workflow: host-guarantee the agent-start/agent-end pairing on every stop path

agent-end was worker-authored only, so a start already forwarded to
observers lost its paired end whenever the worker could no longer speak
- the grace force-settle terminating a wedged script, or an unexpected
worker death - stranding progress consumers with agents that never
finish (ds-review-bot finding on #233).

The host now keeps a ledger of forwarded starts and funnels every
agent-end through one gate: worker-reported ends pair (and clear) their
entry, and both termination paths drain the remainder as synthesized
'cancelled' ends BEFORE the run settles, so ends always precede
workflow/end. A real settlement racing the force-settle loses to the
synthesized cancellation - the same first-wins override onResult applies
to the run's own result.
This commit is contained in:
imccyu
2026-07-09 23:05:43 +08:00
parent 053a2b9403
commit 64d0703c40
6 changed files with 164 additions and 11 deletions

View File

@@ -32,7 +32,7 @@ Values LEAVING the script (hook options/schemas, the script's return) are materi
Per-run limits: a concurrency semaphore (`maxConcurrentAgents`), a total-`agent()` cap (`maxTotalAgents`), and a per-call item cap (`maxItemsPerCall`), all config. `cancel()` posts the cancel to the worker (its hooks start throwing `CANCELLED`; the script dies at its next await) and cancels every host-side child NOW on **both seam channels** — the shared request signal aborts AND each registered child's explicit `cancel()` is called host-side, because the seam leaves a provider free to honor either channel and a worker wedged in a synchronous spin could not relay its own per-child cancel RPCs (those later land as idempotent no-ops). The grace then arms: a run still unsettled `disposeGraceMs` later force-settles `cancelled` and the worker is **terminated**. A cancellation that lands before the body runs (the ready→go handshake) reports `cancelled` without executing anything; a worker `result` racing an in-flight host cancellation reports `cancelled` too (first-wins settlement — the seam-visible result had not settled when cancellation was requested); post-cancel `phase`/`log` narration is suppressed host-side, while cancelled children still deliver their paired `agent-end`.
A worker that dies unexpectedly (an OOM, a script reaching `process.exit` through the documented vm escape) settles the run `stopReason: 'error'` with the exit diagnostics — or `'cancelled'` when a cancel was in flight — and the host-side child registry is what winds every surviving child down. `dispose()` = cancel + immediate host-driven disposal of every registered child (a wedged worker can relay no dispose RPC, so child teardown overlaps the grace instead of starting after it; the worker's own dispose RPCs join the same per-child disposal) + bounded wait (result, then child-registry quiescence, capped by the grace) + unconditional `worker.terminate()`: the thread never outlives its run. Once a run settles, stray children a script fired without awaiting are cancelled too, and `dispose()` waits for their disposal (bounded by the grace) before returning.
A worker that dies unexpectedly (an OOM, a script reaching `process.exit` through the documented vm escape) settles the run `stopReason: 'error'` with the exit diagnostics — or `'cancelled'` when a cancel was in flight — and the host-side child registry is what winds every surviving child down. `dispose()` = cancel + immediate host-driven disposal of every registered child (a wedged worker can relay no dispose RPC, so child teardown overlaps the grace instead of starting after it; the worker's own dispose RPCs join the same per-child disposal) + bounded wait (result, then child-registry quiescence, capped by the grace) + unconditional `worker.terminate()`: the thread never outlives its run. Once a run settles, stray children a script fired without awaiting are cancelled too, and `dispose()` waits for their disposal (bounded by the grace) before returning. `agent-start`/`agent-end` pairing is host-guaranteed the same way: forwarded starts live in a ledger, worker-reported ends pair them on the graceful paths, and the termination paths (grace force-settle, worker death) synthesize the missing ends (outcome `cancelled`) before the run settles — a start still in flight across the force-settle can surface after `workflow/end`, immediately paired the same way.
**Engine-specific limitations**: worker startup is paid per run; on a termination path `agentsStarted` reports the HOST-observed count (accepted `child-start`s — calls still queued worker-side for a concurrency slot are unknowable then); and a returned promise or thenable resolves per JavaScript semantics BEFORE materialization — that is what makes an un-awaited `return agent('x')` work — with the value-boundary guard applying to the resolution.

View File

@@ -22,7 +22,11 @@
* survivor when the worker dies or is terminated mid-flight. The three
* paths share ONE disposal per child (memoized by callId; the seam's
* dispose() is idempotent anyway, the memo keeps the bookkeeping and the
* containment warn single). On a termination path `agentsStarted` reports the
* containment warn single). Lifecycle pairing is host-guaranteed the same
* way: every forwarded `agent-start` lives in a ledger, and a start the
* dead or terminated worker never paired is closed by a synthesized
* `agent-end` (outcome `'cancelled'`) before the run settles. On a
* termination path `agentsStarted` reports the
* HOST-observed count (accepted `child-start` messages) — `agent()` calls
* still queued worker-side for a concurrency slot are unknowable then; the
* worker's own count rides the result message on every graceful path.
@@ -37,7 +41,7 @@ import type { Context } from 'cordis'
import type { Agent } from '@deepseek-ai/dsh-agent'
import { assertNever } from '@deepseek-ai/dsh-llm'
import type { SubagentRun } from '@deepseek-ai/dsh-subagent'
import type { WorkflowMeta, WorkflowResult, WorkflowRun, WorkflowRunId } from '@deepseek-ai/dsh-workflow'
import type { WorkflowAgentEndInfo, WorkflowAgentInfo, WorkflowMeta, WorkflowResult, WorkflowRun, WorkflowRunId } from '@deepseek-ai/dsh-workflow'
import { renderThrown } from './realm.ts'
import type { ExecutionObserver } from './runtime.ts'
import { HostToWorkerType, WorkerToHostType } from './protocol.ts'
@@ -92,6 +96,8 @@ export class WorkerRun implements WorkflowRun {
private readonly children = new Map<number, SubagentRun>()
/** In-flight child disposals by callId — the memo that gives every path (worker RPC, dispose(), reap) ONE shared disposal per child. */
private readonly childDisposals = new Map<number, Promise<void>>()
/** Started-but-not-ended agents by seq — the pairing ledger the HOST guarantees (see {@link endAgent}). */
private readonly liveAgents = new Map<number, WorkflowAgentInfo>()
private readonly quiescenceWaiters: (() => void)[] = []
/** The per-run abort fanout every child start request carries. */
private readonly controller = new AbortController()
@@ -155,6 +161,10 @@ export class WorkerRun implements WorkflowRun {
// ChildCancel relay (those later RPCs land as idempotent no-ops).
for (const run of this.children.values()) run.cancel(this.cancelReason)
this.graceTimer = setTimeout(() => {
// The worker may no longer speak (it is about to be terminated): pair
// every stranded start before the run settles, so ends precede
// workflow/end.
this.endStrandedAgents()
this.settleResult(this.cancelledResult(this.hostStarted))
void this.worker.terminate()
}, this.disposeGraceMs)
@@ -225,13 +235,15 @@ export class WorkerRun implements WorkflowRun {
if (this.cancelReason === undefined) this.observer.log(message.message)
break
case WorkerToHostType.AgentStart:
this.liveAgents.set(message.info.seq, message.info)
this.observer.agentStart(message.info)
break
case WorkerToHostType.AgentEnd:
// NOT suppressed on cancel: cancelled children report their paired
// agent-end with outcome 'cancelled' (the one-pair-per-started-child
// contract holds on every stop path).
this.observer.agentEnd(message.info)
// agent-end with outcome 'cancelled'. The gate (with the termination
// paths' synthesis) is what makes the one-pair-per-started-child
// contract hold on every stop path.
this.endAgent(message.info)
break
case WorkerToHostType.ChildStart:
this.onChildStart(message.callId, message.request)
@@ -373,6 +385,10 @@ export class WorkerRun implements WorkflowRun {
private onWorkerDeath(message: string): void {
// Whatever the worker left behind must not leak — abort + dispose it all.
if (this.children.size > 0) this.reapChildren('workflow worker gone')
// The thread is gone: no more worker-authored agent-ends can arrive —
// pair every stranded start (a start that crossed between the grace
// force-settle and this exit included) before the run settles.
this.endStrandedAgents()
// settleResult no-ops on an already-settled run (the expected exit after
// a dispose's terminate lands here too).
if (this.cancelReason !== undefined) {
@@ -382,6 +398,34 @@ export class WorkerRun implements WorkflowRun {
this.settleResult({ value: null, stopReason: 'error', error: message, agentsStarted: this.hostStarted })
}
/**
* The single agent-end emission gate: forwards `end` iff its start is still
* unpaired in the ledger, so every forwarded `workflow/agent-start` gets
* EXACTLY one `workflow/agent-end` — the worker's own report where it can
* speak, a host-synthesized one where it cannot ({@link endStrandedAgents}).
* @param end - the settlement to emit (worker-reported or synthesized).
*/
private endAgent(end: WorkflowAgentEndInfo): void {
/* v8 ignore next -- a real end still in flight across the grace force-settle: not orderable in-process */
if (!this.liveAgents.delete(end.seq)) return
this.observer.agentEnd(end)
}
/**
* Synthesize the missing `agent-end` for every started-but-unpaired agent,
* outcome `'cancelled'`: the reap cancels every child, and a real
* settlement racing the force-settle loses to the cancellation — the same
* first-wins override {@link onResult} applies to the run's own result.
* Called where the worker can no longer speak (the grace force-settle,
* worker death), BEFORE settleResult, so the paired ends reach observers
* before `workflow/end`.
*/
private endStrandedAgents(): void {
for (const info of [...this.liveAgents.values()]) {
this.endAgent({ ...info, outcome: 'cancelled' })
}
}
private cancelledResult(agentsStarted: number): WorkflowResult {
// cancel() is the only writer of cancelReason and every caller checks it
// first; the fallback guards the type, not a reachable path.

View File

@@ -597,6 +597,73 @@ describe('dsh-workflow-workerthread', () => {
// The memo: the host drive and the worker's RPC share one disposal.
expect(provider.runs[0]!.disposeCalls).toBe(1)
})
it('the grace force-settle pairs every stranded start: a host-synthesized cancelled agent-end lands before workflow/end', async () => {
const { ctx, parent, provider } = await setup({ manual: true, config: { provider: 'stub', maxConcurrentAgents: 8, disposeGraceMs: 300 } })
const ends: { seq: number; outcome: string }[] = []
const order: string[] = []
ctx.on('workflow/agent-start', (_info, agent) => { order.push(`start:${agent.seq}`) })
ctx.on('workflow/agent-end', (_info, agent) => {
ends.push({ seq: agent.seq, outcome: agent.outcome })
order.push(`end:${agent.seq}`)
})
ctx.on('workflow/end', () => { order.push('run-end') })
const handle = ctx.workflows.start({
// 'slow' starts and its agent-start crosses to observers (the awaited
// 'fast' call keeps the worker loop turning), then the script seizes
// the loop: the wedged worker can never author slow's agent-end —
// only the host's ledger can close the pair.
...scripted(`
const p = agent('slow')
await agent('fast')
const end = Date.now() + 1500
while (Date.now() < end) {}
return 'raced'
`),
parent,
})
await vi.waitFor(() => { expect(order.filter(entry => entry.startsWith('start:')).length).toBe(2) })
const fast = provider.runs.find(run => (run.request.prompt[0] as { text?: string }).text === 'fast')!
fast.settle(text('fast done'))
handle.cancel('stop now')
const result = await handle.result
expect(result.stopReason).toBe('cancelled')
// fast's end is the worker's own report; slow's is host-synthesized at
// the force-settle — exactly one end per started seq, no third event.
expect(ends).toEqual([
{ seq: 2, outcome: 'completed' },
{ seq: 1, outcome: 'cancelled' },
])
// Both ends reached observers BEFORE workflow/end: a progress consumer
// can finalize its state at run-end without dangling agents.
expect(order.indexOf('run-end')).toBe(order.length - 1)
await handle.dispose()
}, 15_000)
it('graceful cancellation keeps pairing worker-authored: exactly one agent-end per start, nothing synthesized on top', async () => {
const { ctx, parent, provider } = await setup({ manual: true })
const ends: { seq: number; outcome: string }[] = []
const order: string[] = []
ctx.on('workflow/agent-end', (_info, agent) => {
ends.push({ seq: agent.seq, outcome: agent.outcome })
order.push(`end:${agent.seq}`)
})
ctx.on('workflow/end', () => { order.push('run-end') })
const handle = ctx.workflows.start({
...scripted("await parallel([() => agent('a'), () => agent('b')])\nreturn 'unreachable'"),
parent,
})
await vi.waitFor(() => { expect(provider.runs.length).toBe(2) })
handle.cancel('user stop')
const result = await handle.result
expect(result.stopReason).toBe('cancelled')
// The live worker reported both pairs itself; the ledger must not add
// a synthesized duplicate on any path that settles inside the grace.
expect(ends.map(end => end.outcome)).toEqual(['cancelled', 'cancelled'])
expect(new Set(ends.map(end => end.seq)).size).toBe(2)
expect(order.indexOf('run-end')).toBe(order.length - 1)
await handle.dispose()
})
})
describe('worker death', () => {
@@ -669,6 +736,45 @@ describe('dsh-workflow-workerthread', () => {
await handle.dispose()
}, 15_000)
it('a worker death pairs every stranded start: the synthesized cancelled agent-end precedes the error workflow/end', async () => {
const { ctx, parent, provider } = await setup({ manual: true })
const ends: { seq: number; outcome: string }[] = []
const order: string[] = []
ctx.on('workflow/agent-start', (_info, agent) => { order.push(`start:${agent.seq}`) })
ctx.on('workflow/agent-end', (_info, agent) => {
ends.push({ seq: agent.seq, outcome: agent.outcome })
order.push(`end:${agent.seq}`)
})
ctx.on('workflow/end', () => { order.push('run-end') })
const handle = ctx.workflows.start({
// Same choreography as the force-settle pairing test, but the worker
// DIES (the documented vm escape) instead of being terminated: the
// exit path must close slow's pair from the ledger too. The escaped
// setTimeout lets the already-posted messages flush before the kill.
...scripted(`
const p = agent('slow')
await agent('fast')
const proc = ${ESCAPE}
const st = globalThis.constructor.constructor('return setTimeout')()
await new Promise(resolve => st(resolve, 150))
proc.exit(7)
`),
parent,
})
await vi.waitFor(() => { expect(order.filter(entry => entry.startsWith('start:')).length).toBe(2) })
const fast = provider.runs.find(run => (run.request.prompt[0] as { text?: string }).text === 'fast')!
fast.settle(text('fast done'))
const result = await handle.result
expect(result.stopReason).toBe('error')
expect(result.error).toContain('exit code 7')
expect(ends).toEqual([
{ seq: 2, outcome: 'completed' },
{ seq: 1, outcome: 'cancelled' },
])
expect(order.indexOf('run-end')).toBe(order.length - 1)
await handle.dispose()
}, 15_000)
it('a dispose ack racing the worker death is dropped, not crashed (post after exit)', async () => {
// Slow child disposal: the ack resolves only AFTER the worker died, so
// it has nowhere to go and must be dropped silently (the workerGone

View File

@@ -85,7 +85,10 @@ declare module 'cordis' {
'workflow/agent-start'(info: WorkflowRunInfo, agent: WorkflowAgentInfo): void
/**
* One `agent()` call settled (clean result, child failure, or run
* cancellation). Paired with {@link Events['workflow/agent-start']}.
* cancellation). Paired with {@link Events['workflow/agent-start']} by
* `agent.seq`, exactly once per started call on every stop path — on an
* engine termination path (a worker killed past its grace) the end is
* engine-synthesized with outcome `'cancelled'`.
* @param info - the run's identity snapshot.
* @param agent - the call identity plus its outcome.
* @mode emit