@esso0428/pi-subagents 0.17.5 → 0.17.7

package/dist/agent-manager.js

@@ -1,108 +1,57 @@
 /**
  * agent-manager.ts — Tracks agents, background execution, resume support.
  *
- * Background agents are subject to a configurable concurrency limit (default: 4).
- * Excess agents are queued and auto-started as running agents complete.
- * Foreground agents bypass the queue (they block the parent anyway).
+ * There are two independent concurrency pools, never one:
+ *
+ * - Background (`maxConcurrent`, default 10) bounds detached agents.
+ * - Foreground (`maxConcurrentForeground`, default 0 = unlimited) bounds
+ *   agents a caller is blocking on inline — `spawnAndWait`.
+ *
+ * Independent by design: a foreground agent blocks the parent anyway, so
+ * charging it to the background pool would let a saturated pool starve the main
+ * session of work it could have done itself. Excess agents in either pool are
+ * queued and auto-started as slots free up. Nested children take no slot in
+ * either — see `occupiesPoolSlot` / `occupiesForegroundSlot`.
  */
 import { randomUUID } from "node:crypto";
 import { statSync } from "node:fs";
 import { isAbsolute } from "node:path";
-import { readAgentHistory } from "./agent-history.js";
-import { readAgentRecoveryCheckpoints, removeAgentRecoveryCheckpoint, writeAgentRecoveryCheckpoint, } from "./agent-recovery.js";
+import { agentHistoryLocator, createAgentHistoryPath, readAgentHistory, streamAgentHistory, writeAgentHistoryInitialEntry, } from "./agent-history.js";
+import { readAgentRecoveryCheckpoints, writeAgentRecoveryCheckpoint, } from "./agent-recovery.js";
 import { resumeAgent, runAgent } from "./agent-runner.js";
+import { assignHandle, handleBase } from "./mention.js";
+import { describeModel } from "./model-resolver.js";
+import { writeInitialEntry } from "./output-file.js";
 import { addUsage } from "./usage.js";
-import { cleanupWorktree, createWorktree, pruneWorktrees, } from "./worktree.js";
-/** Default max concurrent background agents. */
-const DEFAULT_MAX_CONCURRENT = 4;
-const RESTORABLE_STATUSES = new Set(["completed", "steered", "stopped", "aborted", "error"]);
-const THINKING_LEVELS = new Set(["minimal", "low", "medium", "high", "xhigh", "max", "off"]);
-/** Narrow persisted strings before putting them into runtime/UI state. */
-function isSafePersistedString(value, maxLength) {
-    return typeof value === "string" && value.length > 0 && value.length <= maxLength && !/[\0\r\n]/.test(value);
-}
-function isFiniteTimestamp(value) {
-    return typeof value === "number" && Number.isFinite(value) && Number.isInteger(value) && value >= 0;
-}
-function isValidUsage(value) {
-    if (!value || typeof value !== "object")
-        return false;
-    const usage = value;
-    return ["input", "output", "cacheWrite"].every((key) => {
-        const n = usage[key];
-        return typeof n === "number" && Number.isFinite(n) && n >= 0;
-    });
-}
-function isSafeTranscriptLocator(value) {
-    return typeof value === "string"
-        && /^\.pi-subagents\/agent-transcripts\/[^/]+\.jsonl$/.test(value)
-        && !value.includes("..")
-        && !value.includes("\\")
-        && !value.includes("\0");
-}
-function isSafeInvocation(value) {
-    if (!value || typeof value !== "object")
-        return false;
-    const invocation = value;
-    for (const key of ["modelName", "effectiveModelName"]) {
-        if (invocation[key] !== undefined && !isSafePersistedString(invocation[key], 512))
-            return false;
-    }
-    for (const key of ["thinking", "effectiveThinking"]) {
-        if (invocation[key] !== undefined && (typeof invocation[key] !== "string" || !THINKING_LEVELS.has(invocation[key])))
-            return false;
-    }
-    if (invocation.maxTurns !== undefined && (!Number.isInteger(invocation.maxTurns) || invocation.maxTurns < 0))
-        return false;
-    for (const key of ["isolated", "inheritContext", "runInBackground"]) {
-        if (invocation[key] !== undefined && typeof invocation[key] !== "boolean")
-            return false;
-    }
-    if (invocation.isolation !== undefined && invocation.isolation !== "worktree")
-        return false;
-    return true;
-}
-function cloneInvocation(value) {
-    return value ? {
-        modelName: value.modelName,
-        effectiveModelName: value.effectiveModelName,
-        thinking: value.thinking,
-        effectiveThinking: value.effectiveThinking,
-        maxTurns: value.maxTurns,
-        isolated: value.isolated,
-        inheritContext: value.inheritContext,
-        runInBackground: value.runInBackground,
-        isolation: value.isolation,
-    } : undefined;
-}
-/** Validate one persisted terminal record without constructing runtime handles. */
-export function isRestorableAgentRecord(value) {
-    if (!value || typeof value !== "object")
-        return false;
-    const record = value;
-    if (!isSafePersistedString(record.id, 256)
-        || !isSafePersistedString(record.type, 256)
-        || !isSafePersistedString(record.description, 4096)
-        || typeof record.status !== "string"
-        || !RESTORABLE_STATUSES.has(record.status)
-        || !isFiniteTimestamp(record.startedAt)
-        || !isFiniteTimestamp(record.completedAt)
-        || record.completedAt < record.startedAt)
-        return false;
-    if (record.result !== undefined && !isSafePersistedString(record.result, 2_000_000))
-        return false;
-    if (record.error !== undefined && !isSafePersistedString(record.error, 64_000))
-        return false;
-    if (record.toolUses !== undefined && (!Number.isInteger(record.toolUses) || record.toolUses < 0))
-        return false;
-    if (record.lifetimeUsage !== undefined && !isValidUsage(record.lifetimeUsage))
-        return false;
-    if (record.transcriptPath !== undefined && !isSafeTranscriptLocator(record.transcriptPath))
-        return false;
-    if (record.invocation !== undefined && !isSafeInvocation(record.invocation))
-        return false;
-    return true;
-}
+import { cleanupWorktree, createWorktree, isWorktreeIsolationEnabled, pruneWorktrees, } from "./worktree.js";
+/**
+ * Default max concurrent background agents.
+ *
+ * Raised from 4 when top-level spawns started defaulting to background
+ * (`backgroundByDefault`): foreground agents bypass this pool entirely, so
+ * while foreground was the default a fan-out of six ran six. With background
+ * as the default every top-level agent takes a slot, and a limit of 4 would
+ * have silently queued the tail of exactly the parallel fan-outs the `Agent`
+ * tool description tells the model to send.
+ */
+const DEFAULT_MAX_CONCURRENT = 10;
+/**
+ * Default max concurrent foreground (blocking) agents — `0` = unlimited, the
+ * extension's existing convention for "no ceiling" (`defaultMaxTurns`).
+ *
+ * Off by default because nothing here ever bounded foreground work, and pi
+ * dispatches a message's tool calls through `Promise.all`, so an unqualified
+ * fan-out of blocking `Agent` calls has always run all at once. Users who want
+ * it bounded — chiefly local models, where parallel agents thrash the prompt
+ * cache (#253) — opt in; everyone else keeps today's behaviour exactly.
+ */
+const DEFAULT_MAX_CONCURRENT_FOREGROUND = 0;
+/**
+ * How many evicted agents stay addressable by name. Only a bound on memory —
+ * a session that spawns hundreds of agents shouldn't retain every one — and
+ * far above the handful anyone keeps in their head.
+ */
+const MAX_TOMBSTONES = 100;
 /**
  * Validate a caller-supplied SpawnOptions.cwd. `undefined`/`null` mean "unset"
  * (parent cwd). Anything else must be an absolute path to an existing
@@ -126,28 +75,145 @@ function assertValidSpawnCwd(cwd) {
         throw new Error(`SpawnOptions.cwd is not a directory: "${cwd}"`);
     }
 }
+/**
+ * Whether a record occupies one of the `maxConcurrent` background slots.
+ * Nested children don't: their parent already holds a slot, so counting (and
+ * therefore queueing) them would deadlock a parent that waits on its own child.
+ *
+ * Note this bounds nothing horizontally — the depth cap limits how DEEP nesting
+ * goes, not how WIDE. A parent's only limit on concurrent children is that each
+ * spawn costs it a turn, which is unbounded when max turns is unlimited.
+ */
+function occupiesPoolSlot(record) {
+    return !!record.isBackground && isTopLevelAgent(record);
+}
+/**
+ * Whether a record is one of the session's own agents, rather than something
+ * another agent or a workflow owns.
+ *
+ * The single definition behind every user-facing surface — the fleet list, the
+ * widget, the `/agents` menus, `@handle` resolution, and the completion events
+ * and session entries. An owned child reports through its owner, so surfacing
+ * it separately would double-count the same work in the places a person reads.
+ */
+export function isTopLevelAgent(record) {
+    return record.parentAgentId === undefined && record.workflowId === undefined;
+}
+/**
+ * Whether a record occupies one of the `maxConcurrentForeground` slots.
+ *
+ * Keyed on `blocking` — a caller awaiting this record inline — rather than on
+ * `isBackground === false`, because `spawn()` is also the funnel for DETACHED
+ * starts (cross-extension RPC, `@handle` mentions, the registry) that may pass
+ * `isBackground: false` and are documented to run immediately regardless. Those
+ * block nobody, so bounding them buys nothing and would park a record with no
+ * one waiting to release it.
+ *
+ * Nested children are excluded for the same reason as `occupiesPoolSlot`, and
+ * more sharply: their parent is blocked *awaiting them*, so queueing a child
+ * behind its own parent is a guaranteed deadlock rather than a possible one.
+ * Enforced here rather than at the call site so no caller can reintroduce it.
+ *
+ * A workflow's children go out through `spawnAndWait` and so are `blocking`
+ * too, and are excluded on the same `isTopLevelAgent` test as the background
+ * pool: the run already caps how many of its agents run at once, and charging
+ * them here as well would let one fan-out queue behind a limit meant for the
+ * session's own work.
+ *
+ * Like the background pool this bounds width at the top level only — a parent's
+ * own fan-out is limited by nothing but its turn budget.
+ */
+function occupiesForegroundSlot(record) {
+    return !!record.blocking && isTopLevelAgent(record);
+}
+/** Best-effort ceiling on one child's shutdown handlers, so teardown can't strand a quit. */
+const CHILD_SHUTDOWN_TIMEOUT_MS = 3_000;
+/**
+ * Close the extension lifecycle `runAgent` opened with `bindExtensions`, then dispose.
+ *
+ * `AgentSession.dispose()` only calls `ExtensionRunner.invalidate()` — pi emits the event
+ * itself in `AgentSessionRuntime.dispose()` beforehand, and this is the one place that binds
+ * extensions onto a session without going through that path. Without the emit, everything an
+ * extension armed in `session_start` leaks once per spawn, and its next tick throws
+ * `assertActive()` from a bare timer callback — an uncaughtException that kills pi (#242).
+ */
+async function shutdownChildSession(session) {
+    try {
+        const runner = session?.extensionRunner;
+        // Optional all the way down: on a pi without the getter, or a stubbed session from a
+        // partial `onSessionCreated`, skip the emit — the same degrade as before this fix.
+        if (runner?.hasHandlers?.("session_shutdown")) {
+            // Raced, not awaited outright. `emit` runs every handler serially with no timeout of
+            // its own, and dispose() is reached from pi's own `session_shutdown` with the TUI
+            // already torn down — one hung handler would leave a dead terminal.
+            await Promise.race([
+                runner.emit({ type: "session_shutdown", reason: "quit" }),
+                new Promise(resolve => setTimeout(resolve, CHILD_SHUTDOWN_TIMEOUT_MS).unref()),
+            ]);
+        }
+    }
+    catch { /* a partial session must degrade, not take the teardown down with it */ }
+    // Always, even on timeout: disposal is what this function ultimately exists to do.
+    try {
+        session?.dispose?.();
+    }
+    catch { /* ignore */ }
+}
 export class AgentManager {
     agents = new Map();
     cleanupInterval;
     onComplete;
     onStart;
     onCompact;
+    onUsage;
     maxConcurrent;
+    maxConcurrentForeground = DEFAULT_MAX_CONCURRENT_FOREGROUND;
     /** Base repos worktrees were created from — so dispose() can prune them all,
      *  not just the parent repo (caller-supplied cwd can target other repos). */
     worktreeRepos = new Set();
     /** Project cwd for each record's durable checkpoint. */
     recoveryCwds = new Map();
-    /** Queue of background agents waiting to start. */
+    /**
+     * Startup phases, keyed by agent id. `spawn()` still returns synchronously,
+     * but an agent using worktree isolation is not running yet when it does —
+     * copying the repo is an awaited git call. This is what `awaitStartup` hands
+     * callers that must fail their tool call on a startup failure, and what
+     * `waitForAll` waits on while a record is "running" with no `promise` yet.
+     * Entries are dropped once the run is underway, and kept (rejected) after a
+     * startup failure so a late `awaitStartup` still sees it.
+     */
+    startups = new Map();
+    /**
+     * Evicted agents that can still be reached by name, keyed by handle. Outlives
+     * the 10-minute record cleanup — that timer exists to bound memory, not to
+     * expire a conversation the user might still want — and is cleared alongside
+     * completed records on session start/switch.
+     */
+    tombstones = new Map();
+    /**
+     * Agents waiting to start, tagged with the pool they wait on. One queue for
+     * both pools: `drainQueue` picks the earliest entry whose own pool has room,
+     * so neither can head-of-line-block the other, and every removal path
+     * (`abort`, `abortAll`, `dispose`) stays a single filter.
+     *
+     * `release` wakes a caller blocked in `spawnAndWait`, and is fired once the
+     * entry's `start` has SETTLED rather than at drain time: startup is async
+     * now, so releasing earlier would wake the caller before `record.promise`
+     * exists and it would read a still-starting agent as one that never ran.
+     * Removing an entry from this array MUST release it — a queued record has no
+     * promise to await, and pi has no tool-execution timeout to bail the caller
+     * out.
+     */
     queue = [];
     /** Number of currently running background agents. */
     runningBackground = 0;
-    /** Prevent late promise settlement from decrementing a replacement run. */
-    runningBackgroundIds = new Set();
-    constructor(onComplete, maxConcurrent = DEFAULT_MAX_CONCURRENT, onStart, onCompact) {
+    /** Number of currently running foreground (blocking) agents. */
+    runningForeground = 0;
+    constructor(onComplete, maxConcurrent = DEFAULT_MAX_CONCURRENT, onStart, onCompact, onUsage) {
         this.onComplete = onComplete;
         this.onStart = onStart;
         this.onCompact = onCompact;
+        this.onUsage = onUsage;
         this.maxConcurrent = maxConcurrent;
         // Cleanup completed agents after 10 minutes (but keep sessions for resume)
         this.cleanupInterval = setInterval(() => this.cleanup(), 60_000);
@@ -162,51 +228,192 @@ export class AgentManager {
     getMaxConcurrent() {
         return this.maxConcurrent;
     }
-    finishBackground(id) {
-        if (!this.runningBackgroundIds.delete(id))
-            return;
-        this.runningBackground = Math.max(0, this.runningBackground - 1);
+    /** Update the max concurrent foreground (blocking) agents limit. 0 = unlimited. */
+    setMaxConcurrentForeground(n) {
+        // Floor 0, not 1: unlimited is a meaningful value here and the default.
+        this.maxConcurrentForeground = Math.max(0, n);
+        // Start queued agents if the new limit allows — including everything, when
+        // the limit is cleared back to unlimited mid-run.
+        this.drainQueue();
+    }
+    getMaxConcurrentForeground() {
+        return this.maxConcurrentForeground;
+    }
+    /**
+     * Which pool a spawn is charged to, or undefined for one that is charged to
+     * neither (nested children, detached non-background spawns).
+     *
+     * Nothing here queues when the limit is unset — `poolHasRoom` reports an
+     * unlimited pool as always having room, so that alone is what keeps the
+     * default path identical. The `> 0` guard is belt and braces on top: it also
+     * keeps the counter from churning and the settle path from calling a drain
+     * that would find nothing to do. Both are unobservable, which is why no test
+     * pins them; the observable half — that the default start stays synchronous —
+     * is pinned in `test/foreground-concurrency.test.ts`.
+     */
+    poolFor(record) {
+        if (occupiesPoolSlot(record))
+            return "background";
+        if (this.maxConcurrentForeground > 0 && occupiesForegroundSlot(record))
+            return "foreground";
+        return undefined;
+    }
+    poolHasRoom(pool) {
+        return pool === "background"
+            ? this.runningBackground < this.maxConcurrent
+            : this.maxConcurrentForeground === 0 || this.runningForeground < this.maxConcurrentForeground;
+    }
+    /**
+     * Spawn an agent and return its ID immediately (for background use).
+     * If the concurrency limit is reached, the agent is queued.
+     *
+     * The id comes back synchronously, but with `isolation: "worktree"` the agent
+     * is not running yet when it does — the repo copy is an awaited git call.
+     * Callers that must fail a tool call on a startup failure await
+     * `awaitStartup(id)`; everyone else sees it on the record (status "error").
+     */
+    spawn(pi, ctx, type, prompt, options) {
+        // Validate before the queue branch — a queued spawn should fail at the
+        // call, not minutes later at drain. Throw (not warn): programmatic callers
+        // can fix and retry; the RPC layer converts throws into error envelopes.
+        assertValidSpawnCwd(options.cwd);
+        const id = randomUUID().slice(0, 17);
+        const abortController = new AbortController();
+        const record = {
+            id,
+            type,
+            // Owned children — nested, or a workflow's — are filtered out of every
+            // top-level surface, so no handle: nothing can address them and they must
+            // not consume a name a top-level sibling could otherwise take.
+            handle: !isTopLevelAgent(options)
+                ? undefined
+                // A reclaimed handle is used as-is: it belongs to the conversation this
+                // spawn is reopening, and re-deriving it would lose the numbering.
+                : options.reclaim?.handle ?? assignHandle(handleBase(type), this.takenHandles()),
+            description: options.description,
+            // Reclaimed here, or filled in below from `name` — in which case it must
+            // see the handle this record just took, since both come out of the same
+            // namespace.
+            alias: isTopLevelAgent(options) ? options.reclaim?.alias : undefined,
+            // Overwritten below when the spawn is actually queued; a foreground spawn
+            // that queues flips to "queued" there rather than being guessed at here,
+            // since the pool decision needs the finished record.
+            status: options.isBackground ? "queued" : "running",
+            toolUses: 0,
+            startedAt: Date.now(),
+            abortController,
+            lifetimeUsage: { input: 0, output: 0, cacheWrite: 0, cost: 0 },
+            compactionCount: 0,
+            // Raw tri-state (not coerced to a boolean): true = background, false =
+            // foreground (has an inline tool-result surface), undefined = caller never
+            // declared it (e.g. a cross-extension RPC spawn). The widget's background-
+            // only filter excludes only explicit `false`, so undefined agents — which
+            // have no inline surface — stay visible instead of vanishing.
+            isBackground: options.isBackground,
+            // Whether anyone is awaiting this agent is a property of the agent, not
+            // of the call that made it — and both settle paths need it long after
+            // `options` has stopped being the interesting object.
+            blocking: options.blocking,
+            invocation: options.invocation,
+            depth: options.depth ?? 1,
+            parentAgentId: options.parentAgentId,
+            workflowId: options.workflowId,
+            maxSubagentDepth: options.maxSubagentDepth,
+            rootSessionId: options.rootSessionId,
+        };
+        this.agents.set(id, record);
+        this.recoveryCwds.set(id, ctx.cwd);
+        // Durable history is manager-owned so every spawn path (Agent, scheduler,
+        // RPC, mention, and Workflow) has the same recoverable seam. Attach before
+        // any caller callback can start wiring output or observe the id.
+        this.attachDurableTranscript(record, id, prompt, ctx.cwd, options.outputTranscript !== false);
+        // After the insert, so `takenHandles()` already counts this record's own
+        // handle — a spawn named after its own type gets `explore-2`, not a
+        // duplicate `explore` that would make resolution ambiguous.
+        if (record.handle !== undefined && record.alias === undefined && options.name !== undefined) {
+            record.alias = assignHandle(handleBase(options.name), this.takenHandles());
+        }
+        const args = { pi, ctx, type, prompt, options };
+        const pool = this.poolFor(record);
+        if (pool !== undefined && !options.bypassQueue && !this.poolHasRoom(pool)) {
+            // Queue it — started when a running agent in the same pool completes.
+            // Idempotent for background (already "queued"); the flip that matters is
+            // a blocking foreground spawn, optimistically marked "running" above.
+            record.status = "queued";
+            // A queued record never reaches startAgent's signal wiring, so arm the
+            // parent abort here or Esc could not release the position.
+            if (!this.armQueuedAbort(id, options.signal))
+                return id;
+            let release;
+            record.startGate = new Promise(resolve => { release = resolve; });
+            this.queue.push({
+                id,
+                pool,
+                start: () => this.launch(id, record, args, pool),
+                release: () => release(),
+            });
+            options.onQueued?.(id, this.queue.filter(e => e.pool === pool).length - 1);
+            return id;
+        }
+        this.launch(id, record, args, undefined);
+        return id;
+    }
+    /**
+     * Attach the project-local transcript once for a newly-created record.
+     * Repeated calls are harmless: deterministic paths and an existing file keep
+     * the initial user entry intact, which is important for resume and retries.
+     */
+    attachDurableTranscript(record, id, prompt, cwd, outputTranscript) {
+        try {
+            const historyFile = createAgentHistoryPath(cwd, id);
+            writeAgentHistoryInitialEntry(historyFile, id, prompt, cwd);
+            if (outputTranscript)
+                writeInitialEntry(historyFile, id, prompt, cwd);
+            record.historyFile = historyFile;
+            record.transcriptPath = agentHistoryLocator(cwd, historyFile);
+        }
+        catch (err) {
+            // A read-only project must not prevent the agent from running. The
+            // checkpoint still records the spawn metadata and the warning makes the
+            // loss of durable history visible to the host.
+            console.warn(`[pi-subagents] failed to attach durable transcript for ${id}: ${err instanceof Error ? err.message : String(err)}`);
+        }
+        this.checkpoint(record);
     }
     checkpointStatus(record) {
         return record.status;
     }
     makeCheckpoint(record) {
-        const checkpoint = {
+        return {
             version: 1,
             id: record.id,
             type: record.type,
             description: record.description,
             status: this.checkpointStatus(record),
             startedAt: record.startedAt,
-            toolUses: record.toolUses,
-            lifetimeUsage: { ...record.lifetimeUsage },
-            compactionCount: record.compactionCount,
             ...(record.completedAt !== undefined && { completedAt: record.completedAt }),
-            // A durable transcript is the source of truth for partial/full output.
-            // Avoid duplicating potentially sensitive or very large result text.
             ...(!record.transcriptPath && record.result !== undefined && { result: record.result }),
             ...(record.error !== undefined && { error: record.error }),
+            toolUses: record.toolUses,
+            lifetimeUsage: { ...record.lifetimeUsage },
+            compactionCount: record.compactionCount,
             ...(record.transcriptPath !== undefined && { transcriptPath: record.transcriptPath }),
-            ...(record.invocation !== undefined && { invocation: cloneInvocation(record.invocation) }),
+            ...(record.invocation !== undefined && { invocation: { ...record.invocation } }),
         };
-        return checkpoint;
     }
     checkpoint(record) {
         const cwd = this.recoveryCwds.get(record.id);
-        if (!cwd)
-            return;
-        writeAgentRecoveryCheckpoint(cwd, this.makeCheckpoint(record));
+        if (cwd)
+            writeAgentRecoveryCheckpoint(cwd, this.makeCheckpoint(record));
     }
-    flushOutput(record) {
-        if (!record.outputCleanup)
-            return;
-        try {
-            record.outputCleanup();
-        }
-        catch { /* recovery must remain best effort */ }
-        record.outputCleanup = undefined;
+    /** Checkpoint a record after external transcript wiring. */
+    checkpointRecord(id) {
+        const record = this.agents.get(id);
+        if (record)
+            this.checkpoint(record);
     }
-    /** Set the durable transcript locator and checkpoint the current state. */
+    /** Register durable transcript metadata for compatibility with callers that
+     * attach a pre-existing history (for example a restored session). */
     setTranscript(id, historyFile, transcriptPath, cwd) {
         const record = this.agents.get(id);
         if (!record)
@@ -217,111 +424,171 @@ export class AgentManager {
             this.recoveryCwds.set(id, cwd);
         this.checkpoint(record);
     }
-    /** Checkpoint one record explicitly (used after transcript setup). */
-    checkpointRecord(id) {
-        const record = this.agents.get(id);
-        if (record)
-            this.checkpoint(record);
-    }
-    /**
-     * Reload durable records from this project's checkpoint directory. A
-     * running/queued checkpoint means the process was killed before it could
-     * write its stopped state; treat it as stopped and retain its transcript.
-     * SIGKILL cannot run a final flush/checkpoint, so this active snapshot is
-     * necessarily the last recoverable state.
-     */
+    /** Restore active checkpoints as stopped partial history after a restart. */
     restoreRecovered(cwd) {
         for (const checkpoint of readAgentRecoveryCheckpoints(cwd)) {
+            // A session_start can fire again in the same process (resume/switch).
+            // Never replace the live in-memory record with its older checkpoint: the
+            // checkpoint may intentionally omit `result` once durable history exists.
+            if (this.agents.has(checkpoint.id))
+                continue;
             if (!checkpoint.transcriptPath || !readAgentHistory(cwd, checkpoint.transcriptPath))
                 continue;
             const status = checkpoint.status === "running" || checkpoint.status === "queued"
-                ? "stopped"
-                : checkpoint.status;
-            const completedAt = checkpoint.completedAt ?? Date.now();
-            const existing = this.agents.get(checkpoint.id);
-            if (existing) {
-                // Parent-branch records can still carry an unread in-memory result.
-                // Never replace that richer record with the checkpoint's transcript
-                // stub during the same session. Merge only durable locator metadata.
-                if (!existing.transcriptPath && checkpoint.transcriptPath) {
-                    existing.transcriptPath = checkpoint.transcriptPath;
-                }
-                this.recoveryCwds.set(checkpoint.id, cwd);
-                continue;
-            }
-            this.agents.set(checkpoint.id, this.createRestoredRecord({
-                ...checkpoint,
+                ? "stopped" : checkpoint.status;
+            const record = {
+                id: checkpoint.id,
+                type: checkpoint.type,
+                description: checkpoint.description,
                 status,
-                completedAt,
-            }));
-            this.recoveryCwds.set(checkpoint.id, cwd);
+                result: checkpoint.result,
+                error: checkpoint.error,
+                toolUses: checkpoint.toolUses,
+                startedAt: checkpoint.startedAt,
+                completedAt: checkpoint.completedAt ?? Date.now(),
+                transcriptPath: checkpoint.transcriptPath,
+                historyFile: undefined,
+                invocation: checkpoint.invocation,
+                lifetimeUsage: { ...checkpoint.lifetimeUsage },
+                compactionCount: checkpoint.compactionCount,
+            };
+            this.agents.set(record.id, record);
+            this.recoveryCwds.set(record.id, cwd);
         }
     }
     /**
-     * Spawn an agent and return its ID immediately (for background use).
-     * If the concurrency limit is reached, the agent is queued.
+     * Restore terminal records from durable history without creating sessions.
+     * Invalid and live records are ignored so recovery cannot replace active work.
      */
-    spawn(pi, ctx, type, prompt, options) {
-        // Validate before the queue branch — a queued spawn should fail at the
-        // call, not minutes later at drain. Throw (not warn): programmatic callers
-        // can fix and retry; the RPC layer converts throws into error envelopes.
-        assertValidSpawnCwd(options.cwd);
-        const id = randomUUID().slice(0, 17);
-        const abortController = new AbortController();
-        const record = {
-            id,
-            type,
-            description: options.description,
-            status: options.isBackground ? "queued" : "running",
-            toolUses: 0,
-            startedAt: Date.now(),
-            abortController,
-            lifetimeUsage: { input: 0, output: 0, cacheWrite: 0 },
-            compactionCount: 0,
-            // Raw tri-state (not coerced to a boolean): true = background, false =
-            // foreground (has an inline tool-result surface), undefined = caller never
-            // declared it (e.g. a cross-extension RPC spawn). The widget's background-
-            // only filter excludes only explicit `false`, so undefined agents — which
-            // have no inline surface — stay visible instead of vanishing.
-            isBackground: options.isBackground,
-            invocation: options.invocation,
-        };
-        this.agents.set(id, record);
-        this.recoveryCwds.set(id, ctx.cwd);
-        // Give callers a chance to create the durable transcript before the first
-        // checkpoint. This closes the small spawn→attach window in which a queued
-        // or running agent could be left recoverable only as metadata.
-        try {
-            options.onSpawned?.(id);
-            this.checkpoint(record);
-        }
-        catch (err) {
-            this.agents.delete(id);
-            this.recoveryCwds.delete(id);
-            removeAgentRecoveryCheckpoint(ctx.cwd, id);
-            throw err;
-        }
-        const args = { pi, ctx, type, prompt, options };
-        if (options.isBackground && !options.bypassQueue && this.runningBackground >= this.maxConcurrent) {
-            // Queue it — will be started when a running agent completes
-            this.queue.push({ id, args });
-            return id;
+    restoreCompleted(records) {
+        const terminal = new Set(["completed", "steered", "stopped", "aborted", "error"]);
+        const restoredIds = new Set();
+        for (const candidate of records) {
+            const id = typeof candidate.id === "string" && candidate.id.length > 0 ? candidate.id : undefined;
+            const status = candidate.status;
+            if (!id || !status || !terminal.has(status))
+                continue;
+            if (this.agents.has(id) && !restoredIds.has(id))
+                continue;
+            const type = typeof candidate.type === "string" ? candidate.type : undefined;
+            const description = typeof candidate.description === "string" ? candidate.description : undefined;
+            const startedAt = candidate.startedAt;
+            if (!type || description === undefined || typeof startedAt !== "number" || !Number.isFinite(startedAt))
+                continue;
+            if (candidate.completedAt !== undefined && (typeof candidate.completedAt !== "number" || !Number.isFinite(candidate.completedAt)))
+                continue;
+            if (candidate.transcriptPath !== undefined && (typeof candidate.transcriptPath !== "string" || candidate.transcriptPath.includes("..") || candidate.transcriptPath.startsWith("/")))
+                continue;
+            const restored = {
+                ...candidate,
+                id,
+                type,
+                description,
+                status,
+                toolUses: typeof candidate.toolUses === "number" && Number.isFinite(candidate.toolUses) ? candidate.toolUses : 0,
+                startedAt,
+                completedAt: candidate.completedAt ?? Date.now(),
+                lifetimeUsage: candidate.lifetimeUsage ? { ...candidate.lifetimeUsage } : { input: 0, output: 0, cacheWrite: 0 },
+                compactionCount: typeof candidate.compactionCount === "number" && Number.isFinite(candidate.compactionCount) ? candidate.compactionCount : 0,
+                session: undefined,
+                abortController: undefined,
+                promise: undefined,
+                startGate: undefined,
+                outputCleanup: undefined,
+                historyCleanup: undefined,
+            };
+            this.agents.set(id, restored);
+            restoredIds.add(id);
         }
-        // startAgent can throw (e.g. strict worktree-isolation failure) — clean
-        // up the record so callers don't see an orphan in `listAgents()`.
-        try {
-            this.startAgent(id, record, args);
+    }
+    /**
+     * Wire a parent abort signal for a record that is about to be QUEUED.
+     * `startAgent` does this for running agents, and a queued record never gets
+     * there, so without this Esc could not release a queue position.
+     *
+     * Returns false when the signal is ALREADY aborted, in which case the record
+     * is stopped here and must not be enqueued: `addEventListener` never fires on
+     * an aborted signal, so a `spawnAndWait` on it would wait forever — pi has no
+     * tool-execution timeout to bail it out.
+     *
+     * The listener is left in place when the agent starts. `startAgent` adds its
+     * own, so both fire on a later abort, but `abort()` on an already-stopped
+     * record is a no-op — so detaching would only be tidiness, and tidiness the
+     * `abortAll`/`dispose` paths could not offer anyway.
+     */
+    armQueuedAbort(id, signal) {
+        if (signal === undefined)
+            return true;
+        if (signal.aborted) {
+            const record = this.agents.get(id);
+            if (record) {
+                record.status = "stopped";
+                record.completedAt = Date.now();
+                this.flushOutput(record);
+                this.checkpoint(record);
+            }
+            return false;
         }
-        catch (err) {
-            this.agents.delete(id);
-            this.recoveryCwds.delete(id);
-            removeAgentRecoveryCheckpoint(ctx.cwd, id);
+        signal.addEventListener("abort", () => this.abort(id), { once: true });
+        return true;
+    }
+    /**
+     * Kick off an agent's startup and register it under `startups`. The returned
+     * promise never rejects — the failure is delivered through `awaitStartup`,
+     * and to the record.
+     *
+     * @param queuedPool - The pool this start was QUEUED on, or undefined for an
+     *   immediate start. A queue drain can be minutes after `spawn()` returned,
+     *   and nobody is awaiting `awaitStartup` by then, so a failure has to live
+     *   on the record as status "error" — what drainQueue did when the throw was
+     *   still synchronous. An immediate start instead drops the record, exactly
+     *   as the throw out of `spawn()` did: no orphan in `listAgents()`, and the
+     *   handle goes back.
+     */
+    launch(id, record, args, queuedPool) {
+        const startup = this.startAgent(id, record, args).then(() => { this.startups.delete(id); }, (err) => {
+            this.startups.delete(id);
+            if (queuedPool !== undefined) {
+                // Mirrors settleRun: an inline caller gets this failure as a throw
+                // out of spawnAndWait, so an unconsumed record would ALSO nudge the
+                // session about it — the same failure reported twice.
+                if (queuedPool === "foreground")
+                    record.resultConsumed = true;
+                record.status = "error";
+                record.error = err instanceof Error ? err.message : String(err);
+                record.completedAt = Date.now();
+                this.flushOutput(record);
+                this.checkpoint(record);
+                this.onComplete?.(record);
+            }
+            else {
+                this.agents.delete(id);
+            }
+            // The agent never kept its slot (startAgent gives it back on failure),
+            // so anything queued behind it can go now.
+            this.drainQueue();
             throw err;
-        }
-        return id;
+        });
+        this.startups.set(id, startup);
+        // Nothing is obliged to await `startups` — swallow the rejection once here
+        // so an unawaited startup can't take the process down, and hand callers
+        // (drainQueue) that swallowed promise.
+        return startup.catch(() => { });
+    }
+    /**
+     * Resolves once the agent is actually running, and rejects with the startup
+     * failure (strict worktree isolation) that `spawn()` used to throw before the
+     * repo copy became async. Resolves immediately for an agent that is already
+     * running, still queued, or unknown — so callers can await it unconditionally.
+     *
+     * Call it in the same tick as the `spawn()` it belongs to: a failed startup
+     * takes its record (and this entry) with it, exactly as the throw did.
+     */
+    awaitStartup(id) {
+        return this.startups.get(id) ?? Promise.resolve();
     }
     /** Actually start an agent (called immediately or from queue drain). */
-    startAgent(id, record, { pi, ctx, type, prompt, options }) {
+    async startAgent(id, record, { pi, ctx, type, prompt, options }) {
         // Re-validate a caller-supplied cwd: queued spawns can start minutes after
         // spawn()'s check, and the directory may be gone by then (TOCTOU). Same
         // curated errors; drainQueue parks a throw on the record as an error.
@@ -330,13 +597,46 @@ export class AgentManager {
         // repo and both cleanup calls below MUST agree on this value forever.
         const customCwd = options.cwd ?? undefined; // null (RPC "unset") → undefined
         const baseCwd = customCwd ?? ctx.cwd;
+        // Take the running state — and with it the concurrency slot — BEFORE the
+        // first await. Creating a worktree is an awaited git call, and drainQueue
+        // reads the pool counters synchronously in a loop: incrementing after the
+        // await would let it start every queued agent at once while the first is
+        // still copying its repo. Claiming "running" here also keeps abort() and
+        // abortAll() able to reach an agent whose worktree is still being created.
+        //
+        // The pool is resolved ONCE, here, and carried to `settleRun` below:
+        // `poolFor` reads `maxConcurrentForeground`, which the user can change from
+        // `/agents → Settings` mid-run, so recomputing it at settle time would
+        // decrement a pool this run never charged (counter underflow, limit
+        // silently lifted) or skip the decrement for one it did (leaked slot —
+        // every later blocking spawn queues forever). The two startup exits below
+        // never reach `settleRun`, so they hand the slot back themselves.
+        const pool = this.poolFor(record);
+        const releaseSlot = () => {
+            if (pool === "background")
+                this.runningBackground--;
+            else if (pool === "foreground")
+                this.runningForeground--;
+        };
+        record.status = "running";
+        record.startedAt = Date.now();
+        record.startGate = undefined;
+        if (pool === "background")
+            this.runningBackground++;
+        else if (pool === "foreground")
+            this.runningForeground++;
+        this.checkpoint(record);
         // Worktree isolation: try to create a temporary git worktree. Strict —
-        // fail loud if not possible (no silent fallback to main tree). Done
-        // BEFORE state mutation so a throw doesn't leave the record half-running.
+        // fail loud if not possible (no silent fallback to main tree). Done BEFORE
+        // the run is kicked off so a failure doesn't leave a half-running agent.
+        // The project switch is enforced here as well as at the tool boundary
+        // because cross-extension RPC forwards its options unvalidated — a schema
+        // that omits the field can't stop a caller that never saw the schema.
         let worktreeCwd;
-        if (options.isolation === "worktree") {
-            const wt = createWorktree(baseCwd, id);
+        if (options.isolation === "worktree" && isWorktreeIsolationEnabled()) {
+            const wt = await createWorktree(pi, baseCwd, id);
             if (!wt) {
+                releaseSlot();
                 throw new Error('Cannot run with isolation: "worktree" — not a git repo, no commits yet, or `git worktree add` failed. ' +
                     'Initialize git and commit at least once, or omit `isolation`.');
             }
@@ -349,21 +649,32 @@ export class AgentManager {
             // subdirectory, silently dropping extensions/skills.
             worktreeCwd = customCwd !== undefined ? wt.workPath : wt.path;
             this.worktreeRepos.add(baseCwd);
-        }
-        record.status = "running";
-        record.startedAt = Date.now();
-        this.checkpoint(record);
-        if (options.isBackground) {
-            this.runningBackground++;
-            this.runningBackgroundIds.add(id);
+            // No longer "running" means a stop landed while the copy was being made
+            // (abort(), abortAll()) — a window that did not exist when creation was
+            // synchronous. The record is already terminal, so launching the run would
+            // burn tokens on work nobody is waiting for: discard the fresh (and by
+            // definition unchanged) worktree instead.
+            if (record.status !== "running") {
+                releaseSlot();
+                record.worktreeResult = await cleanupWorktree(pi, baseCwd, wt, options.description);
+                this.drainQueue();
+                return;
+            }
         }
         this.onStart?.(record);
         // Wire parent abort signal to stop the subagent when the parent is interrupted
         let detachParentSignal;
         if (options.signal) {
-            const onParentAbort = () => this.abort(id);
-            options.signal.addEventListener("abort", onParentAbort, { once: true });
-            detachParentSignal = () => options.signal.removeEventListener("abort", onParentAbort);
+            // A queued spawn can start minutes after the caller handed us its signal,
+            // by which time it may already be aborted — and `addEventListener` would
+            // never fire, leaving a child the parent can no longer reach.
+            if (options.signal.aborted)
+                this.abort(id);
+            else {
+                const onParentAbort = () => this.abort(id);
+                options.signal.addEventListener("abort", onParentAbort, { once: true });
+                detachParentSignal = () => options.signal.removeEventListener("abort", onParentAbort);
+            }
         }
         const detach = () => { detachParentSignal?.(); detachParentSignal = undefined; };
         const promise = runAgent(ctx, type, prompt, {
@@ -374,13 +685,20 @@ export class AgentManager {
             isolated: options.isolated,
             inheritContext: options.inheritContext,
             thinkingLevel: options.thinkingLevel,
+            structuredOutput: options.structuredOutput,
+            resumeSessionFile: options.resumeSessionFile,
+            nested: options.parentAgentId !== undefined,
+            workflow: options.workflowId !== undefined,
             // Worktree wins for the working dir (the agent must run in the copy —
             // which, with a custom cwd, was created from that target). Config stays
             // with the parent project when a caller-supplied cwd is in play; it must
             // stay undefined otherwise so plain worktree runs keep resolving config
             // (incl. relative extension paths and memory) inside the worktree copy.
             cwd: worktreeCwd ?? customCwd,
-            configCwd: customCwd !== undefined ? ctx.cwd : undefined,
+            // Set iff a worktree was created (see above) — names the directory the
+            // copy came from, so the prompt can tell the agent not to work there.
+            worktreeBase: worktreeCwd ? baseCwd : undefined,
+            configCwd: options.configCwd ?? (customCwd !== undefined ? ctx.cwd : undefined),
             signal: record.abortController.signal,
             onToolActivity: (activity) => {
                 if (activity.type === "end")
@@ -391,6 +709,7 @@ export class AgentManager {
             onTextDelta: options.onTextDelta,
             onAssistantUsage: (usage) => {
                 addUsage(record.lifetimeUsage, usage);
+                this.onUsage?.(record, usage);
                 options.onAssistantUsage?.(usage);
             },
             onCompaction: (info) => {
@@ -398,14 +717,48 @@ export class AgentManager {
                 this.onCompact?.(record, info);
                 options.onCompaction?.(info);
             },
+            nestedRuntime: {
+                manager: this,
+                parentAgentId: id,
+                depth: record.depth ?? 1,
+                maxSubagentDepth: record.maxSubagentDepth,
+            },
             onSessionCreated: (session) => {
                 record.session = session;
-                const model = session.model;
-                record.invocation = {
-                    ...(record.invocation ?? {}),
-                    ...(model && { effectiveModelName: model.name ?? model.id }),
-                    effectiveThinking: session.thinkingLevel,
-                };
+                // Capture now, while the session object exists: after eviction this
+                // path is the only thing that can reopen the conversation, and an
+                // in-memory session reports undefined, which correctly means
+                // "nothing to come back to".
+                // Optional chaining, not defensiveness for its own sake: this is the
+                // only field read off the session at creation, so an older pi or a
+                // stubbed session must degrade to "not resumable" rather than throw
+                // and take the whole spawn down with it.
+                record.sessionFile = session.sessionManager?.getSessionFile?.();
+                // Same reason, different field: the model and thinking level are only
+                // knowable once pi has resolved its defaults and clamped the level to
+                // what the model supports. Writing them back here makes the record
+                // authoritative, so every surface reads one place instead of each
+                // re-deriving "session, else the request" for itself.
+                if (session.model) {
+                    record.invocation ??= {};
+                    // Read the kept request first: a caller's level survives being clamped
+                    // AND, one line later, being replaced by the effective one.
+                    const requested = record.invocation.requestedThinking ?? record.invocation.thinking;
+                    Object.assign(record.invocation, describeModel(session.model));
+                    // Guarded for the reason above: a session that reports no level keeps
+                    // the request rather than losing it. Overwriting unconditionally would
+                    // turn an older or stubbed session into a blank `thinking:` tag, which
+                    // is worse than the stale-but-true value it replaced.
+                    if (session.thinkingLevel) {
+                        record.invocation.thinking = session.thinkingLevel;
+                        if (requested && requested !== session.thinkingLevel) {
+                            record.invocation.requestedThinking = requested;
+                        }
+                    }
+                }
+                if (record.historyFile) {
+                    record.historyCleanup = streamAgentHistory(session, record.historyFile, record.id, ctx.cwd);
+                }
                 // Flush any steers that arrived before the session was ready
                 if (record.pendingSteers?.length) {
                     for (const msg of record.pendingSteers) {
@@ -413,15 +766,11 @@ export class AgentManager {
                     }
                     record.pendingSteers = undefined;
                 }
-                options.onSessionCreated?.(session);
                 this.checkpoint(record);
+                options.onSessionCreated?.(session);
             },
         })
-            .then(({ responseText, session, aborted, steered, failure }) => {
-            // A disposed manager no longer owns this run. Avoid late callbacks
-            // mutating a dead session or emitting completion side effects.
-            if (this.agents.get(id) !== record)
-                return responseText;
+            .then(async ({ responseText, session, aborted, steered, failure, structuredJson, structuredRetried }) => {
             // Don't overwrite status if externally stopped via abort()
             if (record.status !== "stopped") {
                 // Precedence: a hard abort keeps "aborted"; then a failed final turn
@@ -439,54 +788,46 @@ export class AgentManager {
                 }
             }
             record.result = responseText;
+            // Kept beside `result`, never inside it: `result` is prose meant for a
+            // reader — it is previewed, transcribed, and appended to below — while
+            // this is a machine-readable payload one caller asked for by schema.
+            record.structuredJson = structuredJson;
+            record.structuredRetried = structuredRetried;
             record.session = session;
             record.completedAt ??= Date.now();
             detach();
-            // Final flush of streaming output file
-            if (record.outputCleanup) {
-                try {
-                    record.outputCleanup();
-                }
-                catch { /* ignore */ }
-                record.outputCleanup = undefined;
-            }
+            // Flush both optional output and durable history before terminal state
+            // is checkpointed or completion is observable.
+            this.flushOutput(record);
             // Clean up worktree if used
             if (record.worktree) {
-                const wtResult = cleanupWorktree(baseCwd, record.worktree, options.description);
+                // The one moment the child's tree still exists and the child is done
+                // writing to it. try/catch, not decoration: a hook that throws must
+                // not leave the worktree behind.
+                if (options.onBeforeWorktreeCleanup) {
+                    try {
+                        await options.onBeforeWorktreeCleanup(record.worktree.path);
+                    }
+                    catch { /* ignore — never block cleanup */ }
+                }
+                const wtResult = await cleanupWorktree(pi, baseCwd, record.worktree, options.description);
                 record.worktreeResult = wtResult;
                 if (wtResult.hasChanges && wtResult.branch) {
                     // With a caller-supplied cwd the branch lives in THAT repo, not the
                     // parent session's — say so, or the orchestrator merges in the wrong repo.
                     const repoNote = customCwd !== undefined ? ` in \`${baseCwd}\`` : "";
+                    // Appended to the prose only. A structured child's caller parses
+                    // `structuredJson`, which stays untouched — but `result` is also
+                    // what a human reads, so the note still belongs on it.
                     record.result = (record.result ?? "") +
                         `\n\n---\nChanges saved to branch \`${wtResult.branch}\`${repoNote}. Merge with: \`git merge ${wtResult.branch}\`${customCwd !== undefined ? ` (run in \`${baseCwd}\`)` : ""}`;
                 }
             }
-            this.checkpoint(record);
-            // Fire onComplete for foreground agents too — lifecycle symmetry.
-            // Mark resultConsumed so the callback skips notifications (result returned inline).
-            if (!options.isBackground) {
-                record.resultConsumed = true;
-                try {
-                    this.onComplete?.(record);
-                }
-                catch { /* ignore completion side-effect errors */ }
-            }
-            else {
-                this.finishBackground(id);
-                try {
-                    this.onComplete?.(record);
-                }
-                catch { /* ignore completion side-effect errors */ }
-                this.drainQueue();
-            }
+            this.abortOwnedChildren(id);
+            this.settleRun(record, true, pool);
             return responseText;
         })
-            .catch((err) => {
-            // A disposed manager no longer owns this run. Avoid late callbacks
-            // mutating a dead session or emitting completion side effects.
-            if (this.agents.get(id) !== record)
-                return "";
+            .catch(async (err) => {
             // Don't overwrite status if externally stopped via abort()
             if (record.status !== "stopped") {
                 record.status = "error";
@@ -494,108 +835,284 @@ export class AgentManager {
             record.error = err instanceof Error ? err.message : String(err);
             record.completedAt ??= Date.now();
             detach();
-            // Final flush of streaming output file on error
-            if (record.outputCleanup) {
-                try {
-                    record.outputCleanup();
-                }
-                catch { /* ignore */ }
-                record.outputCleanup = undefined;
-            }
+            // Preserve partial assistant/tool history before recording the error.
+            this.flushOutput(record);
             // Best-effort worktree cleanup on error
             if (record.worktree) {
                 try {
-                    const wtResult = cleanupWorktree(baseCwd, record.worktree, options.description);
+                    const wtResult = await cleanupWorktree(pi, baseCwd, record.worktree, options.description);
                     record.worktreeResult = wtResult;
                 }
                 catch { /* ignore cleanup errors */ }
             }
-            this.checkpoint(record);
-            // Fire onComplete for foreground agents too — lifecycle symmetry.
-            // Mark resultConsumed so the callback skips notifications (result returned inline).
-            if (!options.isBackground) {
-                record.resultConsumed = true;
-                this.onComplete?.(record);
-            }
-            else {
-                this.finishBackground(id);
-                this.onComplete?.(record);
-                this.drainQueue();
-            }
+            this.abortOwnedChildren(id);
+            this.settleRun(record, false, pool);
             return "";
         });
         record.promise = promise;
+        // Notify caller that spawn is complete (record is in the map, promise is set).
+        // Called synchronously — onSessionCreated fires asynchronously inside runAgent.
+        // Used by spawnAndWait to let the caller set up output files before streaming
+        // starts. Read off the options, so a spawn that started from a queue drain
+        // still reaches the caller that queued it.
+        options.onSpawned?.(id);
     }
-    /** Start queued agents up to the concurrency limit. */
+    /**
+     * The shared tail of both settle paths: release whatever pool slot the run
+     * held, notify, and let the queue drain into the freed slot.
+     *
+     * The decrement lives HERE and nowhere else. `abort()` on a running record
+     * only fires its controller and leaves the run to settle normally, so
+     * decrementing there too would double-free — permanently lifting the limit.
+     *
+     * Foreground agents fire `onComplete` for lifecycle symmetry, with
+     * `resultConsumed` set so the callback skips notifications the inline result
+     * already delivered.
+     *
+     * @param guardCallback swallow a throwing `onComplete` (the success path does;
+     *   the error path historically did not, and keeps not doing so).
+     * @param pool the pool this run was CHARGED TO at start time — passed in, not
+     *   recomputed, so a mid-run change to `maxConcurrentForeground` can't make
+     *   the release disagree with the acquire.
+     */
+    settleRun(record, guardCallback, pool) {
+        // Terminal state is not durable until the stream has flushed. The
+        // checkpoint deliberately follows this call so stop/error/partial runs can
+        // be reopened after the live session is released.
+        this.flushOutput(record);
+        this.checkpoint(record);
+        if (!record.isBackground)
+            record.resultConsumed = true;
+        if (pool === "background")
+            this.runningBackground--;
+        else if (pool === "foreground")
+            this.runningForeground--;
+        if (guardCallback) {
+            try {
+                this.onComplete?.(record);
+            }
+            catch { /* ignore completion side-effect errors */ }
+        }
+        else {
+            this.onComplete?.(record);
+        }
+        // The isBackground half reproduces the pre-pool condition exactly — a
+        // background settle has always drained, even for a nested child that held
+        // no slot — so that path is unchanged whether or not the foreground pool is
+        // on. The `pool` half only adds the drain a freed FOREGROUND slot needs.
+        // A drain with nothing freed is a no-op anyway, but "no-op" is a claim
+        // about reachability, and matching the old condition needs no such claim.
+        if (record.isBackground || pool !== undefined)
+            this.drainQueue();
+    }
+    flushOutput(record) {
+        if (record.outputCleanup) {
+            try {
+                record.outputCleanup();
+            }
+            catch { /* best effort */ }
+            record.outputCleanup = undefined;
+        }
+        if (record.historyCleanup) {
+            try {
+                record.historyCleanup();
+            }
+            catch { /* best effort */ }
+            record.historyCleanup = undefined;
+        }
+    }
+    /**
+     * Stop the nested children a settled parent owns. Nested records are hidden
+     * from the UI and only their owner can consume them, so a child outliving its
+     * parent would burn tokens unseen with no way to reach it. Grandchildren are
+     * covered transitively — each abort lands in that child's own settle path.
+     */
+    abortOwnedChildren(parentId) {
+        for (const [id, record] of this.agents) {
+            if (record.parentAgentId === parentId)
+                this.abort(id);
+        }
+    }
+    /**
+     * Start queued agents up to each pool's concurrency limit.
+     *
+     * `findIndex` on the entry's OWN pool rather than `shift`: with one queue
+     * serving two independent limits, a saturated foreground pool at the head
+     * would otherwise stall every background agent behind it. Taking the earliest
+     * eligible entry keeps FIFO within each pool, which is what callers see.
+     */
     drainQueue() {
-        while (this.queue.length > 0 && this.runningBackground < this.maxConcurrent) {
-            const next = this.queue.shift();
+        for (;;) {
+            const i = this.queue.findIndex(e => this.poolHasRoom(e.pool));
+            if (i === -1)
+                return;
+            const [next] = this.queue.splice(i, 1);
             const record = this.agents.get(next.id);
-            if (!record || record.status !== "queued")
+            // Stale entries (aborted while queued) are not started — but are still
+            // released, since nothing else will.
+            if (!record || record.status !== "queued") {
+                next.release();
                 continue;
-            try {
-                this.startAgent(next.id, record, next.args);
-            }
-            catch (err) {
-                // Late failure (e.g. strict worktree-isolation) — surface on the record
-                // so the user/agent can see it via /agents, then keep draining.
-                record.status = "error";
-                record.error = err instanceof Error ? err.message : String(err);
-                record.completedAt = Date.now();
-                this.checkpoint(record);
-                this.onComplete?.(record);
             }
+            // Detached, and never rejects: a late failure (e.g. strict worktree
+            // isolation) lands on the record inside `launch`, exactly as the
+            // synchronous throw did here before, and draining continues either way.
+            //
+            // The release waits for that startup to SETTLE rather than firing here.
+            // Startup is async now, so a release at drain time would wake a blocked
+            // `spawnAndWait` while `record.promise` was still undefined, and it would
+            // read a perfectly healthy agent as one that never ran.
+            void next.start().then(() => next.release(), () => next.release());
+        }
+    }
+    /**
+     * Remove queued entries and wake anyone blocked on them. The single point
+     * that enforces "leaving the queue releases the waiter" — a missed release is
+     * an unbounded hang, not a failed call.
+     */
+    dequeue(pred) {
+        const kept = [];
+        for (const entry of this.queue) {
+            if (pred(entry))
+                entry.release();
+            else
+                kept.push(entry);
         }
+        this.queue = kept;
     }
     /**
      * Spawn an agent and wait for completion (foreground use).
-     * Foreground agents bypass the concurrency queue.
+     * Charged to the foreground pool (`maxConcurrentForeground`), which is
+     * unlimited by default; never to the background one.
      * Returns { id, record } so callers can access the agent ID.
      *
-     * @param onSpawned - Called synchronously after spawn(), before onSessionCreated fires.
-     *   Use this to set record.outputFile so streamToOutputFile can pick it up.
+     * @param onSpawned - Called synchronously once the run is kicked off, before
+     *   onSessionCreated fires. Use this to set record.outputFile so
+     *   streamToOutputFile can pick it up.
      */
     async spawnAndWait(pi, ctx, type, prompt, options, onSpawned) {
+        // `blocking` is what maxConcurrentForeground bounds, and this is its only
+        // source. onSpawned rides on the options rather than on a field of this
+        // manager: a queued spawn starts at drain time, long after any install/
+        // restore pair around this call would have put the field back — and it now
+        // fires after an await (worktree creation) even on the immediate path.
         const id = this.spawn(pi, ctx, type, prompt, {
             ...options,
             isBackground: false,
+            blocking: true,
             onSpawned,
         });
         const record = this.agents.get(id);
-        await record.promise;
+        // Queued: nothing to await yet — the promise appears when the drain starts
+        // it. The gate resolves (never rejects) on every path out of the queue,
+        // start and abort alike, so a rejection can never escape into the caller's
+        // tool `execute` and take down pi's whole Promise.all tool batch.
+        if (record.status === "queued")
+            await record.startGate;
+        // The run promise only exists once startup is past its awaited repo copy —
+        // without this the call would return before the agent had started at all.
+        // A startup failure (strict worktree isolation) rejects here, which is what
+        // the immediate path owes its caller: pi only marks a tool result failed
+        // when `execute` throws. A queued spawn's failure landed on the record
+        // instead (nobody was awaiting `startups` at drain time) and is rethrown
+        // below, so the contract is the same either way.
+        await this.awaitStartup(id);
+        // undefined when it was aborted while queued, or stopped mid-copy, and so
+        // never ran — the record is already terminal with a completedAt, which is
+        // what the caller renders.
+        if (record.promise)
+            await record.promise;
+        // A record that ended "error" without ever getting a promise never ran: the
+        // same startup failure spawn() rethrows on the immediate path (#179). Keep
+        // one contract rather than letting queue pressure decide whether a strict
+        // worktree failure throws or returns as a result.
+        if (record.promise === undefined && record.status === "error") {
+            throw new Error(record.error ?? "Agent failed to start");
+        }
         return { id, record };
     }
     /**
      * Resume an existing agent session with a new prompt.
      */
-    async resume(id, prompt, signal) {
+    async resume(id, prompt, signal, options) {
         const record = this.agents.get(id);
         if (!record?.session)
             return undefined;
+        // Background resume: settle asynchronously and notify on completion exactly
+        // like a background spawn, returning immediately with the record still
+        // "running" — or "queued" when at the concurrency limit. Previously
+        // run_in_background was ignored on resume (the Agent tool's resume branch
+        // returned before its background branch, and resume() only ever awaited
+        // inline), so a resumed agent always blocked the caller until it finished.
+        if (options?.isBackground) {
+            // Never re-enter a run that is still in flight. Detaching means the caller
+            // gets control back while the record stays "running", so nothing stops the
+            // model from resuming the same agent again. Starting a second run would
+            // overwrite record.abortController — orphaning the live run beyond the
+            // reach of `/agents` stop and abortAll() — double-count the pool slot, and
+            // then reject from session.prompt() with "Agent is already processing",
+            // whose settle path would abort the LIVE run's children and report a
+            // failure for a run that is still going. Refuse instead, leaving the
+            // record untouched; the caller decides whether to wait or steer.
+            if (record.status === "running" || record.status === "queued")
+                return undefined;
+            record.isBackground = true;
+            record.resultConsumed = false;
+            record.result = undefined;
+            record.error = undefined;
+            record.completedAt = undefined;
+            record.status = "queued";
+            const start = () => this.startResume(id, record, prompt, signal, options);
+            if (occupiesPoolSlot(record) && !this.poolHasRoom("background")) {
+                // At the concurrency limit — queue it, drains when a slot frees. A
+                // detached resume has no inline caller, hence nothing to release. The
+                // queue is shared with spawns, whose startup is async, so entries are
+                // promise-shaped even though a resume starts synchronously; failures
+                // land on the record here, since drainQueue no longer catches.
+                this.queue.push({
+                    id,
+                    pool: "background",
+                    start: async () => {
+                        try {
+                            start();
+                        }
+                        catch (err) {
+                            record.status = "error";
+                            record.error = err instanceof Error ? err.message : String(err);
+                            record.completedAt = Date.now();
+                            this.onComplete?.(record);
+                        }
+                    },
+                    release: () => { },
+                });
+            }
+            else {
+                start();
+            }
+            return record;
+        }
+        // Foreground resume: run inline and return the settled record.
         record.status = "running";
         record.startedAt = Date.now();
         record.completedAt = undefined;
         record.result = undefined;
         record.error = undefined;
-        const resumedModel = record.session.model;
-        record.invocation = {
-            ...(record.invocation ?? {}),
-            ...(resumedModel && { effectiveModelName: resumedModel.name ?? resumedModel.id }),
-            effectiveThinking: record.session.thinkingLevel,
-        };
-        this.checkpoint(record);
         try {
             const { text, failure } = await resumeAgent(record.session, prompt, {
                 onToolActivity: (activity) => {
                     if (activity.type === "end")
                         record.toolUses++;
+                    options?.onToolActivity?.(activity);
                 },
                 onAssistantUsage: (usage) => {
                     addUsage(record.lifetimeUsage, usage);
+                    this.onUsage?.(record, usage);
+                    options?.onAssistantUsage?.(usage);
                 },
                 onCompaction: (info) => {
                     record.compactionCount++;
                     this.onCompact?.(record, info);
+                    options?.onCompaction?.(info);
                 },
                 signal,
             });
@@ -606,16 +1123,121 @@ export class AgentManager {
                 record.error = failure;
             record.result = text;
             record.completedAt = Date.now();
-            this.checkpoint(record);
         }
         catch (err) {
             record.status = "error";
             record.error = err instanceof Error ? err.message : String(err);
             record.completedAt = Date.now();
-            this.checkpoint(record);
         }
+        // Same contract as the spawn settle paths: children spawned during the
+        // resumed turn must not outlive it — nothing else can see or reach them.
+        this.abortOwnedChildren(id);
         return record;
     }
+    /**
+     * Start a background resume run: detached, settling and notifying like
+     * startAgent's background path. Invoked immediately, or from drainQueue when
+     * a concurrency slot frees. The session already exists (resume reuses it), so
+     * there is no onSessionCreated to hang per-run wiring off — callers use
+     * `options.onStarted`, which fires on both the immediate and the drained path.
+     */
+    startResume(id, record, prompt, parentSignal, options) {
+        if (!record.session)
+            return;
+        record.status = "running";
+        record.startedAt = Date.now();
+        if (occupiesPoolSlot(record))
+            this.runningBackground++;
+        this.onStart?.(record);
+        // Fresh abort controller so /agents stop and steering target THIS run rather
+        // than the previous one's settled controller.
+        const abortController = new AbortController();
+        record.abortController = abortController;
+        // Optional, and NOT what the Agent tool passes for a detached resume: a
+        // parent signal aborts on the parent's own interrupt (user Esc), which is
+        // right for a foreground run whose result the caller is awaiting, and wrong
+        // for a detached one — background spawns omit it for exactly this reason.
+        let detachParentSignal;
+        if (parentSignal) {
+            const onParentAbort = () => this.abort(id);
+            parentSignal.addEventListener("abort", onParentAbort, { once: true });
+            detachParentSignal = () => parentSignal.removeEventListener("abort", onParentAbort);
+        }
+        // Per-run durable history starts at the existing session tail. The prompt
+        // and all messages produced by this resumed run are then flushed by the
+        // same manager-owned seam as a fresh spawn.
+        if (record.historyFile) {
+            const cwd = this.recoveryCwds.get(id);
+            if (cwd) {
+                const startIndex = Array.isArray(record.session.messages) ? record.session.messages.length : 0;
+                record.historyCleanup = streamAgentHistory(record.session, record.historyFile, id, cwd, startIndex);
+            }
+        }
+        // Per-run side effects (optional `.output` streaming) — see ResumeOptions.onStarted.
+        // After the record is in its running shape, before the run is kicked off.
+        try {
+            options.onStarted?.();
+        }
+        catch { /* ignore caller wiring errors */ }
+        const settle = () => {
+            detachParentSignal?.();
+            detachParentSignal = undefined;
+            // Final flush of streaming files. The durable history stream is owned by
+            // the manager; the optional `.output` stream is caller-wired.
+            this.flushOutput(record);
+            // Children spawned during the resumed turn must not outlive it.
+            this.abortOwnedChildren(id);
+            if (occupiesPoolSlot(record))
+                this.runningBackground--;
+            try {
+                this.onComplete?.(record);
+            }
+            catch { /* ignore completion side-effect errors */ }
+            this.drainQueue();
+        };
+        const promise = resumeAgent(record.session, prompt, {
+            onToolActivity: (activity) => {
+                if (activity.type === "end")
+                    record.toolUses++;
+                options.onToolActivity?.(activity);
+            },
+            onAssistantUsage: (usage) => {
+                addUsage(record.lifetimeUsage, usage);
+                this.onUsage?.(record, usage);
+                options.onAssistantUsage?.(usage);
+            },
+            onCompaction: (info) => {
+                record.compactionCount++;
+                this.onCompact?.(record, info);
+                options.onCompaction?.(info);
+            },
+            signal: abortController.signal,
+        })
+            .then(({ text, failure }) => {
+            // Don't overwrite status if externally stopped via abort().
+            if (record.status !== "stopped") {
+                // Same contract as the spawn path (#144): a failed final turn is an
+                // error, not a completion — but the resumed text stays available.
+                record.status = failure ? "error" : "completed";
+                if (failure)
+                    record.error = failure;
+            }
+            record.result = text;
+            record.completedAt ??= Date.now();
+            settle();
+            return text;
+        })
+            .catch((err) => {
+            if (record.status !== "stopped") {
+                record.status = "error";
+                record.error = err instanceof Error ? err.message : String(err);
+            }
+            record.completedAt ??= Date.now();
+            settle();
+            return "";
+        });
+        record.promise = promise;
+    }
     /**
      * Send a steering message to an agent from the UI (mirrors the steer_subagent
      * tool). A live session delivers it now — it interrupts the agent after its
@@ -643,52 +1265,91 @@ export class AgentManager {
     getRecord(id) {
         return this.agents.get(id);
     }
-    listAgents() {
-        return [...this.agents.values()].sort((a, b) => b.startedAt - a.startedAt);
-    }
-    /** Restore terminal records persisted by a parent branch without runtime handles. */
-    restoreCompleted(records) {
-        const latest = new Map();
-        for (const value of records) {
-            if (isRestorableAgentRecord(value)) {
-                latest.set(value.id, this.createRestoredRecord(value));
-            }
+    /** Handles already in use, so a fresh spawn can pick an unclaimed one. */
+    takenHandles() {
+        const taken = new Set();
+        for (const record of this.agents.values()) {
+            if (record.handle)
+                taken.add(record.handle);
+            if (record.alias)
+                taken.add(record.alias);
+        }
+        // Tombstones hold their names too: an evicted `@explore` is still
+        // resurrectable, so a later Explore must become `explore-2` rather than
+        // shadowing a conversation the user can still reach.
+        for (const entry of this.tombstones.values()) {
+            taken.add(entry.handle);
+            if (entry.alias)
+                taken.add(entry.alias);
         }
-        for (const [id, restored] of latest) {
-            const existing = this.agents.get(id);
-            if (existing?.status === "running" || existing?.status === "queued")
+        return taken;
+    }
+    /**
+     * Resolve an `@name` from the prompt. Matches a top-level agent's handle
+     * case-insensitively, preferring one that can still be steered and otherwise
+     * the most recently started (which is the one a resume should continue), then
+     * falls back to an exact agent id so `@<agentId>` works too.
+     */
+    resolveMention(name) {
+        const wanted = name.toLowerCase();
+        let fallback;
+        for (const record of this.agents.values()) {
+            if (record.parentAgentId !== undefined)
                 continue;
-            this.agents.set(id, restored);
+            // Handle and alias share one namespace, so at most one agent answers a
+            // name and it makes no difference which of the two matched.
+            if (record.handle?.toLowerCase() !== wanted && record.alias?.toLowerCase() !== wanted)
+                continue;
+            if (record.status === "running" || record.status === "queued")
+                return { kind: "live", record };
+            if (!fallback || record.startedAt > fallback.startedAt)
+                fallback = record;
+        }
+        if (fallback)
+            return { kind: "live", record: fallback };
+        const byId = this.agents.get(name);
+        if (byId?.parentAgentId === undefined && byId !== undefined)
+            return { kind: "live", record: byId };
+        // Only once nothing live answers: a tombstone is a conversation to reopen,
+        // and reopening one while its record still exists would fork the session.
+        for (const entry of this.tombstones.values()) {
+            if (entry.handle.toLowerCase() === wanted || entry.alias?.toLowerCase() === wanted || entry.id === name) {
+                return { kind: "tombstone", entry };
+            }
         }
+        return undefined;
     }
-    createRestoredRecord(record) {
-        return {
-            id: record.id,
-            type: record.type,
-            description: record.description,
-            status: record.status,
-            result: record.result,
-            error: record.error,
-            toolUses: record.toolUses ?? 0,
-            startedAt: record.startedAt,
-            completedAt: record.completedAt,
-            transcriptPath: record.transcriptPath,
-            invocation: cloneInvocation(record.invocation),
-            lifetimeUsage: record.lifetimeUsage
-                ? { ...record.lifetimeUsage }
-                : { input: 0, output: 0, cacheWrite: 0 },
-            compactionCount: record.compactionCount ?? 0,
-        };
+    /**
+     * Forget an evicted agent, by handle. For the case where its session file has
+     * gone: the entry can then only ever fail, while still holding the name
+     * against the type that would otherwise start a fresh agent under it.
+     *
+     * A *successful* resume does not drop its tombstone — the live record it
+     * creates already wins in `resolveMention`, and overwrites the entry in place
+     * when it is itself evicted.
+     */
+    dropTombstone(handle) {
+        this.tombstones.delete(handle);
+    }
+    /** Evicted agents whose conversation can still be reopened, newest first. */
+    listTombstones() {
+        return [...this.tombstones.values()].sort((a, b) => b.completedAt - a.completedAt);
+    }
+    listAgents() {
+        return [...this.agents.values()].sort((a, b) => b.startedAt - a.startedAt);
     }
     abort(id) {
         const record = this.agents.get(id);
         if (!record)
             return false;
-        // Remove from queue if queued
+        // Remove from queue if queued. No decrement — the slot was never taken —
+        // and no onComplete, matching what a queued background abort has always
+        // done; a blocking caller learns of the stop from its own tool result.
         if (record.status === "queued") {
-            this.queue = this.queue.filter(q => q.id !== id);
+            this.dequeue(q => q.id === id);
             record.status = "stopped";
             record.completedAt = Date.now();
+            this.flushOutput(record);
             this.checkpoint(record);
             return true;
         }
@@ -699,15 +1360,48 @@ export class AgentManager {
         record.completedAt = Date.now();
         this.flushOutput(record);
         this.checkpoint(record);
-        this.finishBackground(id);
-        this.drainQueue();
         return true;
     }
     /** Dispose a record's session and remove it from the map. */
     removeRecord(id, record) {
-        record.session?.dispose?.();
+        this.tombstone(record);
+        const session = record.session;
+        // Detached before the shutdown starts, so the record leaves the map at once and
+        // nothing can observe a session that is half torn down.
         record.session = undefined;
         this.agents.delete(id);
+        // A failed startup keeps its (rejected) entry so a late awaitStartup still
+        // sees it; drop it with the record so the map can't grow unbounded.
+        this.startups.delete(id);
+        // Fire-and-forget is right here and only here: this runs from the 60s cleanup timer
+        // and from `clearCompleted()` on session boundaries, with the process staying alive,
+        // so handlers get their full window. The quit path awaits instead — see dispose().
+        void shutdownChildSession(session);
+    }
+    /**
+     * Preserve enough of a departing record for `@handle` to reopen its
+     * conversation later. Nothing to keep unless it has both a handle to be
+     * addressed by and a session file to reopen — an in-memory session leaves no
+     * transcript, so the mention would have nothing to continue from.
+     */
+    tombstone(record) {
+        if (!record.handle || !record.sessionFile)
+            return;
+        this.tombstones.set(record.handle, {
+            handle: record.handle,
+            alias: record.alias,
+            id: record.id,
+            type: record.type,
+            description: record.description,
+            sessionFile: record.sessionFile,
+            completedAt: record.completedAt ?? Date.now(),
+        });
+        // Bound the memory a long session can accumulate. Oldest first, since the
+        // agent someone still wants to reach is the one they used most recently.
+        while (this.tombstones.size > MAX_TOMBSTONES) {
+            const oldest = [...this.tombstones.values()].reduce((a, b) => (a.completedAt <= b.completedAt ? a : b));
+            this.tombstones.delete(oldest.handle);
+        }
     }
     cleanup() {
         const cutoff = Date.now() - 10 * 60_000;
@@ -716,29 +1410,6 @@ export class AgentManager {
                 continue;
             if ((record.completedAt ?? 0) >= cutoff)
                 continue;
-            // A durable transcript is the source of truth for history. Release the
-            // live session after the TTL, but retain a lightweight record so opening
-            // history again in this session does not silently lose its identity or
-            // locator. Records without durable storage remain eligible for eviction.
-            if (record.transcriptPath) {
-                try {
-                    record.session?.dispose?.();
-                }
-                catch { /* ignore cleanup failures */ }
-                record.session = undefined;
-                try {
-                    record.outputCleanup?.();
-                }
-                catch { /* ignore cleanup failures */ }
-                record.outputCleanup = undefined;
-                record.outputFile = undefined;
-                record.historyFile = undefined;
-                // The durable transcript is the source of truth after the TTL. Keep
-                // only the small identity/status record in memory; get_subagent_result
-                // reloads the final answer from transcriptPath on demand.
-                record.result = undefined;
-                continue;
-            }
             this.removeRecord(id, record);
         }
     }
@@ -756,6 +1427,13 @@ export class AgentManager {
                 continue;
             this.removeRecord(id, record);
         }
+        // Unconditional: both callers are session boundaries (`session_start` and
+        // `session_before_switch`), and `skipUnconsumed` only spares records whose
+        // results the LLM has yet to read — it does not make the sweep partial in
+        // the sense that matters here. A new session means new handles, or
+        // `@explore` would silently reach an agent the user never started. Claude
+        // Code resets its registry on `/clear` for the same reason.
+        this.tombstones.clear();
     }
     /** Whether any agents are still running or queued. */
     hasRunning() {
@@ -770,20 +1448,19 @@ export class AgentManager {
             if (record) {
                 record.status = "stopped";
                 record.completedAt = Date.now();
+                this.flushOutput(record);
                 this.checkpoint(record);
                 count++;
             }
         }
-        this.queue = [];
-        // Abort running agents. Flush before checkpointing so a catchable
-        // shutdown/session switch leaves the latest assistant message available.
+        this.dequeue(() => true);
+        // Abort running agents
         for (const record of this.agents.values()) {
             if (record.status === "running") {
                 record.abortController?.abort();
                 record.status = "stopped";
                 record.completedAt = Date.now();
                 this.flushOutput(record);
-                this.finishBackground(record.id);
                 this.checkpoint(record);
                 count++;
             }
@@ -796,39 +1473,76 @@ export class AgentManager {
         // agents finish they start queued ones, which need awaiting too.
         while (true) {
             this.drainQueue();
-            const pending = [...this.agents.values()]
-                .filter(r => r.status === "running" || r.status === "queued")
-                .map(r => r.promise)
-                .filter(Boolean);
+            const pending = [];
+            for (const record of this.agents.values()) {
+                if (record.status !== "running" && record.status !== "queued")
+                    continue;
+                // An agent whose worktree is still being created is "running" with no
+                // `promise` yet — without its startup the wait would return too early.
+                const startup = this.startups.get(record.id);
+                if (startup)
+                    pending.push(startup);
+                if (record.promise)
+                    pending.push(record.promise);
+            }
             if (pending.length === 0)
                 break;
             await Promise.allSettled(pending);
         }
     }
-    dispose() {
+    /**
+     * @param pi - Needed to run `git worktree prune`, which is async now and so
+     *   cannot be reached through a stored spawn argument at shutdown. Omitting
+     *   it (tests, teardown of a manager that never spawned) skips the prune.
+     */
+    async dispose(pi) {
         clearInterval(this.cleanupInterval);
-        // Clear queue
-        this.queue = [];
+        // Keep the pre-shutdown active snapshot marked active in the checkpoint. A
+        // process can still be killed after dispose starts; restoreRecovered turns
+        // that stale active marker into a stopped record, while the transcript has
+        // already received the terminal abort/stop event below.
+        const activeBeforeDispose = new Set([...this.agents.values()]
+            .filter(record => record.status === "running" || record.status === "queued")
+            .map(record => record.id));
+        // Mark live work stopped and flush durable history before child sessions are
+        // shut down. The caller also invokes abortAll(), but dispose is deliberately
+        // safe and complete when used on its own.
+        this.abortAll();
+        // Clear queue — via dequeue, so anyone blocked in spawnAndWait is woken
+        // rather than left awaiting a gate nothing will ever resolve.
+        this.dequeue(() => true);
         for (const record of this.agents.values()) {
-            record.session?.dispose();
+            this.flushOutput(record);
+            this.checkpoint(record);
+            if (activeBeforeDispose.has(record.id)) {
+                const activeSnapshot = this.makeCheckpoint(record);
+                delete activeSnapshot.completedAt;
+                activeSnapshot.status = "running";
+                const cwd = this.recoveryCwds.get(record.id);
+                if (cwd)
+                    writeAgentRecoveryCheckpoint(cwd, activeSnapshot);
+            }
         }
+        const sessions = [...this.agents.values()].map(record => record.session);
         this.agents.clear();
         this.recoveryCwds.clear();
-        this.runningBackgroundIds.clear();
-        this.runningBackground = 0;
-        // Prune any orphaned git worktrees (crash recovery)
-        try {
-            pruneWorktrees(process.cwd());
-        }
-        catch { /* ignore */ }
-        // Also prune repos that caller-supplied cwds created worktrees in — a clean
-        // exit with in-flight agents would otherwise leave stale registrations there.
-        for (const repo of this.worktreeRepos) {
-            try {
-                pruneWorktrees(repo);
-            }
-            catch { /* ignore */ }
+        this.startups.clear();
+        if (pi) {
+            // Prune any orphaned git worktrees (crash recovery). Detached: dispose runs
+            // on the shutdown path, which cannot wait for git. Started before the awaited
+            // shutdown below rather than after it, so the git calls have that window to
+            // finish in instead of racing the process exit that follows.
+            const prune = (repo) => { pruneWorktrees(pi, repo).catch(() => { }); };
+            prune(process.cwd());
+            // Also prune repos that caller-supplied cwds created worktrees in — a clean
+            // exit with in-flight agents would otherwise leave stale registrations there.
+            for (const repo of this.worktreeRepos)
+                prune(repo);
         }
+        // Awaited, unlike the eviction path: pi awaits this extension's `session_shutdown`
+        // handler and the process exits right after it returns, so anything left unawaited
+        // here never runs at all. Bounded — each call carries its own ceiling, concurrently.
+        await Promise.all(sessions.map(session => shutdownChildSession(session)));
     }
 }
 //# sourceMappingURL=agent-manager.js.map