider registry, request/result/descriptor types, and lifecycle events. |
| `@deepseek-ai/dsh-subagent-spawn` | Fresh in-process child, with cold resume. |
| `@deepseek-ai/dsh-subagent-fork` | In-process child seeded with completed parent turns, with cold resume. |
| `@deepseek-ai/dsh-subagent-acp` | Fresh out-of-process ACP child (one-shot). |
| `@deepseek-ai/dsh-subagent-control` | Continuable-child orchestration: durable ids, descriptors, Task-backed activation. |
| `@deepseek-ai/dsh-tool-subagent` | Model-facing delegation tool over one configured provider. |
| `@deepseek-ai/dsh-tool-subagent-control` | The globally named `send_message` follow-up tool. |
Multiple providers may coexist under different names. This lets a deployment expose, for example, a cheap in-process child and an isolated ACP child without changing the service contract.
## Service API
`SubagentService` has four main operations:
`SubagentService` has five main operations:
| Member | Meaning |
|---|---|
@@ -28,8 +30,9 @@ Multiple providers may coexist under different names. This lets a deployment exp
| `getProvider(name)` | Return the provider, or `undefined` when absent. |
| `list()` | Return provider names in insertion order. |
| `start(name, request)` | Validate requested capabilities and semantic values, then await the provider until a real child is ready. Fulfillment returns a holder-owned `SubagentRun`; rejection means the provider has already cleaned every partial startup resource. |
| `resume(name, request)` | Capability-checked dispatch to `provider.resume?()` with the same run lifecycle observation as `start`. The caller (the control service) has already loaded the child, folded its descriptor, and authorized the parent; this seam stays collection-, Task-, and persistence-agnostic. |
`SubagentStartRequest.signal` is required and is the canonical cancellation channel. An abort before publication makes `start()` reject after rollback; an abort after publication cancels the live child. The request may also select a model, require structured output, cap delegation depth, restrict child tools, or set a child persona.
`SubagentStartRequest.signal` is required and is the canonical cancellation channel. An abort before publication makes `start()` reject after rollback; an abort after publication cancels the live child. The request may also select a model, require structured output, cap delegation depth, restrict child tools, set a child persona, or carry a resolved `continuation` (the control-allocated stable child id plus its durable descriptor), which requires the provider's `resume` capability.
Same-process requests, descriptors, results, and event payloads are trusted typed values borrowed as immutable. The service does not clone or freeze them; serialization and hostile-input validation belong at actual process, worker, persistence, and model boundaries.
@@ -42,23 +45,27 @@ Start-time features are advertised in `provider.capabilities` because the servic
- `toolFilter` — apply the requested child tool restriction.
- `persona` — apply a per-child persona.
Runtime features are optional methods whose presence is the capability check: `SubagentRun.steer?` delivers strictly to the actively running child turn (it throws rather than queueing when the child is not running), and `SubagentProvider.resume?` reconstructs a persisted continuable child. A run represents one disposable activation, so it deliberately has no cold-resume operation — a disposed run cannot be reconstructed after restart.
## The durable descriptor
The seam owns the versioned `subagent/descriptor` session event vocabulary (`src/descriptor.ts`): `snapshotSubagentDescriptor()` validates and detaches the declared composition before any Task exists, and `foldSubagentDescriptor()` recovers it from a loaded child log. The payload records the provider name, resolved child `agentOptions.provider`/`model`, and optional `persona`/`toolFilter` — explicit fields, never the merge-extensible `AgentOptions` object, so an unrelated extension value cannot break continuation. It omits `subagentDepth` (the persisted header's `delegationDepth` is the monotone floor) and `outputSchema` (an activation's result contract). The event is log-only: no `surfaceOp`, absent from model history, and retained by the append-only log across compaction.
## Delegation depth
The seam owns the depth vocabulary shared by implementations and consumers: the `AgentOptions.subagentDepth` declaration, `assertSubagentMaxDepth`, and `delegationDepthOf(agent)`. The persisted `SessionHeader.delegationDepth` is authoritative and monotone — runtime options may deepen the count but never lower it, so a resumed child cannot be re-counted as top-level.
Runtime features are optional methods on `SubagentRun`: `sendMessage?` steers a live child, while `resume?` asynchronously creates a continuation run. Method presence is the capability check.
`inheritsParentContext` is descriptive rather than enforceable. It says only whether the child sees completed parent conversation history (`fork` does; `spawn` and ACP do not), not whether it inherits tools, services, or authority.
## Ownership and lifecycle
`provider.start(request): Promise<SubagentRun>` is the ownership-transfer boundary. Before fulfillment, the provider owns setup and must cancel, roll back, and quiesce partial resources on every failure. After fulfillment, the caller owns the run and must call `dispose()` on every path.
`provider.start(request): Promise<SubagentRun>` is the ownership-transfer boundary. Before fulfillment, the provider owns setup and must cancel, roll back, and quiesce partial resources on every failure. After fulfillment, the caller owns the run and must call `dispose()` on every path. `provider.resume?(request)` shares the same contract for a resumed activation.
`SubagentRun.result` resolves to `{ output, structured?, stopReason }`. Child-level failures resolve with a non-`completed` reason; only an infrastructure fault that the seam cannot represent may reject. `dispose()` is idempotent, cancels remaining work, and waits for the child resources to quiesce.
A local run publishes an ordinary child agent/session before `start()` fulfills, returns that shared session id as `SubagentRun.id`, exposes the exact child as `SubagentRun.localAgent`, and records `request.parent.session.id` in the child's `parentSession` header. Remote providers instead mint a parent-scoped lifecycle id and return `localAgent: undefined`.
A local run publishes an ordinary child agent/session before `start()` fulfills, returns that shared session id as `SubagentRun.id`, exposes the exact child as `SubagentRun.localAgent`, and records `request.parent.session.id` in the child's `parentSession` header. A continuable start publishes exactly the control-allocated `continuation.sessionId`. Remote providers instead mint a parent-scoped lifecycle id and return `localAgent: undefined`.
The service emits `subagent/start` only after `start()` has fulfilled. It attaches the result observer before that synchronous notification, so even an already-settled child still produces `subagent/start` before `subagent/end`. The pair shares a service-minted `runId`; its `local` flag is snapshotted from the provider's exact `localAgent`, so observers never infer run identity or locality from reusable provider/session names.
The service emits `subagent/start` only after `start()` or `resume()` has fulfilled. It attaches the result observer before that synchronous notification, so even an already-settled child still produces `subagent/start` before `subagent/end`. The pair shares a service-minted `runId`; its `local` flag is snapshotted from the provider's exact `localAgent`, so observers never infer run identity or locality from reusable provider/session names.
Run events are scoped to the delegating parent. Every listener is independently contained: a synchronous throw or rejected returned promise is logged without starving peer listeners or changing the run.
@@ -66,17 +73,17 @@ Provider additions and removals also emit `subagent/provider-added` and `subagen
## Collection model
The model-facing tool collects synchronously by default: it awaits the child result and disposes the run before returning. Background delegation does not change this seam; the consumer registers startup and the eventual run with the generic `ctx.tasks` runtime, then collection and cancellation use the shared task tools. See the [background subagent tasks Agent Note](../../../.agents/notes/implemented/feature/2026-07-08-background-subagent-tasks.md), the [capability-seam Agent Note](../../../.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.md), and `src/types.ts` for the complete contracts.
The model-facing tool collects synchronously by default: it awaits the child result and disposes the run before returning. Background delegation does not change this seam; `@deepseek-ai/dsh-subagent-control` registers each activation with the generic `ctx.tasks` runtime, then collection and cancellation use the shared task tools. See the [background subagent tasks Agent Note](../../../.agents/notes/implemented/feature/2026-07-08-background-subagent-tasks.md), the [continuable background subagents Agent Note](../../../.agents/notes/implemented/feature/2026-07-21-continuable-background-subagents.md), the [capability-seam Agent Note](../../../.agents/notes/implemented/feature/2026-06-21-subagent-capability-seam.md), and `src/types.ts` for the complete contracts.
## Model Experience
Indirectly, through `dsh-tool-subagent`, which renders provider-specific schemas and foreground or generic-background results while child working context remains child-only.
Indirectly, through `dsh-tool-subagent` and `dsh-tool-subagent-control`, which render provider-specific schemas and foreground, background, or follow-up results while child working context remains child-only.
#### KV Cache effect
No direct invalidation; the named consumer owns any request-prefix changes.
No direct invalidation; the named consumers own any request-prefix changes.
## Known Limitations and Deferred Work
- **Runtime steering and continuation are seam-only capabilities** — `sendMessage` and `resume` have no model-facing consumer in the current tool.
- **ACP children remain one-shot** — `AcpProvider.resume` requires persisting the remote session id in provider-specific descriptor data and a per-child continuation advertisement, since ACP `loadSession` support is negotiated per child rather than established by the provider method's presence.
- **Lifecycle events are observe-only** — a run-affecting `subagent/end` continuation or decision surface waits for a concrete consumer.