@namzu/sdk 12.2.0 → 13.1.0

package/src/manager/agent/__tests__/an-env-reaches-the-child-it-was-set-for.test.ts

@@ -0,0 +1,196 @@
+import { describe, expect, it } from 'vitest'
+
+import { EMPTY_TOKEN_USAGE } from '../../../constants/limits.js'
+import { AgentRegistry } from '../../../registry/agent/definitions.js'
+import { DefaultCapacityValidator } from '../../../session/handoff/capacity.js'
+import { SessionSummaryMaterializer } from '../../../session/summary/materialize.js'
+import { WorkspaceBackendRegistry } from '../../../session/workspace/registry.js'
+import { InMemorySessionStore } from '../../../store/session/memory.js'
+import { InMemoryThreadStore } from '../../../store/thread/memory.js'
+import type { BaseAgentConfig, BaseAgentResult } from '../../../types/agent/base.js'
+import type { Agent } from '../../../types/agent/core.js'
+import type { AgentDefinition } from '../../../types/agent/factory.js'
+import type { AgentTaskContext } from '../../../types/agent/task.js'
+import type { AgentId, TenantId } from '../../../types/ids/index.js'
+import type { SummaryId } from '../../../types/session/ids.js'
+import { ZERO_COST } from '../../../utils/cost.js'
+import { ThreadManager } from '../../thread/lifecycle.js'
+import { AgentManager } from '../lifecycle.js'
+
+/**
+ * A delegate registered the normal way inherited no environment at all.
+ *
+ * `AgentManager` builds a child's config on two branches. The bare-config
+ * branch — taken only when a definition has NO `configBuilder` — has always
+ * carried `env`. The `configBuilder` branch, which is what a host registering
+ * a real agent actually uses, never stamped it. So a run given an environment
+ * handed its delegates none of it.
+ *
+ * This is the third field to go the same way: `parentSpan` and `resumeHandler`
+ * are both stamped after the builder returns, each with a comment saying the
+ * builder is written by whoever registered the agent and cannot be expected to
+ * forward something it was never told about. `env` was missed, and it went
+ * unnoticed because a missing environment does not fail — the child just runs
+ * somewhere else.
+ *
+ * These drive the real `AgentManager` rather than re-implementing the merge.
+ * A test that restates the logic proves the logic agrees with itself; what has
+ * to hold is that the manager applies it.
+ */
+
+const TENANT = 'tnt_env' as TenantId
+
+/** Records the config it is run with, so the assertion is on what shipped. */
+function recordingAgent(seen: { config?: BaseAgentConfig }) {
+	return {
+		type: 'reactive',
+		metadata: {
+			id: 'worker',
+			name: 'worker',
+			version: '1.0.0',
+			category: 'general',
+			description: 'records its config',
+			type: 'reactive',
+			capabilities: {},
+		},
+		async run(_input: unknown, config: BaseAgentConfig): Promise<BaseAgentResult> {
+			seen.config = config
+			return {
+				runId: 'run_child',
+				status: 'completed',
+				result: 'ok',
+				usage: { ...EMPTY_TOKEN_USAGE },
+				cost: { ...ZERO_COST },
+				iterations: 1,
+				durationMs: 0,
+				messages: [],
+			} as BaseAgentResult
+		},
+		async cancel() {},
+		getCapabilities() {
+			return {} as never
+		},
+	} as unknown as Agent<BaseAgentConfig, BaseAgentResult>
+}
+
+/** A definition WITH a configBuilder — the branch that dropped `env`. */
+function definitionWithBuilder(
+	agent: Agent<BaseAgentConfig, BaseAgentResult>,
+	builderEnv?: Record<string, string>,
+): AgentDefinition {
+	return {
+		info: {
+			id: 'worker',
+			name: 'worker',
+			version: '1.0.0',
+			category: 'general',
+			description: 'a worker',
+			tools: [],
+			defaults: { model: 'test', tokenBudget: 1_000 },
+		},
+		typedAgent: agent,
+		configBuilder: () =>
+			({
+				model: 'test',
+				tokenBudget: 1_000,
+				timeoutMs: 10_000,
+				...(builderEnv ? { env: builderEnv } : {}),
+			}) as BaseAgentConfig,
+	} as AgentDefinition
+}
+
+async function spawnWith(options: {
+	builderEnv?: Record<string, string>
+	overrideEnv?: Record<string, string>
+}): Promise<BaseAgentConfig | undefined> {
+	const seen: { config?: BaseAgentConfig } = {}
+	const store = new InMemorySessionStore()
+	const threadStore = new InMemoryThreadStore()
+	const project = await store.createProject({ tenantId: TENANT, name: 'p' }, TENANT)
+	const thread = await threadStore.createThread({ projectId: project.id, title: 't' }, TENANT)
+	const parentActor = { kind: 'agent', agentId: 'sup' as AgentId, tenantId: TENANT } as const
+	const parentSession = await store.createSession(
+		{ threadId: thread.id, projectId: project.id, currentActor: parentActor },
+		TENANT,
+	)
+	await store.updateSession({ ...parentSession, status: 'active' }, TENANT)
+
+	const registry = new AgentRegistry()
+	registry.register(definitionWithBuilder(recordingAgent(seen), options.builderEnv))
+
+	let n = 0
+	const manager = new AgentManager(registry, undefined, {
+		sessionStore: store,
+		threadManager: new ThreadManager({ threadStore, sessionStore: store }),
+		workspaceRegistry: new WorkspaceBackendRegistry(),
+		capacity: new DefaultCapacityValidator(store),
+		summaryMaterializer: new SessionSummaryMaterializer({
+			store,
+			generateSummaryId: () => `sum_${++n}` as SummaryId,
+		}),
+	})
+
+	const context: AgentTaskContext = {
+		parentRunId: 'run_parent' as never,
+		parentAgentId: 'sup',
+		parentAbortController: new AbortController(),
+		depth: 0,
+		budgetTracker: { total: 100_000, remaining: 100_000 },
+		tenantId: TENANT,
+		threadId: thread.id,
+		sessionId: parentSession.id,
+		projectId: project.id,
+		parentActor,
+	} as AgentTaskContext
+
+	const task = await manager.sendMessage(
+		{
+			agentId: 'worker',
+			input: { messages: [], workingDirectory: '/tmp' } as never,
+			parentSessionId: parentSession.id,
+			tenantId: TENANT,
+			projectId: project.id,
+			parentActor,
+			...(options.overrideEnv ? { configOverrides: { env: options.overrideEnv } } : {}),
+		} as never,
+		context,
+	)
+
+	await manager.waitForCompletion(task.taskId)
+	return seen.config
+}
+
+describe('an environment survives a configBuilder that never heard of it', () => {
+	it('reaches a child whose builder sets none', async () => {
+		const config = await spawnWith({ overrideEnv: { API_BASE: 'https://staging' } })
+
+		expect(config?.env).toEqual({ API_BASE: 'https://staging' })
+	})
+
+	it('keeps the builder keys the caller did not restate', async () => {
+		// Why this is a merge and not an assignment: `configOverrides` is a
+		// Partial, so replacing the map would drop everything the builder set,
+		// and a caller overriding one variable is not saying the rest should
+		// vanish.
+		const config = await spawnWith({
+			builderEnv: { REGION: 'eu', TIER: 'free' },
+			overrideEnv: { TIER: 'paid' },
+		})
+
+		expect(config?.env).toEqual({ REGION: 'eu', TIER: 'paid' })
+	})
+
+	it('leaves the builder environment alone when nothing overrides it', async () => {
+		const config = await spawnWith({ builderEnv: { REGION: 'eu' } })
+
+		expect(config?.env).toEqual({ REGION: 'eu' })
+	})
+
+	it('gives a child with no environment undefined rather than an empty map', async () => {
+		// An agent that never had an environment must not gain an empty object
+		// that every reader then has to tell apart from a real one.
+		const config = await spawnWith({})
+
+		expect(config?.env).toBeUndefined()
+	})
+})

package/src/manager/agent/lifecycle.ts

@@ -82,6 +82,25 @@ interface ChildSpawnRecord {
 	workspaceRef?: WorkspaceRef
 }
 
+/**
+ * Combine a child's own environment with what its parent passed down.
+ *
+ * Per key, override winning — not whole-value replacement, which would drop
+ * every key a `configBuilder` set and the caller did not happen to restate.
+ * `configOverrides` is a `Partial`, so replacement reads as "the caller
+ * supplied an environment" when what they supplied was one variable.
+ *
+ * Returns `undefined` when both sides are empty, so an agent that never had an
+ * environment does not gain an empty object it then has to be checked for.
+ */
+function mergeEnv(
+	base: Readonly<Record<string, string>> | undefined,
+	override: Readonly<Record<string, string>> | undefined,
+): Record<string, string> | undefined {
+	if (!base && !override) return undefined
+	return { ...base, ...override }
+}
+
 export class AgentManager {
 	private registry: AgentRegistry
 	private instances: Map<TaskId, AgentTask> = new Map()
@@ -289,6 +308,23 @@ export class AgentManager {
 			// covered the top-level run and nothing it delegated.
 			const inheritedHandler = options.configOverrides?.resumeHandler ?? context.resumeHandler
 			if (inheritedHandler) childConfig.resumeHandler = inheritedHandler
+
+			// And the environment, for the third time and the same reason.
+			//
+			// The bare-config branch below has always carried `env`; this one
+			// never did, so a delegate registered WITH a `configBuilder` — the
+			// normal way, and what every host in this repo does — silently ran
+			// with none of the environment its parent had been given. The
+			// builder is written by whoever registered the agent and cannot be
+			// expected to forward a field it was never told about, which is
+			// exactly why `parentSpan` and `resumeHandler` are stamped here too.
+			//
+			// Merged per key rather than replaced. `configOverrides` is a
+			// `Partial`, so assigning the whole map would drop every key the
+			// builder set and the caller did not restate — the override wins per
+			// key, the same direction it already wins for `model` and `effort`.
+			const inheritedEnv = mergeEnv(childConfig.env, options.configOverrides?.env)
+			if (inheritedEnv) childConfig.env = inheritedEnv
 		} else {
 			this.log.warn('No configBuilder, using bare config', {
 				agentId: options.agentId,
@@ -641,6 +677,24 @@ export class AgentManager {
 			if (this.deps.workspaceRegistry.has(backend)) {
 				const driver = this.deps.workspaceRegistry.get(backend)
 				workspaceRef = await driver.create({ label: subSession.id })
+
+				// Write the workspace onto the record that outlives this process.
+				//
+				// The ref was kept only on the in-memory `ChildSpawnRecord`, so
+				// `SubSession.workspaceId` stayed `null` for every spawn-created
+				// child — and `ArchivalManager` resolves a workspace only when
+				// that field is set (`session/retention/archive.ts`). The one
+				// record that could have named the workspace said there was none,
+				// which is why the archival path could never act as a backstop
+				// for a leaked worktree.
+				//
+				// After `create` rather than in `createSubSession`, because the
+				// sub-session id is the workspace's label — the workspace cannot
+				// exist before the record it is named after. Inside the try, so
+				// the compensating rollback below covers it like every other
+				// mutation here.
+				subSession = { ...subSession, workspaceId: workspaceRef.id }
+				await store.updateSubSession(subSession, context.tenantId)
 			}
 		} catch (err) {
 			// Compensating rollback order is mandated by the store's
@@ -764,9 +818,9 @@ export class AgentManager {
 						)
 					}
 				} else {
-					// Non-success: mark sub-session failed and, when we own a
-					// workspace, dispose it. Dispose errors are logged but not
-					// propagated — the sub-session state is already persisted.
+					// Non-success: mark sub-session failed. Disposal is shared with
+					// the success branch below — the workspace was provisioned for
+					// this child either way, and it is this manager that owns it.
 					const subSession = await store.getSubSession(
 						spawnRecord.subSessionId,
 						spawnRecord.tenantId,
@@ -774,21 +828,25 @@ export class AgentManager {
 					if (subSession) {
 						await store.updateSubSession({ ...subSession, status: 'failed' }, spawnRecord.tenantId)
 					}
-					if (spawnRecord.workspaceRef) {
-						const backend = spawnRecord.workspaceRef.meta.backend
-						if (this.deps.workspaceRegistry.has(backend)) {
-							await this.deps.workspaceRegistry
-								.get(backend)
-								.dispose(spawnRecord.workspaceRef)
-								.catch((disposeErr) => {
-									this.log.warn('Workspace dispose failed', {
-										backend,
-										error: toErrorMessage(disposeErr),
-									})
-								})
-						}
-					}
 				}
+
+				// A child is done with its workspace however it ended.
+				//
+				// This used to live only in the branch above, so both dispose
+				// sites in this class were failure paths and a child that
+				// SUCCEEDED released nothing. `.namzu/worktrees/` then grew once
+				// per successful delegation — the more reliable the workers, the
+				// faster it filled, which is the opposite of the signal a leak
+				// usually gives.
+				//
+				// It runs after the summary is sealed and the sub-session flipped
+				// to `idle`, so nothing the terminalization path reads is gone
+				// before it reads it. It also runs BEFORE the `subsession_idled`
+				// emission below: a listener cannot reach into the workspace from
+				// that event. Stated rather than hedged — no consumer does today,
+				// and holding a worktree open for a hypothetical one is what this
+				// is fixing.
+				await this.disposeChildWorkspace(spawnRecord)
 			} catch (err) {
 				this.log.error('Sub-session finalization failed', {
 					taskId: agentTask.taskId,
@@ -937,15 +995,38 @@ export class AgentManager {
 				spawnRecord.tenantId,
 			)
 		}
-		if (spawnRecord.workspaceRef) {
-			const backend = spawnRecord.workspaceRef.meta.backend
-			if (this.deps.workspaceRegistry.has(backend)) {
-				await this.deps.workspaceRegistry
-					.get(backend)
-					.dispose(spawnRecord.workspaceRef)
-					.catch(() => undefined)
-			}
-		}
+		await this.disposeChildWorkspace(spawnRecord)
+	}
+
+	/**
+	 * Release the workspace this manager provisioned for a child.
+	 *
+	 * Called on every terminal path, success included. `has(backend)` before
+	 * `get(backend)` because the registry is deny-by-default and throws on an
+	 * unknown kind — a driver deregistered mid-run must not turn cleanup into
+	 * an exception on a child that already finished.
+	 *
+	 * Never throws. Disposal is cleanup, not part of the child's result: the
+	 * sub-session state is already persisted by the time this runs, and
+	 * failing here would report a delegation that worked as one that did not.
+	 * The failure is logged instead, because a worktree that could not be
+	 * removed is an operator's problem and silence is how it stays one.
+	 */
+	private async disposeChildWorkspace(spawnRecord: ChildSpawnRecord): Promise<void> {
+		if (!spawnRecord.workspaceRef) return
+		const backend = spawnRecord.workspaceRef.meta.backend
+		if (!this.deps.workspaceRegistry.has(backend)) return
+		await this.deps.workspaceRegistry
+			.get(backend)
+			.dispose(spawnRecord.workspaceRef)
+			.catch((disposeErr) => {
+				this.log.warn('Workspace dispose failed', {
+					backend,
+					workspaceId: spawnRecord.workspaceRef?.id,
+					subSessionId: spawnRecord.subSessionId,
+					error: toErrorMessage(disposeErr),
+				})
+			})
 	}
 
 	private markCanceled(taskId: TaskId): void {

package/src/public-runtime.ts

@@ -483,6 +483,12 @@ export {
 
 export {
 	AncestryCycleError,
+	// Exported with the CAS it announces. A host that opts into
+	// `expectedOwnerVersion` has to be able to tell "somebody else took this
+	// session" from any other failure, and string-matching a message is not a
+	// contract — which is the state `ThreadClosedError` and its siblings are
+	// still in, and a reason not to add a fourth.
+	StaleSessionError,
 	TenantIsolationError,
 	WorkspaceBackendError,
 } from './session/errors.js'

package/src/session/__tests__/integration/a-successful-delegation-disposes-its-workspace.test.ts

@@ -0,0 +1,270 @@
+/**
+ * A worktree provisioned for a delegated child outlived the child that used it.
+ *
+ * `finalizeChild` had two dispose sites and both were failure paths — the
+ * non-success branch, and the rollback in `failSubSession`. The success branch
+ * disposed nothing, so `.namzu/worktrees/` grew once per successful delegation:
+ * the more reliable the workers, the faster it filled.
+ *
+ * The backstop could not fire either. `ArchivalManager` resolves a workspace
+ * only when `SubSession.workspaceId` is set, and for a spawn-created
+ * sub-session that field was written `null` and never updated —
+ * `provisionSpawn` kept the ref on the in-memory `ChildSpawnRecord` and nowhere
+ * else. So the one record that could have named the leaked worktree said there
+ * was none.
+ *
+ * Both halves are pinned here, and the failure path is re-asserted alongside
+ * them: a test that only counted disposals would pass on the old code by
+ * reading the failure branch's disposal and calling it the success branch's.
+ */
+
+import { describe, expect, it } from 'vitest'
+import { EMPTY_TOKEN_USAGE } from '../../../constants/limits.js'
+import { AgentManager } from '../../../manager/agent/lifecycle.js'
+import { ThreadManager } from '../../../manager/thread/lifecycle.js'
+import { AgentRegistry } from '../../../registry/agent/definitions.js'
+import { InMemorySessionStore } from '../../../store/session/memory.js'
+import { InMemoryThreadStore } from '../../../store/thread/memory.js'
+import type {
+	AgentCapabilities,
+	AgentInput,
+	BaseAgentConfig,
+	BaseAgentResult,
+} from '../../../types/agent/base.js'
+import type { Agent } from '../../../types/agent/core.js'
+import type { AgentDefinition } from '../../../types/agent/factory.js'
+import type { AgentTaskContext, SendMessageOptions } from '../../../types/agent/task.js'
+import type { RunId, TenantId, UserId, WorkspaceId } from '../../../types/ids/index.js'
+import { createAssistantMessage } from '../../../types/message/index.js'
+import type { ActorRef } from '../../../types/session/actor.js'
+import type { SummaryId } from '../../../types/session/ids.js'
+import type { WorkspaceRef } from '../../../types/workspace/ref.js'
+import { ZERO_COST } from '../../../utils/cost.js'
+import { DefaultCapacityValidator } from '../../handoff/capacity.js'
+import { SessionSummaryMaterializer } from '../../summary/materialize.js'
+import type {
+	BranchWorkspaceParams,
+	CreateWorkspaceParams,
+	WorkspaceBackendDriver,
+	WorkspaceInspection,
+} from '../../workspace/driver.js'
+import { WorkspaceBackendRegistry } from '../../workspace/registry.js'
+
+const tenant = 'tnt_alpha' as TenantId
+
+const capabilities: AgentCapabilities = {
+	supportsTools: false,
+	supportsStreaming: false,
+	supportsConcurrency: false,
+	supportsSubAgents: false,
+}
+
+/** A child that settles the way `outcome` says, so both branches are reachable. */
+function buildAgent(
+	id: string,
+	outcome: 'completed' | 'failed',
+): Agent<BaseAgentConfig, BaseAgentResult> {
+	return {
+		type: 'reactive',
+		metadata: {
+			type: 'reactive',
+			id,
+			name: id,
+			version: '1.0.0',
+			category: 'test',
+			description: id,
+			capabilities,
+		},
+		run: async (_input: AgentInput, _config: BaseAgentConfig): Promise<BaseAgentResult> => ({
+			runId: 'run_child' as RunId,
+			status: outcome,
+			usage: { ...EMPTY_TOKEN_USAGE },
+			cost: { ...ZERO_COST },
+			iterations: 1,
+			durationMs: 1,
+			messages: [createAssistantMessage('child did the work')],
+			result: 'child did the work',
+		}),
+		cancel: async () => undefined,
+		getCapabilities: () => capabilities,
+	}
+}
+
+function buildDefinition(agent: Agent<BaseAgentConfig, BaseAgentResult>): AgentDefinition {
+	return {
+		info: {
+			id: agent.metadata.id,
+			name: agent.metadata.name,
+			version: agent.metadata.version,
+			category: agent.metadata.category,
+			description: agent.metadata.description,
+			tools: [],
+			defaults: { model: 'test', tokenBudget: 1_000 },
+		},
+		typedAgent: agent,
+	}
+}
+
+/** Provisions successfully and records every ref it is asked to dispose. */
+class RecordingWorkspaceDriver implements WorkspaceBackendDriver {
+	readonly kind = 'git-worktree' as const
+	readonly created: WorkspaceRef[] = []
+	readonly disposed: WorkspaceId[] = []
+	private counter = 0
+
+	async create(params: CreateWorkspaceParams): Promise<WorkspaceRef> {
+		const ref: WorkspaceRef = {
+			id: `wsp_test_${++this.counter}` as WorkspaceId,
+			meta: {
+				backend: 'git-worktree',
+				repoRoot: '/tmp/repo',
+				branch: `namzu/${params.label ?? 'unlabelled'}`,
+				worktreePath: `/tmp/repo/.namzu/worktrees/${params.label ?? 'unlabelled'}`,
+			},
+			createdAt: new Date(),
+		}
+		this.created.push(ref)
+		return ref
+	}
+
+	async branch(_source: WorkspaceRef, _params: BranchWorkspaceParams): Promise<WorkspaceRef> {
+		throw new Error('unused in this test')
+	}
+
+	async dispose(ref: WorkspaceRef): Promise<void> {
+		this.disposed.push(ref.id)
+	}
+
+	async inspect(_ref: WorkspaceRef): Promise<WorkspaceInspection> {
+		throw new Error('unused in this test')
+	}
+}
+
+/**
+ * Stands up a Project → Thread → parent Session and an AgentManager wired to a
+ * recording workspace driver. `outcome` decides how the delegated child ends;
+ * `registerBackend: false` leaves the registry empty, which is the supported
+ * lazy-provisioning configuration rather than an error (pattern doc §7.1).
+ */
+async function harness(
+	outcome: 'completed' | 'failed',
+	{ registerBackend = true }: { registerBackend?: boolean } = {},
+) {
+	const store = new InMemorySessionStore()
+	const threadStore = new InMemoryThreadStore()
+	const project = await store.createProject({ tenantId: tenant, name: 'workspace-project' }, tenant)
+	const thread = await threadStore.createThread(
+		{ projectId: project.id, title: 'workspace-topic' },
+		tenant,
+	)
+
+	const userActor: ActorRef = { kind: 'user', userId: 'usr_root' as UserId, tenantId: tenant }
+
+	const parentSession = await store.createSession(
+		{ threadId: thread.id, projectId: project.id, currentActor: userActor },
+		tenant,
+	)
+	await store.updateSession({ ...parentSession, status: 'active' }, tenant)
+
+	let summaryCounter = 0
+	const materializer = new SessionSummaryMaterializer({
+		store,
+		generateSummaryId: () => `sum_test_${++summaryCounter}` as SummaryId,
+	})
+
+	const registry = new AgentRegistry()
+	registry.register(buildDefinition(buildAgent('worker', outcome)))
+
+	const workspaceRegistry = new WorkspaceBackendRegistry()
+	const driver = new RecordingWorkspaceDriver()
+	if (registerBackend) workspaceRegistry.register(driver)
+
+	const manager = new AgentManager(registry, undefined, {
+		sessionStore: store,
+		summaryMaterializer: materializer,
+		workspaceRegistry,
+		capacity: new DefaultCapacityValidator(store),
+		threadManager: new ThreadManager({ threadStore, sessionStore: store }),
+	})
+
+	const taskContext: AgentTaskContext = {
+		parentRunId: 'run_parent' as RunId,
+		parentAgentId: 'supervisor',
+		parentAbortController: new AbortController(),
+		depth: 0,
+		budgetTracker: { total: 100_000, remaining: 100_000 },
+		tenantId: tenant,
+		threadId: thread.id,
+		sessionId: parentSession.id,
+		projectId: project.id,
+		parentActor: userActor,
+	}
+
+	const options: SendMessageOptions = {
+		agentId: 'worker',
+		input: { messages: [], workingDirectory: '/tmp' },
+		parentSessionId: parentSession.id,
+		tenantId: tenant,
+		projectId: project.id,
+		parentActor: userActor,
+		workspaceBackend: 'git-worktree',
+	}
+
+	return { store, manager, driver, parentSession, options, taskContext }
+}
+
+describe('a delegated child does not outlive its workspace', () => {
+	it('disposes the workspace when the child SUCCEEDS', async () => {
+		const { manager, driver, options, taskContext } = await harness('completed')
+
+		const task = await manager.sendMessage(options, taskContext)
+		await manager.waitForCompletion(task.taskId)
+
+		expect(manager.getState(task.taskId)).toBe('completed')
+		expect(driver.created).toHaveLength(1)
+		// The assertion the leak fails: one workspace made, the same one released.
+		expect(driver.disposed).toEqual([driver.created[0]?.id])
+	})
+
+	it('still disposes the workspace when the child FAILS', async () => {
+		const { manager, driver, options, taskContext } = await harness('failed')
+
+		const task = await manager.sendMessage(options, taskContext)
+		await manager.waitForCompletion(task.taskId)
+
+		expect(driver.created).toHaveLength(1)
+		expect(driver.disposed).toEqual([driver.created[0]?.id])
+	})
+
+	it('records the workspace on the sub-session, so archival can find it', async () => {
+		const { store, manager, driver, parentSession, options, taskContext } =
+			await harness('completed')
+
+		const task = await manager.sendMessage(options, taskContext)
+		await manager.waitForCompletion(task.taskId)
+
+		const [subSession] = await store.getChildren(parentSession.id, tenant)
+		expect(subSession).toBeDefined()
+		// Was `null` on every spawn-created sub-session, which is what made
+		// `ArchivalManager`'s `sub.workspaceId &&` guard unreachable here.
+		expect(subSession?.workspaceId).toBe(driver.created[0]?.id)
+	})
+
+	it('leaves workspaceId null when no backend is registered', async () => {
+		// Lazy provisioning stays legal (pattern doc §7.1): an unregistered
+		// backend is not an error, and the record must not claim a workspace
+		// that was never made.
+		const { store, manager, driver, parentSession, options, taskContext } = await harness(
+			'completed',
+			{ registerBackend: false },
+		)
+
+		const task = await manager.sendMessage(options, taskContext)
+		await manager.waitForCompletion(task.taskId)
+
+		expect(manager.getState(task.taskId)).toBe('completed')
+		expect(driver.created).toHaveLength(0)
+		const [subSession] = await store.getChildren(parentSession.id, tenant)
+		expect(subSession?.workspaceId).toBeNull()
+	})
+})

package/src/session/__tests__/integration/handoff-single-e2e.test.ts

@@ -114,13 +114,22 @@ describe('Integration — single-recipient handoff E2E', () => {
 		expect(reloaded?.previousActors).toHaveLength(1)
 		expect(reloaded?.previousActors[0]).toEqual(sourceActor)
 
-		// Wired assertion: updateSession was invoked at least with the lock
-		// transition + the final commit. That sequence is what makes the
-		// handoff "atomic at the store layer".
-		// Expect at minimum locked (v0) followed by committed (v1).
+		// The lock MOVES the version, and this assertion used to say it did not.
+		//
+		// It expected `locked` at v0 — the version the source was already at —
+		// which is precisely why the lock was not a lock: a second handoff
+		// holding the same snapshot saw an unchanged `ownerVersion`, passed the
+		// check, and locked the session again. Both provisioned a worktree and
+		// one erased the other.
+		//
+		// The bump moved to the lock and the commit now keeps it, so a handoff
+		// still consumes exactly one version and `committedOwnerVersion` is
+		// unchanged at 1 — asserted above, and that is the host-visible
+		// contract. What changed is only the intermediate state, which is the
+		// state that had to become visible.
 		expect(updateCalls).toEqual(
 			expect.arrayContaining([
-				{ status: 'locked', ownerVersion: 0 },
+				{ status: 'locked', ownerVersion: 1 },
 				{ status: 'idle', ownerVersion: 1 },
 			]),
 		)