@bubblebrain-ai/bubble 0.0.46 → 0.0.47

package/dist/agent/subagent/runtime.js

@@ -0,0 +1,789 @@
+/**
+ * Subagent thread lifecycle + dynamic-workflow execution.
+ *
+ * Extracted from Agent: spawn/wait/send/close, the scheduler dispatch and its
+ * terminal callbacks, admission, ingestion, worktree-backed child construction
+ * and the QuickJS workflow driver all lived on the Agent class alongside the
+ * conversation loop, sharing its private fields.
+ *
+ * Two rules keep the boundary honest:
+ *  - Everything the runtime needs from the parent arrives through
+ *    `SubagentRuntimeParent` as a LIVE accessor, never a construction-time
+ *    snapshot: `/model` reassigns the parent's provider/model/thinking
+ *    mid-session and children must inherit the current values.
+ *  - Nothing is pushed back into the parent. Tool updates and ingestion
+ *    notices are PULLED (`drainToolUpdates` / `drainIngestionNotices`) at the
+ *    points in Agent.run() that previously called the private flush methods,
+ *    matching ResultIntegrator.drainNotices.
+ *
+ * `new Agent(...)` stays on Agent, reached through `parent.createChild`. That
+ * is the only genuine call back into the parent, and keeping it there is what
+ * guarantees the child-construction argument set — including the fields that
+ * must stay ABSENT (onMessageAppend, sessionID, routing catalog) — has exactly
+ * one definition site.
+ */
+import { randomUUID } from "node:crypto";
+import { buildSystemPrompt } from "../../system-prompt.js";
+import { buildToolPromptOptions } from "../../tools/index.js";
+import { appendOutputSchemaInstructions, buildSchemaCorrectionPrompt, validateStructuredSummary } from "../structured-output.js";
+import { runWorkflow, WorkflowConcurrencyGate } from "../workflow/runtime.js";
+import { assignAgentNickname, discoverAgentProfiles, findAgentProfile, validateAgentProfileTools } from "../profiles.js";
+import { snapshotSubagentThread, subagentResultFromThread } from "../subagent-control.js";
+import { SubagentStore } from "../subagent-store.js";
+import { SubagentScheduler } from "../subagent-scheduler.js";
+import { ChildRunner, classifySubagentAbortReason } from "../child-runner.js";
+import { ResultIntegrator } from "../result-integrator.js";
+import { SubagentAbortError } from "../abort-errors.js";
+import { createSubagentWorktree, finalizeSubagentWorktree } from "../worktree.js";
+import { createWorktreeChildTools, isolateReadonlyChildFileTools } from "../../tools/child-tools.js";
+import { mergeAgentCategories, parseThinkingLevel } from "../categories.js";
+import { composeAbortSignals } from "../budget-ledger.js";
+export class SubagentRuntime {
+    parent;
+    router;
+    categories;
+    subagentsConfig;
+    /** Public so Agent can expose the same private-field access tests rely on. */
+    store;
+    scheduler;
+    childRunner;
+    resultIntegrator = new ResultIntegrator();
+    pendingUpdates = [];
+    updateWakers = new Set();
+    constructor(deps) {
+        this.parent = deps.parent;
+        this.router = deps.router;
+        this.categories = deps.categories;
+        this.subagentsConfig = deps.config;
+        // Eager, with loadPersisted() in the constructor: list_agents must return
+        // persisted children before any spawn happens this process.
+        this.store = new SubagentStore(deps.persistDir);
+        this.store.loadPersisted();
+        this.scheduler = new SubagentScheduler({
+            maxActiveSubagents: this.subagentsConfig.maxActiveSubagents,
+            launchBurst: this.subagentsConfig.launchBurst,
+            launchIntervalMs: this.subagentsConfig.launchIntervalMs,
+            rateLimitMaxAttempts: this.subagentsConfig.rateLimitMaxAttempts,
+            rateLimitBackoffMs: this.subagentsConfig.rateLimitBackoffMs,
+            transportRetryMaxAttempts: this.subagentsConfig.transportRetryMaxAttempts,
+            transportRetryBackoffMs: this.subagentsConfig.transportRetryBackoffMs,
+            getCategoryLimit: (category) => mergeAgentCategories(this.categories)[category]?.maxConcurrent,
+        });
+        this.childRunner = new ChildRunner({
+            allTools: () => this.parent.allTools(),
+            emit: (record, options, status, event, message) => this.emitSubagentLifecycle(record, options, status, event, message),
+            runLifecycleHook: (record, cwd, eventName, status, error, abortSignal) => this.runSubagentLifecycleHookFor(record, cwd, eventName, status, error, abortSignal),
+            finalizeBlocked: (record, error, options) => this.finalizeSubagentBlocked(record, error, options),
+            createInstance: (record, tools, cwd, forkContext) => this.createSubAgentInstance(record, tools, cwd, forkContext),
+            notifyWaiters: (record) => this.store.notifyWaiters(record),
+            onFinal: (record, options) => {
+                this.reclaimWorktree(record);
+                // Workflow-internal agents are not persisted (they never re-import into
+                // the store on restart) and never ingest into parent context (option C).
+                if (!record.workflowInternal) {
+                    this.store.persist(record);
+                    this.maybeEnqueueIngestion(record, options);
+                }
+            },
+        });
+    }
+    // ---- parent-facing view: everything Agent.run() pulls ----
+    /** True when a background child has queued an update the loop should drain. */
+    hasPendingUpdates() {
+        return this.pendingUpdates.length > 0;
+    }
+    /** Follows a session switch: evicts the old session's final children and loads the new one's. */
+    repointPersistDir(persistDir) {
+        this.store.repoint(persistDir);
+    }
+    /**
+     * Inspects and cleans up a child's worktree at ANY terminal outcome:
+     * unchanged → removed; changed → kept for the parent to review, with a
+     * diff stat in the handoff (§8). Called from ChildRunner's onFinal and
+     * from all three scheduler-terminal callbacks — the scheduler paths used
+     * to skip reclamation entirely, leaking the directory and its git
+     * registration whenever a write child exhausted its retries
+     * (known-defects #1).
+     *
+     * When the directory was removed (= the child left no changes, so there
+     * is no file state to lose), the record's worktree reference is cleared:
+     * a kept reference would send a later resume into a deleted directory —
+     * clearing it makes the resume path rebuild a fresh worktree instead
+     * (see ChildRunner's reuse guard).
+     */
+    reclaimWorktree(record) {
+        if (!record.worktree)
+            return;
+        // finalizeSubagentWorktree is idempotent (already-finalized worktrees
+        // return unchanged); only the call that actually performed finalization
+        // may push the handoff note, or re-finalization would duplicate it.
+        const alreadyFinalized = record.worktree.changed !== undefined;
+        finalizeSubagentWorktree(record.worktree);
+        if (!alreadyFinalized && record.worktree.changed) {
+            record.toolNotes.push(`worktree: changes left in ${record.worktree.path} — review the diff before applying`);
+        }
+        if (record.worktree.removed)
+            record.worktree = undefined;
+    }
+    /** Registers a waker so a blocked tool-execution loop learns about updates. */
+    subscribe(wake) {
+        this.updateWakers.add(wake);
+        return () => this.updateWakers.delete(wake);
+    }
+    /**
+     * Completed background children whose summaries should reach parent context
+     * before the next inference turn (design §5). Pulled, not injected.
+     */
+    drainIngestionNotices() {
+        if (!this.resultIntegrator.hasPending())
+            return [];
+        return this.resultIntegrator.drainNotices(this.store);
+    }
+    async runSubAgent(input, cwd, options) {
+        const record = this.createSubagentThreadRecord({
+            profile: options.profile,
+            task: typeof input === "string" ? input : "(multimodal task)",
+            runId: options.runId,
+            agentId: options.subAgentId,
+            parentToolCallId: options.parentToolCallId,
+            parentToolName: "subagent",
+            nickname: options.nickname,
+            route: options.route ?? this.router.resolve(options.profile, options.category),
+        });
+        this.store.set(record);
+        const approval = options.approval ?? options.profile.approval;
+        const admissionError = this.admitSubagentProfile(record, approval);
+        if (admissionError) {
+            this.finalizeSubagentBlocked(record, admissionError, { directEmit: options.emitUpdate });
+            return subagentResultFromThread(record);
+        }
+        record.promise = this.dispatchSubagentRun(record, input, cwd, {
+            approval,
+            abortSignal: options.abortSignal,
+            forkContext: options.forkContext,
+            directEmit: options.emitUpdate,
+        });
+        await record.promise;
+        return subagentResultFromThread(record);
+    }
+    async spawnSubAgent(input, cwd, options) {
+        const route = options.route
+            ?? this.router.resolve(options.profile, options.category, { model: options.model, effort: options.effort });
+        // Early validation (design §7): throws reach the model as a tool error it
+        // can correct this turn, instead of a late provider-factory failure.
+        const routeNote = this.router.validateForDispatch(route);
+        const record = this.createSubagentThreadRecord({
+            profile: options.profile,
+            task: typeof input === "string" ? input : "(multimodal task)",
+            parentToolCallId: options.parentToolCallId,
+            parentToolName: "spawn_agent",
+            route,
+        });
+        if (routeNote)
+            record.toolNotes.push(routeNote);
+        this.store.set(record);
+        const approval = options.approval ?? record.profile.approval;
+        // Admission validation runs before queueing (design §4.2): a request that
+        // would block never consumes a queue slot.
+        const admissionError = this.admitSubagentProfile(record, approval);
+        if (admissionError) {
+            this.finalizeSubagentBlocked(record, admissionError, { queueUpdates: true });
+            return this.snapshotSubagent(record);
+        }
+        this.queueSubagentUpdate(record, "queued", undefined, `Queued ${record.nickname} (${record.profile.name})`);
+        record.promise = this.dispatchSubagentRun(record, input, cwd, {
+            approval,
+            abortSignal: options.abortSignal,
+            forkContext: options.forkContext,
+            queueUpdates: true,
+        });
+        void record.promise.finally(() => this.store.notifyWaiters(record));
+        return this.snapshotSubagent(record);
+    }
+    async waitSubAgents(options = {}) {
+        const targets = this.resolveSubagentTargets(options.agentIds);
+        if (targets.length === 0)
+            return [];
+        const completed = targets.filter((record) => isFinalSubagentStatus(record.status));
+        if (completed.length > 0) {
+            for (const record of completed)
+                this.store.markDelivered(record.agentId);
+            return completed.map((record) => this.snapshotSubagent(record));
+        }
+        const timeoutMs = normalizeWaitTimeout(options.timeoutMs);
+        let waiter;
+        await Promise.race([
+            new Promise((resolve) => {
+                waiter = resolve;
+                for (const record of targets) {
+                    record.waiters.add(resolve);
+                }
+            }).finally(() => {
+                if (waiter) {
+                    for (const record of targets) {
+                        record.waiters.delete(waiter);
+                    }
+                }
+            }),
+            new Promise((resolve) => setTimeout(resolve, timeoutMs)),
+        ]);
+        const finished = targets.filter((record) => isFinalSubagentStatus(record.status));
+        for (const record of finished)
+            this.store.markDelivered(record.agentId);
+        return (finished.length > 0 ? finished : targets).map((record) => this.snapshotSubagent(record));
+    }
+    async sendSubAgentInput(agentId, input, cwd, options = {}) {
+        const record = this.store.get(agentId);
+        if (!record) {
+            throw new Error(`Unknown subagent: ${agentId}`);
+        }
+        if (record.status === "running" || record.status === "queued") {
+            if (!options.interrupt) {
+                throw new Error(`Subagent ${agentId} is still running. Call wait_agent first or pass interrupt:true.`);
+            }
+            record.abortController.abort(new SubagentAbortError(`Subagent ${agentId} interrupted.`, "interrupt"));
+            await record.promise?.catch(() => undefined);
+            record.abortController = new AbortController();
+        }
+        if (record.status === "closed") {
+            throw new Error(`Subagent ${agentId} is closed.`);
+        }
+        record.parentToolCallId = options.parentToolCallId ?? record.parentToolCallId;
+        record.parentToolName = "send_input";
+        record.task = typeof input === "string" ? input : "(multimodal task)";
+        record.summary = "";
+        record.toolNotes = [];
+        record.usage = undefined;
+        record.error = undefined;
+        record.finalReason = undefined;
+        record.deliveredAt = undefined;
+        record.updatedAt = Date.now();
+        if (record.worktree) {
+            // A kept (changed) worktree carries changed/diffStat frozen at the
+            // previous terminal outcome; reset them so the finalize after THIS
+            // run recomputes instead of reporting stale state.
+            record.worktree.changed = undefined;
+            record.worktree.diffStat = undefined;
+        }
+        // A send_input restart is a launch like any other: it goes through the
+        // scheduler's dispatch point and is subject to the same admission limits
+        // (design §4.1) — batch-resuming team members cannot bypass concurrency caps.
+        record.promise = this.dispatchSubagentRun(record, input, cwd, {
+            approval: record.profile.approval,
+            abortSignal: options.abortSignal,
+            queueUpdates: true,
+            reuseAgent: true,
+        });
+        void record.promise.finally(() => this.store.notifyWaiters(record));
+        return this.snapshotSubagent(record);
+    }
+    async closeSubAgent(agentId) {
+        const record = this.store.get(agentId);
+        if (!record) {
+            throw new Error(`Unknown subagent: ${agentId}`);
+        }
+        if (!isFinalSubagentStatus(record.status)) {
+            record.abortController.abort(new SubagentAbortError(`Subagent ${agentId} closed.`, "user_close"));
+            await record.promise?.catch(() => undefined);
+        }
+        record.status = "closed";
+        record.finalReason = record.finalReason ?? "cancelled_user";
+        record.updatedAt = Date.now();
+        this.queueSubagentUpdate(record, "cancelled", undefined, `${record.nickname} closed`);
+        this.store.persist(record);
+        this.store.notifyWaiters(record);
+        return this.snapshotSubagent(record);
+    }
+    /**
+     * Live (non-final) children only — the delegation-nudge gate
+     * (large-task-delegation design §2). listSubAgents() is a grows-only
+     * session history (finished + resumed children stay forever), so it must
+     * never be used to answer "am I currently delegating?".
+     */
+    activeSubAgentCount() {
+        return this.store.activeCount();
+    }
+    listSubAgents() {
+        return this.store.values()
+            .filter((record) => !record.workflowInternal)
+            .map((record) => this.snapshotSubagent(record));
+    }
+    async executeWorkflow(cwd, options) {
+        const profiles = discoverAgentProfiles(cwd, "both").profiles;
+        const runRecords = [];
+        const logs = [];
+        // Per-run isolation (option C review): a concurrency sub-cap below the
+        // global limit so a workflow can't starve interactive subagents.
+        const interactiveReserve = 2;
+        const globalCap = Math.max(1, this.subagentsConfig.maxActiveSubagents ?? 8);
+        const workflowConcurrency = Math.max(1, globalCap - interactiveReserve);
+        const gate = new WorkflowConcurrencyGate(workflowConcurrency);
+        const dispatchAgent = async (spec) => {
+            const baseProfile = findAgentProfile(profiles, spec.opts.agentType ?? "default")
+                ?? findAgentProfile(profiles, "default");
+            if (!baseProfile)
+                return { ok: false, error: "no default subagent profile available" };
+            // Workflow agents are readonly-by-default; mode upgrades come only from the
+            // profile, never from the script (security invariant).
+            const unsupported = baseProfile.mode !== "readonly" && baseProfile.mode !== "write_worktree";
+            if (unsupported)
+                return { ok: false, error: `profile "${baseProfile.name}" mode ${baseProfile.mode} not supported` };
+            // Default-no-network: unattended orchestration of net-capable agents is new
+            // authority in aggregate (option C review), so strip web tools unless the
+            // script opts in with agentType pointing at a profile that includes them.
+            const profile = {
+                ...baseProfile,
+                tools: { ...baseProfile.tools, exclude: [...(baseProfile.tools.exclude ?? []), "web_fetch", "web_search"] },
+            };
+            let route;
+            let routeNote;
+            try {
+                route = this.router.resolve(profile, spec.opts.category, {
+                    model: spec.opts.model,
+                    effort: parseThinkingLevel(spec.opts.effort),
+                });
+                // Dispatch-time validation (§7): resolved routes, never script source text.
+                routeNote = this.router.validateForDispatch(route);
+            }
+            catch (error) {
+                return { ok: false, error: error?.message || String(error) };
+            }
+            const baseTask = spec.opts.schema !== undefined
+                ? appendOutputSchemaInstructions(spec.prompt, spec.opts.schema)
+                : spec.prompt;
+            const record = this.createSubagentThreadRecord({
+                profile,
+                task: baseTask,
+                parentToolCallId: options.parentToolCallId,
+                parentToolName: "run_workflow",
+                route,
+                workflowInternal: true,
+            });
+            record.expectsStructuredOutput = spec.opts.schema !== undefined;
+            if (routeNote)
+                record.toolNotes.push(routeNote);
+            const memberLabel = typeof spec.opts.label === "string" ? spec.opts.label.trim().slice(0, 40) : "";
+            if (memberLabel)
+                record.nickname = memberLabel;
+            runRecords.push(record);
+            this.store.set(record);
+            // Project-local profiles pass the same first-use trust gate as
+            // spawn_agent: a .bubble/agents profile must never gain a side door
+            // into execution just because a script named it (Codex review on #58).
+            // Checked AFTER the record exists so a rejected member still shows up
+            // in the run's counts/snapshots as blocked instead of vanishing.
+            if (options.ensureProfileTrusted) {
+                const blocked = await options.ensureProfileTrusted(baseProfile);
+                if (blocked) {
+                    const message = typeof blocked.content === "string" ? blocked.content : `profile "${baseProfile.name}" requires user approval`;
+                    this.finalizeSubagentBlocked(record, message, { directEmit: options.directEmit, queueUpdates: options.queueUpdates });
+                    return { ok: false, error: message };
+                }
+            }
+            const admissionError = this.admitSubagentProfile(record, profile.approval);
+            if (admissionError) {
+                this.finalizeSubagentBlocked(record, admissionError, { directEmit: options.directEmit, queueUpdates: options.queueUpdates });
+                return { ok: false, error: admissionError };
+            }
+            // Leaf-only concurrency permit (option C review M5): held ONLY around this
+            // agent's dispatch, never across parallel/pipeline composition.
+            await gate.acquire();
+            try {
+                record.promise = this.dispatchSubagentRun(record, baseTask, cwd, {
+                    approval: profile.approval,
+                    abortSignal: options.abortSignal,
+                    directEmit: options.directEmit,
+                    queueUpdates: options.queueUpdates,
+                });
+                await record.promise;
+            }
+            finally {
+                gate.release();
+            }
+            this.store.markDelivered(record.agentId);
+            if (record.status !== "completed") {
+                return { ok: false, error: record.error || `agent ${record.nickname} ended: ${record.finalReason ?? record.status}` };
+            }
+            if (spec.opts.schema === undefined) {
+                return { ok: true, value: record.summary };
+            }
+            // Structured output: validate, one corrective retry, then fall back to raw.
+            // The schema: log lines are the telemetry that decides whether the
+            // deferred terminate-style structured-output tool ever gets reopened.
+            let validated = validateStructuredSummary(record.summary, spec.opts.schema);
+            if (!validated.ok) {
+                logs.push(`schema: ${record.nickname} first output failed validation; sending one corrective retry`);
+                try {
+                    await this.sendSubAgentInput(record.agentId, buildSchemaCorrectionPrompt(spec.opts.schema, record.summary), cwd, { abortSignal: options.abortSignal });
+                    await record.promise?.catch(() => undefined);
+                    validated = validateStructuredSummary(record.summary, spec.opts.schema);
+                }
+                catch {
+                    // resume failed; fall through to raw summary
+                }
+                logs.push(validated.ok
+                    ? `schema: ${record.nickname} corrective retry produced valid output`
+                    : `schema: ${record.nickname} corrective retry still invalid; returning raw summary`);
+            }
+            return { ok: true, value: validated.ok ? validated.value : record.summary };
+        };
+        const result = await runWorkflow({
+            script: options.script,
+            args: options.args,
+            dispatchAgent,
+            onLog: (message) => logs.push(message),
+            onPhase: (title) => logs.push(`— phase: ${title} —`),
+            budget: {
+                // The ledger is pure accounting (no pool limit); scripts see an
+                // unlimited budget unless a future host contract reintroduces one.
+                total: null,
+                spent: () => runRecords.reduce((sum, r) => sum + (r.usage ? r.usage.promptTokens + r.usage.completionTokens : 0), 0),
+                remaining: () => Number.POSITIVE_INFINITY,
+            },
+            signal: options.abortSignal,
+        });
+        return {
+            result,
+            agentCount: runRecords.length,
+            logs,
+            snapshots: runRecords.map((record) => this.snapshotSubagent(record)),
+        };
+    }
+    markSubagentDelivered(agentId) {
+        this.store.markDelivered(agentId);
+    }
+    snapshotSubagent(record) {
+        const snapshot = snapshotSubagentThread(record);
+        if (record.status === "queued") {
+            const queuePosition = this.scheduler.queuePosition(record.agentId);
+            if (queuePosition !== undefined)
+                return { ...snapshot, queuePosition };
+        }
+        return snapshot;
+    }
+    /** Returns the blocking diagnostic message when the profile cannot run, else undefined. */
+    admitSubagentProfile(record, approval) {
+        const diagnostics = validateAgentProfileTools(this.parent.allTools(), record.profile, approval);
+        const blocking = diagnostics.filter((diagnostic) => diagnostic.severity === "error");
+        if (blocking.length === 0)
+            return undefined;
+        return blocking.map((diagnostic) => diagnostic.message).join("\n");
+    }
+    /**
+     * Background children (queueUpdates) get their results ingested before the
+     * parent's next inference turn (design §5); foreground children (team,
+     * legacy task) deliver through their tool result instead.
+     */
+    maybeEnqueueIngestion(record, options) {
+        if (options.queueUpdates) {
+            this.resultIntegrator.enqueue(record.agentId);
+        }
+    }
+    finalizeSubagentBlocked(record, error, emitOptions) {
+        record.status = "blocked";
+        record.finalReason = "blocked";
+        record.error = error;
+        record.updatedAt = Date.now();
+        this.emitSubagentLifecycle(record, emitOptions, "blocked", undefined, error);
+        this.store.persist(record);
+        this.store.notifyWaiters(record);
+    }
+    dispatchSubagentRun(record, input, cwd, options) {
+        record.status = "queued";
+        record.updatedAt = Date.now();
+        const queueSignal = composeAbortSignals([options.abortSignal, record.abortController.signal]);
+        return this.scheduler.dispatch({
+            agentId: record.agentId,
+            category: record.category,
+            signal: queueSignal,
+            run: (ctx) => this.runSubagentThread(record, input, cwd, { ...options, attempt: ctx.attempt }),
+            onCancelledWhileQueued: (reason) => {
+                record.status = "cancelled";
+                record.finalReason = classifySubagentAbortReason(reason, options.abortSignal);
+                record.error = reason instanceof Error ? reason.message : reason ? String(reason) : "Cancelled while queued.";
+                record.updatedAt = Date.now();
+                // This path is NOT only "run never started" — a 429/transport
+                // failure re-queues the entry with the abort listener re-armed, so
+                // an abort during backoff lands here AFTER attempt 1 already ran
+                // (worktree created, SubagentStart fired). Reclaim is therefore a
+                // real leak fix here, and an open Start pair must be closed too
+                // (design §9): only when no Start ever fired is skipping Stop right.
+                if (record.hookStopPending) {
+                    void this.runSubagentLifecycleHookFor(record, cwd, "SubagentStop", record.status, record.error);
+                }
+                this.reclaimWorktree(record);
+                this.emitSubagentLifecycle(record, options, "cancelled", undefined, record.error);
+                this.store.persist(record);
+                this.store.notifyWaiters(record);
+                this.maybeEnqueueIngestion(record, options);
+            },
+            onRateLimitExhausted: (attempts) => {
+                record.status = "failed";
+                record.finalReason = "rate_limited_exhausted";
+                record.error = `Provider rate limit persisted after ${attempts} attempts.`;
+                record.updatedAt = Date.now();
+                // Reclaim before persist so the handoff note lands in the
+                // persisted toolNotes.
+                this.reclaimWorktree(record);
+                void this.runSubagentLifecycleHookFor(record, cwd, "SubagentStop", record.status, record.error);
+                this.emitSubagentLifecycle(record, options, "failed", undefined, record.error);
+                this.store.persist(record);
+                this.store.notifyWaiters(record);
+                this.maybeEnqueueIngestion(record, options);
+            },
+            onTransportRetryExhausted: (attempts) => {
+                record.status = "failed";
+                // failed_transient stays resumable, so the parent can still send_input
+                // to recover the child with its context intact.
+                record.finalReason = "failed_transient";
+                record.error = `Provider transport error persisted after ${attempts} attempts.`;
+                record.updatedAt = Date.now();
+                this.reclaimWorktree(record);
+                void this.runSubagentLifecycleHookFor(record, cwd, "SubagentStop", record.status, record.error);
+                this.emitSubagentLifecycle(record, options, "failed", undefined, record.error);
+                this.store.persist(record);
+                this.store.notifyWaiters(record);
+                this.maybeEnqueueIngestion(record, options);
+            },
+        });
+    }
+    emitSubagentLifecycle(record, options, status, event, message) {
+        const update = this.buildSubagentUpdate(record, status, event, message);
+        options.directEmit?.(update);
+        if (options.queueUpdates) {
+            this.pendingUpdates.push({ id: record.parentToolCallId, name: record.parentToolName, update });
+            this.wakeSubagentUpdateWaiters();
+        }
+    }
+    /** Lets a blocked tool-execution loop drain freshly queued subagent updates. */
+    wakeSubagentUpdateWaiters() {
+        for (const wake of this.updateWakers)
+            wake();
+    }
+    async runSubagentLifecycleHookFor(record, cwd, eventName, status, error, abortSignal) {
+        // Pairing state (design §9), tracked at this single choke point —
+        // every Start/Stop passes through here, so no call site can desync it.
+        // Set synchronously, before the await, to stay ordered with callers.
+        record.hookStopPending = eventName === "SubagentStart";
+        try {
+            await this.parent.runExternalHook({
+                eventName,
+                cwd,
+                runId: record.runId,
+                target: record.profile.name,
+                payload: {
+                    agentId: record.agentId,
+                    nickname: record.nickname,
+                    profile: record.profile.name,
+                    status,
+                    error,
+                },
+            }, abortSignal);
+        }
+        catch {
+            // Subagent lifecycle hooks are observe-only; never fail the subagent.
+            // Observe-only extends to the RESULT, not just errors: the returned
+            // events and modelContext are dropped ON PURPOSE, matching every other
+            // terminal/lifecycle event (Stop, StopFailure, SessionStart/End).
+            // Injecting them would push each child's start/stop hook output into
+            // the parent transcript, once per child in a fan-out. Do not "fix"
+            // this by wiring the result into injectHookModelContext.
+        }
+    }
+    createSubagentThreadRecord(options) {
+        const now = Date.now();
+        const nickname = options.nickname ?? assignAgentNickname(options.profile, this.activeSubagentNicknames());
+        return {
+            agentId: options.agentId ?? randomUUID(),
+            runId: options.runId ?? randomUUID(),
+            nickname,
+            profile: options.profile,
+            category: options.route?.category,
+            route: options.route,
+            workflowInternal: options.workflowInternal,
+            parentToolCallId: options.parentToolCallId,
+            parentToolName: options.parentToolName,
+            status: "queued",
+            task: options.task,
+            summary: "",
+            toolNotes: [],
+            createdAt: now,
+            updatedAt: now,
+            abortController: new AbortController(),
+            waiters: new Set(),
+        };
+    }
+    runSubagentThread(record, input, cwd, options) {
+        return this.childRunner.run(record, input, cwd, options);
+    }
+    async createSubAgentInstance(record, tools, cwd, forkContext) {
+        let childCwd = cwd;
+        let childMode = "plan";
+        if (record.profile.mode === "write_worktree") {
+            // Write children work in a runtime-allocated worktree with fresh tool
+            // instances bound to it (design §8): the parent tree is never touched,
+            // and the tools' own workspace fence enforces containment in code.
+            // NOTE: record.worktree must be assigned before anything below can
+            // throw — ChildRunner's failure path reclaims it off the record.
+            if (!record.worktree) {
+                record.worktree = createSubagentWorktree(cwd, record.agentId);
+            }
+            childCwd = record.worktree.path;
+            childMode = "default";
+            tools = createWorktreeChildTools(childCwd, record.profile.tools.include);
+        }
+        else {
+            // Readonly children share the parent's tool instances; isolate the only
+            // one with mutable file state (read → its FileStateTracker) so concurrent
+            // fan-out members never race shared tool state (design v2 §2).
+            tools = isolateReadonlyChildFileTools(tools);
+        }
+        const childToolNames = tools.map((tool) => tool.name);
+        const route = record.route ?? {
+            providerId: this.parent.providerId,
+            model: this.parent.apiModel,
+            thinkingLevel: this.parent.thinkingLevel,
+            inherited: true,
+        };
+        const provider = await this.resolveProviderForRoute(route);
+        const childSystemPrompt = buildSystemPrompt({
+            agentName: "Bubble",
+            configuredProvider: route.providerId || "none",
+            configuredModel: route.model || "none",
+            configuredModelId: route.providerId && route.model ? `${route.providerId}:${route.model}` : route.model || "none",
+            thinkingLevel: route.thinkingLevel,
+            mode: childMode,
+            workingDir: childCwd,
+            ...buildToolPromptOptions(tools),
+            memoryPrompt: childToolNames.some((name) => name === "memory")
+                ? this.parent.memoryPrompt
+                : undefined,
+            agentProfilePrompt: [
+                `You are subagent ${record.nickname}. Your agent profile is ${record.profile.name}.`,
+                record.profile.mode === "write_worktree"
+                    ? [
+                        "You work inside an isolated git worktree; the parent reviews your diff after you finish.",
+                        "Make your changes, verify them (run tests where possible), and end with a handoff that lists the files you changed and how you verified them.",
+                        "Do not commit, push, or touch anything outside this worktree.",
+                    ].join(" ")
+                    : "",
+                record.profile.prompt,
+            ].filter(Boolean).join("\n\n"),
+        });
+        // Cross-restart resume (design §7): rebuild the child from its persisted
+        // history — including its original system prompt — so send_input continues
+        // with context intact. Consumed here so a later run does not re-seed it.
+        const resumeMessages = record.messages && record.messages.length > 0
+            ? record.messages.map((message) => ({ ...message }))
+            : undefined;
+        if (resumeMessages)
+            record.messages = undefined;
+        return this.parent.createChild({
+            provider,
+            providerId: route.providerId,
+            model: route.model,
+            tools,
+            thinkingLevel: route.thinkingLevel,
+            mode: childMode,
+            maxTurns: record.profile.maxTurns,
+            budgetSource: { runId: record.runId, subAgentId: record.agentId },
+            systemPrompt: childSystemPrompt,
+            subAgentId: record.agentId,
+            resumeMessages,
+            forkContext,
+        });
+    }
+    async resolveProviderForRoute(route) {
+        if (!route.providerId || route.providerId === this.parent.providerId) {
+            return this.parent.provider;
+        }
+        if (!this.parent.providerFactory) {
+            throw new Error([
+                `Subagent route requires provider "${route.providerId}" for model "${route.model}",`,
+                `but the parent agent only has provider "${this.parent.providerId || "none"}" and no provider factory is configured.`,
+            ].join(" "));
+        }
+        return this.parent.providerFactory(route);
+    }
+    buildSubagentUpdate(record, status, event, message) {
+        return {
+            type: "subagent_update",
+            parentToolCallId: record.parentToolCallId,
+            runId: record.runId,
+            subAgentId: record.agentId,
+            agentName: record.profile.name,
+            nickname: record.nickname,
+            category: record.category,
+            route: record.route,
+            status,
+            childEvent: event,
+            summaryDelta: event?.type === "text_delta" ? event.content : undefined,
+            toolName: "name" in (event ?? {}) ? event.name : undefined,
+            toolCallId: "id" in (event ?? {}) ? event.id : undefined,
+            message,
+            metadata: {
+                kind: "subagent",
+                runId: record.runId,
+                subagents: [{
+                        subAgentId: record.agentId,
+                        agentName: record.profile.name,
+                        nickname: record.nickname,
+                        category: record.category,
+                        route: record.route,
+                        status,
+                        profileSource: record.profile.source,
+                        task: record.task,
+                        summary: record.summary,
+                        toolNotes: record.toolNotes,
+                        usage: record.usage,
+                        error: record.error,
+                    }],
+            },
+        };
+    }
+    queueSubagentUpdate(record, status, event, message) {
+        this.pendingUpdates.push({
+            id: record.parentToolCallId,
+            name: record.parentToolName,
+            update: this.buildSubagentUpdate(record, status, event, message),
+        });
+        this.wakeSubagentUpdateWaiters();
+    }
+    /** Queued background-child updates for Agent.run() to emit. */
+    drainToolUpdates() {
+        return this.pendingUpdates.splice(0, this.pendingUpdates.length)
+            .map((pending) => ({
+            type: "tool_update",
+            id: pending.id,
+            name: pending.name,
+            update: pending.update,
+        }));
+    }
+    activeSubagentNicknames() {
+        return this.store.active().map((record) => record.nickname);
+    }
+    resolveSubagentTargets(agentIds) {
+        if (!agentIds || agentIds.length === 0) {
+            return this.store.values().filter((record) => record.status !== "closed" && !record.workflowInternal);
+        }
+        return agentIds.map((id) => {
+            const record = this.store.get(id);
+            if (!record) {
+                throw new Error(`Unknown subagent: ${id}`);
+            }
+            return record;
+        });
+    }
+}
+export function isFinalSubagentStatus(status) {
+    return status === "completed"
+        || status === "failed"
+        || status === "blocked"
+        || status === "cancelled"
+        || status === "closed";
+}
+export function normalizeWaitTimeout(value) {
+    if (typeof value !== "number" || !Number.isFinite(value))
+        return 30_000;
+    return Math.max(100, Math.min(3_600_000, Math.floor(value)));
+}