@namzu/sdk 39.0.0 → 41.0.0

package/src/runtime/jobs/awaited-jobs.ts

@@ -0,0 +1,271 @@
+import type { BackgroundJob, BackgroundJobRegistry } from './registry.js'
+
+/**
+ * What this adapter needs from a registry: the jobs' exits, and one job's
+ * current status. Structural rather than the class, for the reason
+ * `BackgroundJobRegistryRef` is: the kernel passes the real registry, and a
+ * test that needs two exits and nothing else should not have to spawn
+ * processes to produce them.
+ */
+export type AwaitedJobSource = Pick<BackgroundJobRegistry, 'get' | 'onExit'>
+
+/** Exits and the text that puts them in front of the model. See {@link AwaitedJobs.takeDelivery}. */
+export interface DeliverableJobExits {
+	/** The notice text, already queued by the exit that produced it. */
+	readonly text: string
+	/** The exits that text accounts for, taken from the record. */
+	readonly exits: readonly BackgroundJob[]
+}
+
+/**
+ * Background jobs the model SAID it is waiting on.
+ *
+ * `CompletionInbox` gives a run a bounded, zero-token wait for a delegated
+ * task nobody is blocked on, and `holdForOutstandingWork` spends it: the
+ * model stops calling tools, the loop races the inbox against operator input,
+ * and the run only settles once the result is in the transcript or the grace
+ * is gone. Background shell jobs had none of it — so a model that started a
+ * job and then had nothing left to do improvised, and the recorded run
+ * (research/resident/results/2026-09-14-exploration-policy-terra-tui.json)
+ * shows what that costs: six `job read` polls, three `job list` polls and a
+ * `sleep 30`, each one a full context resend.
+ *
+ * This is the same three-method shape the hold already races —
+ * {@link hasPendingWork}, {@link waitForArrival}, {@link drain} — over the
+ * job registry instead of the task gateway.
+ *
+ * **Intent is stated, never inferred.** A job is only outstanding here once
+ * `wait_for_job` named it (see {@link expect}); job existence means nothing.
+ * That is the same distinction `CompletionInbox.expect` draws for a
+ * background task, and it is what keeps a dev server or a file watcher —
+ * started precisely so it would keep running — from holding every run of the
+ * session open for its grace period. The model that wants to wait says so.
+ */
+export class AwaitedJobs {
+	/** Awaited and still running. Nothing else can hold a run open. */
+	private readonly outstanding = new Set<string>()
+	/** Exits since the last {@link drain}. */
+	private exits: BackgroundJob[] = []
+	private readonly arrivals = new Set<() => void>()
+	private detach?: () => void
+
+	constructor(
+		private readonly source: AwaitedJobSource,
+		/** Whose jobs these are — the run or the session the jobs are bound to. */
+		private readonly owner: string,
+		/**
+		 * Whether the exit notice this run queues for the model is still
+		 * unread. A second opinion on {@link noticesDelivered}, and it can
+		 * only ever narrow what counts as pending.
+		 *
+		 * An exit is work only until the model has seen it, and this adapter is
+		 * not what shows it: the text rides out on the next tool result
+		 * (`attachNotice`), which is what normally happens, long before any
+		 * hold opens. {@link noticesDelivered} is how that delivery says so,
+		 * and it is the per-job record — the entry goes, so nothing can count
+		 * it again.
+		 *
+		 * This stays because the two answer different questions. That one is
+		 * reported BY a delivery site, so it is only as good as the sites that
+		 * remember to call it; this reads the channel itself, and a channel
+		 * with nothing queued cannot have news in it whatever the record says.
+		 * It is also safe in one direction only: it can suppress a hold, never
+		 * open one, so a stale `true` here is not a turn the model gets to
+		 * spend — which is the failure this pair exists to prevent.
+		 *
+		 * `CompletionInbox.claim` draws the same line on the task side. It can
+		 * name the task, because the call that delivers a completion knows
+		 * which one it delivered; a notice is text by the time it reaches the
+		 * tool result, so this asks the channel instead.
+		 *
+		 * Absent means nothing else consumes the notices, so a recorded exit is
+		 * unread by definition.
+		 */
+		private readonly unreadNotice?: () => boolean,
+	) {}
+
+	/**
+	 * Start listening.
+	 *
+	 * Returns the detach, and attaching twice is a no-op rather than a second
+	 * subscription — the same rule `CompletionInbox.attach` follows, for the
+	 * same reason: two subscriptions would record each exit twice.
+	 */
+	attach(): () => void {
+		if (this.detach) return this.detach
+		this.detach = this.source.onExit((job) => {
+			// `onExit` is a broadcast over a registry a host may share between
+			// runs, and the filter is the same one the exit-notice channel
+			// applies in `runtime/query/index.ts`.
+			if (job.owner !== this.owner) return
+			// `outstanding` is the gate, so an id nobody awaited is ignored and
+			// a second announcement for the same job cannot queue twice.
+			if (!this.outstanding.delete(job.id)) return
+			this.exits.push(job)
+			for (const wake of [...this.arrivals]) wake()
+		})
+		return this.detach
+	}
+
+	/**
+	 * Say that the model is waiting on this job's exit.
+	 *
+	 * Called by `wait_for_job`, and by nothing else: the wait is the whole
+	 * signal. A job that has ALREADY stopped is not recorded, because the
+	 * call that marks it is the same call that returns its output — holding
+	 * the run open afterwards would buy a turn to read a result the model has
+	 * just read. `CompletionInbox.expect` skips an already-claimed task for
+	 * the same reason.
+	 *
+	 * An id the registry does not know is ignored rather than raised: the
+	 * caller is about to fail on it anyway, and a marking call is not the
+	 * place that decides what an unknown job means.
+	 */
+	expect(id: string): void {
+		let job: BackgroundJob
+		try {
+			job = this.source.get(id)
+		} catch {
+			return
+		}
+		if (job.owner !== this.owner) return
+		if (job.status !== 'running') return
+		this.outstanding.add(id)
+	}
+
+	/**
+	 * Recorded exits the model has not been shown yet.
+	 *
+	 * Both halves of the same question: {@link noticesDelivered} drops the
+	 * entry when its notice goes out, and the `unreadNotice` constructor
+	 * argument asks the channel whether anything is still queued. A queued
+	 * exit whose notice has already been delivered is history, not pending
+	 * work, and an exit nobody awaited is not this adapter's news to carry.
+	 */
+	private get unreadExits(): boolean {
+		if (this.exits.length === 0) return false
+		return this.unreadNotice?.() ?? true
+	}
+
+	/** Whether an awaited job is still running, or one has exited unread. */
+	get hasPendingWork(): boolean {
+		return this.unreadExits || this.outstanding.size > 0
+	}
+
+	/**
+	 * Wait for the next awaited job's exit, deadline or abort, whichever comes
+	 * first. Aborting releases only this waiter; the jobs are untouched.
+	 *
+	 * Bounded by the caller, exactly as `CompletionInbox.waitForArrival` is: a
+	 * job that never exits — a dev server someone did await — must not keep a
+	 * run open, and only the run's own budget knows how long is long enough.
+	 */
+	waitForArrival(timeoutMs: number, signal?: AbortSignal): Promise<void> {
+		if (signal?.aborted) return Promise.resolve()
+		if (this.unreadExits) return Promise.resolve()
+		if (this.outstanding.size === 0) return Promise.resolve()
+
+		return new Promise((resolve) => {
+			const finish = (): void => {
+				clearTimeout(timer)
+				this.arrivals.delete(finish)
+				signal?.removeEventListener('abort', finish)
+				resolve()
+			}
+			const timer = setTimeout(finish, timeoutMs)
+			// `unref` where the runtime has it, so a pending wait never keeps a
+			// process alive past the job it was waiting for.
+			;(timer as { unref?: () => void }).unref?.()
+
+			this.arrivals.add(finish)
+			signal?.addEventListener('abort', finish, { once: true })
+			if (signal?.aborted) finish()
+		})
+	}
+
+	/**
+	 * Say that the queued exit notices have been put in front of the model.
+	 *
+	 * Called by the tool-result delivery (`attachNotice`), which is where an
+	 * exit normally reaches the model; {@link takeDelivery} does the same for
+	 * the hold's own path by taking the entries as it delivers them.
+	 *
+	 * Every recorded exit goes, and that is per-job rather than a blunt
+	 * clear: the channel hands over ALL its queued text at once, and an exit
+	 * is recorded here in the same synchronous announcement that queues its
+	 * notice, so the text just delivered is exactly the notices of the exits
+	 * in hand. An exit recorded after this call keeps its entry, because its
+	 * notice was queued after that text was taken.
+	 *
+	 * Without it the entry outlived the delivery, and the channel-level gate
+	 * above was the only thing standing between a read exit and a model turn
+	 * — a gate that reopens the moment ANY later job, awaited or not, queues
+	 * a notice of its own.
+	 */
+	noticesDelivered(): void {
+		this.exits = []
+	}
+
+	/**
+	 * Take the recorded exits together with the text that delivers them, or
+	 * take neither.
+	 *
+	 * The pairing is the whole method. Draining the exits first and then
+	 * asking for the notice loses them on the branch that finds none: the
+	 * records are gone, no text was written, and the exit the run held itself
+	 * open for is delivered by nobody. So {@link drain} is not called until
+	 * the delivery is certain, and an exit this could not deliver stays in
+	 * hand for the next one.
+	 */
+	takeDelivery(notice: () => string | undefined): DeliverableJobExits | undefined {
+		if (this.exits.length === 0) return undefined
+		const text = notice()
+		if (text === undefined) return undefined
+		return { text, exits: this.drain() }
+	}
+
+	/**
+	 * Take every exit recorded since the last call, leaving none behind.
+	 *
+	 * Draining rather than peeking, for the reason `CompletionInbox.drain`
+	 * gives: an exit that stays queued after being delivered is a duplicate
+	 * waiting to happen. Callers that are delivering want
+	 * {@link takeDelivery}, which will not take them without the notice that
+	 * accounts for them.
+	 */
+	drain(): readonly BackgroundJob[] {
+		if (this.exits.length === 0) return []
+		const exits = this.exits
+		this.exits = []
+		return exits
+	}
+
+	/**
+	 * Awaited jobs still running.
+	 *
+	 * Read when a run ends, so it can say which wait it walked away from.
+	 * Nothing here is stopped by being read — the ids are a statement, and a
+	 * job's lifetime belongs to whoever owns it.
+	 */
+	get outstandingJobIds(): readonly string[] {
+		return [...this.outstanding]
+	}
+
+	/**
+	 * Stop listening. Safe to call more than once.
+	 *
+	 * A run that ends without this leaves its listener on a registry the host
+	 * reuses across runs — the leak `CompletionInbox.close` exists to prevent,
+	 * on the other subsystem.
+	 */
+	close(): void {
+		this.detach?.()
+		this.detach = undefined
+		this.outstanding.clear()
+		this.exits = []
+		// Release anyone still waiting. A closed adapter would otherwise hold
+		// them to their own deadline for an exit that can no longer reach them.
+		for (const wake of [...this.arrivals]) wake()
+		this.arrivals.clear()
+	}
+}

package/src/runtime/jobs/registry.ts

@@ -3,6 +3,7 @@ import { spawn } from 'node:child_process'
 import { SANDBOX_KILL_GRACE_MS } from '../../constants/sandbox/index.js'
 import { killTree } from '../../process/kill-tree.js'
 import { scrubInheritedEnv } from '../../tools/env-scrub.js'
+import { awaitWithAbort } from '../../utils/await-with-abort.js'
 
 /**
  * Work that outlives a tool call, owned by whoever started it.
@@ -310,6 +311,32 @@ export class BackgroundJobRegistry {
 		return entry.record
 	}
 
+	/**
+	 * Await one job's exit — the public counterpart to `onExit`, for a
+	 * caller that wants a single result rather than a standing
+	 * subscription. Resolves at once for a job that has already stopped, so
+	 * a caller that lost the race against a fast-finishing job never blocks
+	 * on a promise that would otherwise never settle.
+	 *
+	 * `signal` is honoured: an aborted wait rejects and detaches rather than
+	 * holding the internal exit promise's continuation open for a job that
+	 * may run for another hour. That only ends the WAIT — the job itself is
+	 * untouched either way, exactly as a timed-out `kill`-less wait leaves
+	 * it. See `wait-for-job-bounds.ts`, the first caller.
+	 */
+	waitForExit(id: string, opts: { signal?: AbortSignal } = {}): Promise<BackgroundJob> {
+		const entry = this.jobs.get(id)
+		if (!entry) throw new UnknownBackgroundJobError({ id })
+		if (entry.record.status !== 'running') return Promise.resolve(entry.record)
+		// `entry` is the same object the `finalize` closure in `start()`
+		// mutates in place, so reading `entry.record` after `exit` settles
+		// sees the final status — the same trick `kill()` already relies on.
+		return awaitWithAbort(
+			entry.exit.then(() => entry.record),
+			opts.signal,
+		)
+	}
+
 	/**
 	 * Output since `fromOffset`, with what the cap dropped stated.
 	 *
@@ -410,6 +437,17 @@ export function bindOwner(
 			readonly workingDirectory: string
 			readonly env?: Record<string, string>
 		}) => JobProcess
+		/**
+		 * Be told that the model said it is waiting on a job's exit.
+		 *
+		 * The registry does not keep this: wait-intent belongs to the RUN that
+		 * expressed it, not to a registry a host may share across runs and
+		 * sessions. The run passes its own recorder here — `AwaitedJobs`, which
+		 * is what the iteration loop holds open for. Absent means nobody is
+		 * listening, and `markAwaited` is then absent from the bound ref rather
+		 * than present and silently doing nothing.
+		 */
+		readonly onAwaited?: (id: string) => void
 	} = {},
 ) {
 	const mine = (id: string): BackgroundJob => {
@@ -443,6 +481,18 @@ export function bindOwner(
 			mine(id)
 			return registry.read(id, opts ?? {})
 		},
+		waitForExit: (id: string, opts?: { signal?: AbortSignal }) => {
+			mine(id)
+			return registry.waitForExit(id, opts ?? {})
+		},
+		...(defaults.onAwaited
+			? {
+					markAwaited: (id: string) => {
+						mine(id)
+						defaults.onAwaited?.(id)
+					},
+				}
+			: {}),
 		kill: async (id: string) => {
 			mine(id)
 			return await registry.kill(id)

package/src/runtime/query/executor.ts

@@ -51,6 +51,8 @@ import type { Logger } from '../../utils/logger.js'
 import { compressShellOutput } from '../../utils/shell-compress.js'
 import { type BackgroundJobRegistry, type JobProcess, bindOwner } from '../jobs/registry.js'
 import { describeVisibleFileEvidence } from './file-evidence-context.js'
+import { seedObservationLedger } from './file-evidence-seed.js'
+import { skippedToolResultText } from './plugin-hooks.js'
 import type { ToolResultObservation } from './project-instructions.js'
 import { ToolCallBudget, assertMaxToolCalls } from './tool-call-budget.js'
 import {
@@ -334,6 +336,16 @@ export interface ToolExecutorConfig {
 	 * session and stops them itself when the session ends.
 	 */
 	backgroundJobOwner?: string
+	/**
+	 * Where `wait_for_job` records that the model is waiting on a job.
+	 *
+	 * A callback rather than the recorder itself: the executor's part is to
+	 * hand the tools a bound ref, and what the run does with the intent —
+	 * hold itself open for the job — is the iteration loop's business. Absent
+	 * means the bound ref has no `markAwaited` at all, so a host that wires no
+	 * recorder gets no hold rather than a marking call that goes nowhere.
+	 */
+	onJobAwaited?: (id: string) => void
 
 	/**
 	 * Where the `skill` tool reads from.
@@ -514,6 +526,8 @@ export class ToolExecutor {
 	/** Set per turn by the orchestrator; see {@link setStepAllowedTools}. */
 	private stepAllowedTools?: readonly string[]
 	private readonly fileReadTracker: FileReadTracker
+	/** A ledger is rebuilt from history at most once; a second pass would re-append its chains. */
+	private fileObservationsSeeded = false
 
 	constructor(
 		config: ToolExecutorConfig,
@@ -547,6 +561,39 @@ export class ToolExecutor {
 		this.config = { ...this.config, sandbox }
 	}
 
+	/**
+	 * Rebuild this run's observation ledger from history a resume restored.
+	 *
+	 * Once, and only from a history that has already been repaired — the ledger
+	 * has to describe what the model is about to be shown, not what was
+	 * checkpointed before the repair removed an abandoned call. `sandboxed` is
+	 * passed rather than read off this executor's config because a resumed run
+	 * restores its history before it acquires a sandbox, so the config does not
+	 * know yet what the run's tool paths will be keyed on.
+	 *
+	 * Awaited, and the only filesystem work anywhere in this feature: the seed
+	 * has to write its entries under the keys the mutation tools will look them
+	 * up under, which on a host means resolving each path through its symlinks
+	 * the way `write` and `edit` do. No file's content is read.
+	 */
+	async seedFileObservations(messages: readonly Message[], sandboxed: boolean): Promise<void> {
+		if (this.fileObservationsSeeded) return
+		this.fileObservationsSeeded = true
+		const report = await seedObservationLedger(messages, this.fileReadTracker, {
+			workingDirectory: this.config.workingDirectory,
+			...(this.config.additionalDirectories
+				? { additionalDirectories: this.config.additionalDirectories }
+				: {}),
+			sandboxed,
+		})
+		this.log.info('Rebuilt the file observation ledger from restored history', {
+			[NAMZU.RUN_ID]: this.config.runId,
+			'namzu.files.witnessed': report.pathsWitnessed,
+			'namzu.files.seen': report.pathsSeen,
+			'namzu.files.replayed_units': report.unitsReplayed,
+		})
+	}
+
 	/** Request-only evidence from the same ledger used by mutation admission. No filesystem I/O. */
 	describeFileEvidence(messages: readonly Message[]): string | undefined {
 		return describeVisibleFileEvidence(
@@ -1342,6 +1389,7 @@ export class ToolExecutor {
 												) as JobProcess,
 										}
 									: {}),
+								...(this.config.onJobAwaited ? { onAwaited: this.config.onJobAwaited } : {}),
 							},
 						),
 					}
@@ -2265,7 +2313,7 @@ export class ToolExecutor {
 					return {
 						kind: 'skip',
 						input: currentInput,
-						output: `Tool ${toolName} skipped by plugin: ${result.reason}`,
+						output: skippedToolResultText(toolName, result.reason),
 					}
 				case 'error':
 					return {