mirror of
https://github.com/deepseek-ai/deepseek-harness
synced 2026-08-15 21:04:50 +00:00
c4bc6e0e38
abort() only kills the in-flight step, so a queued-but-not-yet-started prompt
ran to completion after a cancel and a prompt accepted right after could be
batched into the cancelled turn (the loop merges queued messages into one turn).
This closes TODO(rfc010-cancel-prestep) with a distinct cancel() verb.
cancel() clears the queued + steering FIFOs, aborts the in-flight step, and
drives a turn-scoped marker on the LoopHandle that the driver checks at EVERY
point a turn could start or continue:
- right after the idle wait (window 1): drop the about-to-run turn and settle
whenIdle() waiters directly (no running→idle transition fires, and no
agent/status is emitted, so an ACP listener can't see a spurious idle that
resolves a freshly-queued prompt as cancelled);
- after the synchronous setStatus('running') emit (window 2): a running listener
can cancel in the gap before runTurn;
- in the step-start window (before runStep, after setAbort): a synchronous
turn-start/step-start listener can cancel before any AbortController exists;
- at the continuation gate: a cancel during the continuation waterfall (the
finished step's controller already cleared) ends the turn aborted.
The marker is ARMED only when there is something to cancel (running, an
in-flight step, or queued/steering work) — an idle no-op cancel cannot leave it
set to drop a later prompt — and RESET unconditionally once per loop iteration,
so it governs exactly one turn and never leaks onto the next prompt (even when a
send() lands in the cancelled turn's flush window).
ACP session/cancel now maps to agent.cancel() (keeping the synchronous
settlePrompt). Teardown/disconnect still use abort('disposed') until PR D, so
the ACP README narrows the remaining best-effort window to teardown only.
Tests (agent-loop/cancel.spec.ts) cover every window unit-level (the F1 hang
guard: a whenIdle() waiter registered before a pre-step cancel resolves; the F2
leak guard: idle cancel then a prompt runs; mid-step, continuation, both
pre-step windows, turn-start-listener, steering-cleared, marker-reset). ACP
turns.spec.ts adds the through-bridge tests with NO intervening whenIdle (idle
cancel→prompt runs; mid-stream cancel→immediate next prompt runs) and updates
the stale pre-step test to the queue-aware guarantee. The existing cancel
snapshot golden is byte-identical (it drives the new cancel() path end-to-end
through the real subprocess), so no new golden is needed. 100% coverage.
80 lines
4.7 KiB
Markdown
80 lines
4.7 KiB
Markdown
# dsh-agent-loop
|
|
|
|
THE concrete agent plugin: `ReactLoopAgent` and the loop driver. Implements the `Agent` interface and drives the session/turn/step lifecycle.
|
|
|
|
This is the only package in the harness that contains concrete loop logic. Everything else is an abstract service or a plugin against extension seams — new behavior goes into plugins, not here.
|
|
|
|
## Service: `AgentLoop` (ctx key: `agentLoop`)
|
|
|
|
### Public API
|
|
|
|
- `ctx.agentLoop.create(id: string, options?: AgentOptions): ReactLoopAgent` — config-driven create: an agent on a fresh per-run session id `${id}-session-<uuid>` (no cwd). Used for `cordis.yml`-configured agents. The per-run uuid avoids colliding with the on-disk log a prior run materialized once a durable persistence backend is loaded; each run is a new session (a deliberate demo simplification — a real resume-or-create policy is a TODO). Disposed with the calling fiber.
|
|
|
|
`AgentLoop` also implements the `AgentFactory` seam and registers itself via `ctx.agents.setFactory(this)`, so plugins create/resume agents through `ctx.agents` (the interface):
|
|
|
|
- `ctx.agents.create({ agentId, sessionId, meta?, agentOptions? })` — programmatic create on a caller-supplied `sessionId` (e.g. an ACP-generated id), NOT `${id}-session`.
|
|
- `ctx.agents.resume({ agentId, resumeSessionId, agentOptions? })` — load a persisted session via `ctx.sessionPersistence` ([session persistence](../../docs/rfc/implemented/2026-06-14-session-persistence.md)) and resume an agent on it. The live session id is the resumed id; turn numbering and derived history continue from the loaded log. Requires a session-persistence backend (NOT hard-injected — non-persistent demos still work; `resume` rejects with a clear error when persistence is absent).
|
|
|
|
### Injected services
|
|
|
|
`agents`, `sessions`, `llm`, `tools`, `systemPrompt` — all five interface services.
|
|
|
|
### Configuration (schemastery)
|
|
|
|
```ts
|
|
interface Config {
|
|
agents: Array<{
|
|
id: string // required
|
|
model?: string
|
|
systemPrompt?: string
|
|
}>
|
|
}
|
|
```
|
|
|
|
Agents listed in config are auto-created at startup.
|
|
|
|
### Classes
|
|
|
|
- `ReactLoopAgent` — the concrete `Agent` implementation. Owns the inbox (`Inbox`), the per-step `AbortController`, and the loop driver. Everything observable happens through session events and the `agent/*` event taxonomy.
|
|
- `Inbox` — per-agent queued + steering FIFOs (`enqueue`, `steer`, `drainQueued`, `drainSteering`, `waitForQueued`).
|
|
|
|
### Loop lifecycle (`loop.ts`)
|
|
|
|
One invocation of `runLoop()` drives one agent for its whole lifetime:
|
|
|
|
```
|
|
forever:
|
|
wait for queued messages (idle)
|
|
TURN (error-contained):
|
|
drain queued → 'turn/start' → session('user/message')
|
|
STEP loop:
|
|
drain steering
|
|
assembly = systemPrompt.assemble()
|
|
request = waterfall agent/request
|
|
stream llm.stream(request) → session('assistant/chunk')
|
|
message = waterfall agent/step-result
|
|
session('assistant/message')
|
|
each tool-call: session('tool/call') → tools.execute() → session('tool/result')
|
|
drain steering → session('steering/message')
|
|
cont = waterfall agent/turn-continuation
|
|
if !cont: break
|
|
session('turn/end')
|
|
await session/flush
|
|
re-enqueue leftover steering as queued
|
|
idle unless more queued
|
|
```
|
|
|
|
Error containment: a throwing plugin ends the **turn**, never the loop. Dispose mid-turn emits `agent/status('disposed')` and ends with reason `disposed`. A step that hits the model's output-token ceiling makes the turn end `max-tokens` (the rule: any `max-tokens` step in the turn surfaces as `max-tokens`; `disposed`/`aborted`/`error` still take precedence) — distinct from a clean `completed` stop.
|
|
|
|
Cancellation: `agent.abort()` aborts only the in-flight step; `agent.cancel()` is the broad verb — it clears the queued + steering FIFOs, aborts the in-flight step, and drives a turn-scoped marker the driver checks at every point a turn could start or continue (right after the idle wait, after the `running` flip, before each step, and at the continuation gate) so a turn about to start is dropped. A cancelled turn ends `aborted`; a queued-but-not-started prompt never runs and cannot be batched into the cancelled turn. The marker is reset once per loop iteration, so a cancel governs exactly one turn and never leaks onto a later prompt.
|
|
|
|
### What is NOT here
|
|
|
|
Everything that goes beyond "call the model, run the tools, repeat" belongs to plugins listening on the event taxonomy:
|
|
- Hooks: `agent/request`, `agent/step-result`, `tools/execute`, `agent/turn-continuation`
|
|
- Compaction: `agent/request`
|
|
- Sandbox, permission, plan mode: `tools/execute`
|
|
- Sub-agents: TODO seam on `AgentLoop.create()`
|
|
- Persistence: `session/event` + `session/flush`
|
|
- UI: `agent/stream-chunk` + `agent/*` events
|