fix(FN-5256): keep live task worktrees through pause/resume races

Three independent reliability fixes that all surfaced as the same bug:
live tasks losing their worktrees mid-execution and emitting
`wrong_toplevel` errors.

Fix A — executor stale-self-owned classifier:
`reconcileSelfOwnedActiveSessionForRemoval` now takes a process-active
probe (`executingTaskLock.has`) and a minimum-idle window (default 5s)
in addition to the existing in-memory `activeWorktrees` binding probe.
Recently-registered or still-running entries are refused with
`process-active-refuses` / `too-recent-refuses`, with audit-grade
log lines. Both the pre-remove path
(`reconcileSelfOwnedBeforeRemove`), the post-throw retry in
`removeOwnWorktreeWithReconcile`, and the defensive reconcile in
`removeWorktree` route through the same hardened gates.

Fix B — pause synchronously reaps the agent session:
New `awaitAbortInFlightTaskWork` mirrors the existing fire-and-forget
abort but awaits each `session.abort()` /
`stepExecutor.terminateAllSessions()` /
`workflowSession.abort()`. `parkTaskAfterWorkflowStepPause` calls it
before `moveTask("todo")`, and the `task:updated` user-pause handler
routes through it, so a fast re-dispatch can no longer race a still-
live shell.

Fix C — self-healing realpath + active-task skip:
`reconcileTaskWorktreeMetadata` now realpath-normalizes both sides of
the registry comparison (handling macOS `/private/var/...`) and
refuses to clear `worktree`/`branch` on in-progress or in-review
tasks. The skip emits a new
`task:auto-recover-worktree-metadata-skipped-active` audit event;
executor-level recovery paths remain in charge of active tasks.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.changeset/fix-fn-5256-worktree-liveness.md

@@ -0,0 +1,11 @@
+---
+"@runfusion/fusion": patch
+---
+
+fix(FN-5256): harden three independent code paths that were losing live task worktrees mid-execution and producing `wrong_toplevel` errors.
+
+- Executor stale-self-owned classifier (`reconcileSelfOwnedActiveSessionForRemoval`) now requires two additional signals before dropping a same-task registry entry: a process-active probe (`executingTaskLock.has`) and a minimum-idle window (default 5s) since the entry was registered. This closes the pause/resume race where the new executor cycle hadn't repopulated `activeWorktrees` yet and the old session's registry entry was reaped under a still-live shell. The post-throw reconcile in `removeOwnWorktreeWithReconcile` and the `removeWorktree` defensive reconcile in `worktree-backend.ts` route through the same hardened path.
+
+- Pause-before-park now synchronously awaits agent/step/workflow session disposal via a new `awaitAbortInFlightTaskWork` method, so by the time `parkTaskAfterWorkflowStepPause` calls `moveTask("todo")` the spawned shells are already reaped and any fast re-dispatch sees a clean slate. The user-initiated pause handler on `task:updated` was collapsed onto the same await path for the same reason.
+
+- Self-healing `reconcileTaskWorktreeMetadata` now normalizes both sides via `realpathSync` (with ENOENT fallback) before comparing the task worktree against the registered set, fixing the macOS `/private/var/...` false-stale flag. It additionally refuses to clear `worktree`/`branch` metadata for in-progress or in-review tasks — those go through `task:auto-recover-worktree-metadata-skipped-active` audit events and leave executor-level recovery in charge.

packages/engine/src/__tests__/active-session-registry.test.ts

@@ -94,7 +94,9 @@ describe("activeSessionRegistry", () => {
     activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
 
     expect(
-      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false),
+      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false, {
+        minIdleMs: 0,
+      }),
     ).toEqual({ action: "reconciled" });
     expect(activeSessionRegistry.lookupByPath("/tmp/w1")).toBeNull();
   });
@@ -103,10 +105,75 @@ describe("activeSessionRegistry", () => {
     activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
 
     expect(
-      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false),
+      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false, {
+        minIdleMs: 0,
+      }),
     ).toEqual({ action: "reconciled" });
     expect(
-      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false),
+      reconcileSelfOwnedActiveSessionForRemoval(activeSessionRegistry, "/tmp/w1", "FN-1", () => false, {
+        minIdleMs: 0,
+      }),
     ).toEqual({ action: "no-entry" });
   });
+
+  it("FN-5256: refuses reconcile when processActiveProbe returns true", () => {
+    activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
+
+    const outcome = reconcileSelfOwnedActiveSessionForRemoval(
+      activeSessionRegistry,
+      "/tmp/w1",
+      "FN-1",
+      () => false,
+      { processActiveProbe: () => true, minIdleMs: 0 },
+    );
+    expect(outcome).toEqual({ action: "process-active-refuses", ownerTaskId: "FN-1" });
+    expect(activeSessionRegistry.lookupByPath("/tmp/w1")?.taskId).toBe("FN-1");
+  });
+
+  it("FN-5256: refuses reconcile when registration is younger than minIdleMs", () => {
+    activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
+    const registeredAt = activeSessionRegistry.lookupByPath("/tmp/w1")!.registeredAt;
+
+    const outcome = reconcileSelfOwnedActiveSessionForRemoval(
+      activeSessionRegistry,
+      "/tmp/w1",
+      "FN-1",
+      () => false,
+      { minIdleMs: 5000, now: () => registeredAt + 100 },
+    );
+    expect(outcome).toMatchObject({ action: "too-recent-refuses", ownerTaskId: "FN-1", minIdleMs: 5000 });
+    expect(activeSessionRegistry.lookupByPath("/tmp/w1")?.taskId).toBe("FN-1");
+  });
+
+  it("FN-5256: reconciles when all signals clean (default min-idle window elapsed)", () => {
+    activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
+    const registeredAt = activeSessionRegistry.lookupByPath("/tmp/w1")!.registeredAt;
+
+    const outcome = reconcileSelfOwnedActiveSessionForRemoval(
+      activeSessionRegistry,
+      "/tmp/w1",
+      "FN-1",
+      () => false,
+      {
+        processActiveProbe: () => false,
+        minIdleMs: 5000,
+        now: () => registeredAt + 6000,
+      },
+    );
+    expect(outcome).toEqual({ action: "reconciled" });
+    expect(activeSessionRegistry.lookupByPath("/tmp/w1")).toBeNull();
+  });
+
+  it("FN-5256: live-binding takes precedence over process-active and too-recent", () => {
+    activeSessionRegistry.registerPath("/tmp/w1", { taskId: "FN-1", kind: "executor", ownerKey: "FN-1" });
+
+    const outcome = reconcileSelfOwnedActiveSessionForRemoval(
+      activeSessionRegistry,
+      "/tmp/w1",
+      "FN-1",
+      () => true,
+      { processActiveProbe: () => true, minIdleMs: 5000 },
+    );
+    expect(outcome).toEqual({ action: "live-binding-refuses", ownerTaskId: "FN-1" });
+  });
 });

packages/engine/src/__tests__/executor-pause.test.ts

@@ -2723,3 +2723,112 @@ describe("StepSessionExecutor integration", () => {
   });
 });
 
+describe("FN-5256 awaitAbortInFlightTaskWork pause synchronization", () => {
+  it("synchronously disposes activeSessions on pause via task:updated", async () => {
+    const store = createMockStore();
+    const executor = new TaskExecutor(store, "/tmp/test");
+
+    let abortResolve: (() => void) | undefined;
+    const abortPromise = new Promise<void>((resolve) => {
+      abortResolve = resolve;
+    });
+    const abortSpy = vi.fn(() => abortPromise);
+    const disposeSpy = vi.fn();
+    (executor as any).activeSessions.set("FN-PAUSE-1", {
+      session: { abort: abortSpy, dispose: disposeSpy, steer: vi.fn() },
+      seenSteeringIds: new Set(),
+    });
+
+    const taskUpdatedHandler = (store.on as unknown as ReturnType<typeof vi.fn>).mock.calls
+      .find((call: any[]) => call[0] === "task:updated")?.[1] as (task: any) => Promise<void>;
+    expect(taskUpdatedHandler).toBeTypeOf("function");
+
+    const settled = taskUpdatedHandler({
+      id: "FN-PAUSE-1",
+      title: "pause",
+      description: "pause",
+      column: "in-progress",
+      paused: true,
+      dependencies: [],
+      steps: [],
+      currentStep: 0,
+      log: [],
+      createdAt: new Date().toISOString(),
+      updatedAt: new Date().toISOString(),
+    } satisfies Task);
+
+    // Before abort resolves, dispose must NOT have been called — i.e. handler is
+    // truly awaiting the abort, not fire-and-forgetting it.
+    await new Promise((resolve) => setImmediate(resolve));
+    expect(disposeSpy).not.toHaveBeenCalled();
+
+    abortResolve?.();
+    await settled;
+
+    expect(abortSpy).toHaveBeenCalledTimes(1);
+    expect(disposeSpy).toHaveBeenCalledTimes(1);
+    expect((executor as any).activeSessions.has("FN-PAUSE-1")).toBe(false);
+  });
+
+  it("synchronously reaps workflow-step + step-session surfaces on pause", async () => {
+    const store = createMockStore();
+    const executor = new TaskExecutor(store, "/tmp/test");
+
+    const stepTerminate = vi.fn().mockResolvedValue(undefined);
+    (executor as any).activeStepExecutors.set("FN-PAUSE-2", {
+      terminateAllSessions: stepTerminate,
+    });
+
+    const workflowAbort = vi.fn().mockResolvedValue(undefined);
+    const workflowDispose = vi.fn();
+    (executor as any).activeWorkflowStepSessions.set("FN-PAUSE-2", {
+      abort: workflowAbort,
+      dispose: workflowDispose,
+    });
+
+    const taskUpdatedHandler = (store.on as unknown as ReturnType<typeof vi.fn>).mock.calls
+      .find((call: any[]) => call[0] === "task:updated")?.[1] as (task: any) => Promise<void>;
+
+    await taskUpdatedHandler({
+      id: "FN-PAUSE-2",
+      title: "pause",
+      description: "pause",
+      column: "in-progress",
+      paused: true,
+      dependencies: [],
+      steps: [],
+      currentStep: 0,
+      log: [],
+      createdAt: new Date().toISOString(),
+      updatedAt: new Date().toISOString(),
+    } satisfies Task);
+
+    expect(stepTerminate).toHaveBeenCalledTimes(1);
+    expect(workflowAbort).toHaveBeenCalledTimes(1);
+    expect(workflowDispose).toHaveBeenCalledTimes(1);
+    expect((executor as any).activeStepExecutors.has("FN-PAUSE-2")).toBe(false);
+    expect((executor as any).activeWorkflowStepSessions.has("FN-PAUSE-2")).toBe(false);
+  });
+
+  it("awaitAbortInFlightTaskWork awaits abort before dispose for all surfaces", async () => {
+    const store = createMockStore();
+    const executor = new TaskExecutor(store, "/tmp/test");
+
+    const order: string[] = [];
+    const sessionAbort = vi.fn().mockImplementation(async () => {
+      await new Promise((r) => setTimeout(r, 5));
+      order.push("session-abort");
+    });
+    const sessionDispose = vi.fn(() => order.push("session-dispose"));
+    (executor as any).activeSessions.set("FN-PAUSE-3", {
+      session: { abort: sessionAbort, dispose: sessionDispose, steer: vi.fn() },
+      seenSteeringIds: new Set(),
+    });
+
+    await (executor as any).awaitAbortInFlightTaskWork("FN-PAUSE-3", "test reason");
+
+    expect(order).toEqual(["session-abort", "session-dispose"]);
+    expect((executor as any).activeSessions.has("FN-PAUSE-3")).toBe(false);
+  });
+});
+

packages/engine/src/__tests__/executor-prompt.test.ts

@@ -1939,7 +1939,10 @@ describe("TaskExecutor task:updated listener guards", () => {
     } satisfies Task)).resolves.toBeUndefined();
 
     expect(terminateAllSessions).toHaveBeenCalledTimes(1);
-    expect(executorLog.error).toHaveBeenCalledWith("Uncaught error in task:updated listener:", terminateError);
+    // FN-5256: the pause handler now routes through awaitAbortInFlightTaskWork,
+    // which internally catches/logs the per-surface failure. The error still hits
+    // executorLog.error but via the granular message path.
+    expect(executorLog.error).toHaveBeenCalledWith("Failed to terminate step sessions for FN-001:", terminateError);
   });
 });
 

packages/engine/src/__tests__/executor-worktree-conflict.test.ts

@@ -19,6 +19,8 @@ describe("FN-4973: executor worktree conflict cleanup", () => {
     const executor = new TaskExecutor(store, "/tmp/test");
     store.listTasks.mockResolvedValue([]);
     activeSessionRegistry.registerPath(CONFLICT_PATH, { taskId: "FN-4973", kind: "executor", ownerKey: "FN-4973" });
+    // FN-5256: backdate so the new min-idle window doesn't refuse the reconcile.
+    (activeSessionRegistry.lookupByPath(CONFLICT_PATH) as any).registeredAt = 0;
 
     const removeSpy = vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);
     const result = await (executor as any).cleanupConflictingWorktree(CONFLICT_PATH, "fusion/fn-4973", "FN-4973");
@@ -85,6 +87,8 @@ describe("FN-4973: executor worktree conflict cleanup", () => {
     removeSpy
       .mockImplementationOnce(async () => {
         activeSessionRegistry.registerPath(CONFLICT_PATH, { taskId: "FN-4973", kind: "executor", ownerKey: "FN-4973" });
+        // FN-5256: backdate so the post-throw reconcile is not refused by min-idle.
+        (activeSessionRegistry.lookupByPath(CONFLICT_PATH) as any).registeredAt = 0;
         throw new ActiveSessionWorktreeRemovalError({
           worktreePath: CONFLICT_PATH,
           taskId: "FN-4973",

packages/engine/src/__tests__/reliability-interactions/post-completion-stale-self-owned-binding.test.ts

@@ -25,6 +25,7 @@ describe("FN-5346 reliability interactions: post-completion stale self-owned bin
     const executor = new TaskExecutor(store, ROOT);
     (executor as any).activeWorktrees.set(TASK_ID, PATH);
     activeSessionRegistry.registerPath(PATH, { taskId: TASK_ID, kind: "executor", ownerKey: TASK_ID });
+    (activeSessionRegistry.lookupByPath(PATH) as any).registeredAt = 0;
     const removeSpy = vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);
 
     await executor.cleanup(TASK_ID);
@@ -43,6 +44,7 @@ describe("FN-5346 reliability interactions: post-completion stale self-owned bin
     const store = createMockStore();
     const executor = new TaskExecutor(store, ROOT);
     activeSessionRegistry.registerPath(PATH, { taskId: TASK_ID, kind: "executor", ownerKey: TASK_ID });
+    (activeSessionRegistry.lookupByPath(PATH) as any).registeredAt = 0;
     const removeSpy = vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);
 
     await (executor as any).handleDepAbortCleanup(TASK_ID, PATH);
@@ -96,6 +98,7 @@ describe("FN-5346 reliability interactions: post-completion stale self-owned bin
     const executor = new TaskExecutor(store, ROOT);
     (executor as any).activeWorktrees.set(TASK_ID, PATH);
     activeSessionRegistry.registerPath(PATH, { taskId: TASK_ID, kind: "executor", ownerKey: TASK_ID });
+    (activeSessionRegistry.lookupByPath(PATH) as any).registeredAt = 0;
     const removeSpy = vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);
 
     await executor.cleanup(TASK_ID);

packages/engine/src/__tests__/reliability-interactions/stale-self-owned-active-session-recovery.test.ts

@@ -74,6 +74,8 @@ describe("FN-4973 reliability interactions: stale self-owned active-session reco
     store.listTasks.mockResolvedValue([]);
     const executor = new TaskExecutor(store, "/tmp/test");
     activeSessionRegistry.registerPath(CONFLICT_PATH, { taskId: TASK_ID, kind: "executor", ownerKey: TASK_ID });
+    // FN-5256: backdate so the new min-idle window doesn't refuse the reconcile.
+    (activeSessionRegistry.lookupByPath(CONFLICT_PATH) as any).registeredAt = 0;
 
     vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);
 

packages/engine/src/__tests__/reliability-interactions/stale-self-owned-session-registry.test.ts

@@ -25,6 +25,8 @@ describe("FN-4976: stale self-owned activeSessionRegistry deadlock backstop", ()
     const store = createMockStore();
     store.listTasks.mockResolvedValue([]);
     activeSessionRegistry.registerPath(PATH, { taskId: TASK_ID, kind: "executor", ownerKey: TASK_ID });
+    // FN-5256: backdate so the new min-idle window doesn't refuse the reconcile.
+    (activeSessionRegistry.lookupByPath(PATH) as any).registeredAt = 0;
 
     const executor = new TaskExecutor(store, ROOT);
     const removeSpy = vi.spyOn(worktreePoolModule, "removeWorktree").mockResolvedValue(undefined);

packages/engine/src/__tests__/self-healing-already-merged.real-git.test.ts

@@ -129,7 +129,10 @@ describeIfGit("SelfHealingManager recoverAlreadyMergedReviewTasks (real git)", (
     expect(task.mergeDetails?.mergeConfirmed).toBe(true);
     expect(existsSync(worktreePath)).toBe(false);
     expect(git(repo, "git worktree list")).not.toContain(worktreePath);
-    expect((store as any).recordRunAuditEvent).toHaveBeenCalledTimes(3);
+    // FN-5256: reconcileTaskWorktreeMetadata now normalizes via realpath, so the
+    // (formerly false-stale) macOS realpath mismatch no longer triggers an extra
+    // worktree-metadata-cleared audit event for this in-review task.
+    expect((store as any).recordRunAuditEvent).toHaveBeenCalledTimes(2);
     expect((store as any).recordRunAuditEvent).toHaveBeenCalledWith(
       expect.objectContaining({
         domain: "database",

packages/engine/src/__tests__/self-healing-worktree-metadata.test.ts

@@ -1,4 +1,4 @@
-import { mkdtempSync, mkdirSync, readFileSync, realpathSync, rmSync, writeFileSync } from "node:fs";
+import { mkdtempSync, mkdirSync, readFileSync, realpathSync, rmSync, symlinkSync, writeFileSync } from "node:fs";
 import { fileURLToPath } from "node:url";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
@@ -167,3 +167,98 @@ describe("reconcileTaskWorktreeMetadata matrix", () => {
     expect((store as any).updateTask).not.toHaveBeenCalled();
   });
 });
+
+describe("FN-5256 reconcileTaskWorktreeMetadata reliability", () => {
+  let rootDir = "";
+  let store: TaskStore;
+
+  beforeEach(() => {
+    rootDir = mkdtempSync(join(tmpdir(), "fn-5256-"));
+    git(rootDir, "init -b main");
+    git(rootDir, "config user.name 'Fusion'");
+    git(rootDir, "config user.email 'hi@runfusion.ai'");
+    writeFileSync(join(rootDir, "README.md"), "root\n");
+    git(rootDir, "add README.md");
+    git(rootDir, "commit -m 'init'");
+
+    store = new TaskStore(rootDir, undefined, { inMemoryDb: false });
+  });
+
+  afterEach(() => {
+    try {
+      store?.close();
+    } catch {
+      // noop
+    }
+    if (rootDir) rmSync(rootDir, { recursive: true, force: true });
+    vi.clearAllMocks();
+  });
+
+  it("FN-5256: realpath-normalizes both sides so a symlinked task.worktree is not falsely flagged stale", async () => {
+    await store.createTask({ title: "FN-5256 symlink", description: "symlink case" });
+    const [task] = await store.listTasks();
+
+    mkdirSync(join(rootDir, ".worktrees"), { recursive: true });
+    const realWorktreeDir = join(realpathSync(rootDir), ".worktrees", "real-leaf");
+    const branch = `fusion/${task.id.toLowerCase()}`;
+    git(rootDir, `branch ${branch}`);
+    git(rootDir, `worktree add ${realWorktreeDir} ${branch}`);
+
+    // Create a symlink that points at the registered worktree's parent. The task
+    // metadata persists the symlinked path; the registry will surface realpath.
+    const symlinkParent = join(rootDir, ".worktrees-symlink");
+    symlinkSync(join(rootDir, ".worktrees"), symlinkParent, "dir");
+    const symlinkedTaskWorktree = join(symlinkParent, "real-leaf");
+
+    await store.updateTask(task.id, { worktree: symlinkedTaskWorktree, branch });
+
+    const auditSpy = vi.spyOn(store, "recordRunAuditEvent");
+    const manager = new SelfHealingManager(store, {
+      rootDir,
+      getExecutingTaskIds: () => new Set<string>(),
+    });
+
+    const repaired = await (manager as any).reconcileTaskWorktreeMetadata();
+    expect(repaired).toBe(0);
+
+    const updated = await store.getTask(task.id);
+    expect(updated?.worktree).toBe(symlinkedTaskWorktree);
+    expect(updated?.branch).toBe(branch);
+    expect(auditSpy).not.toHaveBeenCalledWith(
+      expect.objectContaining({ mutationType: "task:auto-recover-worktree-metadata-cleared" }),
+    );
+  });
+
+  it("FN-5256: refuses to clear worktree metadata for an in-progress task with a stale-flagged worktree", async () => {
+    await store.createTask({ title: "FN-5256 in-progress", description: "stale active task" });
+    const [task] = await store.listTasks();
+
+    await store.moveTask(task.id, "todo");
+    await store.moveTask(task.id, "in-progress");
+    // The worktree path is bogus (not registered) but the task is active. The
+    // reconciler must not yank metadata out from under a running task.
+    const stalePath = join(rootDir, ".worktrees", "ghost-leaf");
+    mkdirSync(stalePath, { recursive: true });
+    await store.updateTask(task.id, { worktree: stalePath, branch: `fusion/${task.id.toLowerCase()}` });
+
+    const auditSpy = vi.spyOn(store, "recordRunAuditEvent");
+    const manager = new SelfHealingManager(store, {
+      rootDir,
+      getExecutingTaskIds: () => new Set<string>(),
+    });
+
+    const repaired = await (manager as any).reconcileTaskWorktreeMetadata();
+    expect(repaired).toBe(0);
+
+    const updated = await store.getTask(task.id);
+    expect(updated?.worktree).toBe(stalePath);
+    expect(updated?.branch).toBe(`fusion/${task.id.toLowerCase()}`);
+
+    expect(auditSpy).not.toHaveBeenCalledWith(
+      expect.objectContaining({ mutationType: "task:auto-recover-worktree-metadata-cleared" }),
+    );
+    expect(auditSpy).toHaveBeenCalledWith(
+      expect.objectContaining({ mutationType: "task:auto-recover-worktree-metadata-skipped-active" }),
+    );
+  });
+});

packages/engine/src/__tests__/worktree-backend.test.ts

@@ -873,6 +873,9 @@ describe("removeWorktree", () => {
       reason: RemovalReason.ExecutorDispose,
       expectedOwnerTaskId: "FN-1",
       liveOwnerProbe: () => false,
+      // FN-5256: opt out of the min-idle window so this defensive-reconcile test
+      // is unaffected by the new warm-up gate.
+      reconcileMinIdleMs: 0,
     });
 
     expect(activeSessionRegistry.lookupByPath("/repo/.worktrees/fn-1")).toBeNull();

packages/engine/src/active-session-registry.ts

@@ -16,13 +16,24 @@ export interface ReconcileStaleSelfOwnedResult {
 }
 
 export type LiveBindingProbe = (worktreePath: string, taskId: string) => boolean;
+export type ProcessActiveProbe = (taskId: string) => boolean;
 
 export type SelfOwnedReconcileOutcome =
   | { action: "no-entry" }
   | { action: "foreign-task"; ownerTaskId: string }
   | { action: "live-binding-refuses"; ownerTaskId: string }
+  | { action: "process-active-refuses"; ownerTaskId: string }
+  | { action: "too-recent-refuses"; ownerTaskId: string; ageMs: number; minIdleMs: number }
   | { action: "reconciled" };
 
+/**
+ * FN-5256: default minimum age before a self-owned registry entry can be classified
+ * as stale. Recently-registered entries belong to an executor cycle that is still
+ * warming up (e.g., a pause/resume that hasn't repopulated activeWorktrees yet), so
+ * dropping them races with the live shell that just attached to the worktree.
+ */
+export const DEFAULT_SELF_OWNED_MIN_IDLE_MS = 5000;
+
 export class ActiveSessionRegistry {
   private readonly records = new Map<string, ActiveSessionRecord>();
 
@@ -76,11 +87,29 @@ export class ActiveSessionRegistry {
   }
 }
 
+export interface SelfOwnedReconcileOptions {
+  /**
+   * Process-wide "executor still owns this task" probe. When this returns true the
+   * caller's task is still in the middle of an `execute()` invocation, so dropping
+   * the registry entry would yank the worktree from a live shell (FN-5256).
+   */
+  processActiveProbe?: ProcessActiveProbe;
+  /**
+   * Minimum age (ms since `registeredAt`) before a same-task entry is eligible for
+   * stale reconciliation. Recently-registered entries belong to a warming executor
+   * cycle and must be left alone. Defaults to `DEFAULT_SELF_OWNED_MIN_IDLE_MS`.
+   */
+  minIdleMs?: number;
+  /** Test seam — defaults to `Date.now()`. */
+  now?: () => number;
+}
+
 export function reconcileSelfOwnedActiveSessionForRemoval(
   registry: ActiveSessionRegistry,
   worktreePath: string,
   requestingTaskId: string,
   liveBindingProbe: LiveBindingProbe,
+  options: SelfOwnedReconcileOptions = {},
 ): SelfOwnedReconcileOutcome {
   const record = registry.lookupByPath(worktreePath);
   if (!record) {
@@ -95,6 +124,19 @@ export function reconcileSelfOwnedActiveSessionForRemoval(
     return { action: "live-binding-refuses", ownerTaskId: requestingTaskId };
   }
 
+  if (options.processActiveProbe?.(requestingTaskId)) {
+    return { action: "process-active-refuses", ownerTaskId: requestingTaskId };
+  }
+
+  const minIdleMs = options.minIdleMs ?? DEFAULT_SELF_OWNED_MIN_IDLE_MS;
+  if (minIdleMs > 0) {
+    const now = options.now?.() ?? Date.now();
+    const ageMs = now - record.registeredAt;
+    if (ageMs < minIdleMs) {
+      return { action: "too-recent-refuses", ownerTaskId: requestingTaskId, ageMs, minIdleMs };
+    }
+  }
+
   registry.unregisterPath(worktreePath);
   return { action: "reconciled" };
 }

packages/engine/src/executor.ts

@@ -1196,6 +1196,13 @@ export class TaskExecutor {
       undefined,
       this.getRunContextFor(taskId),
     ).catch(() => undefined);
+    // FN-5256: synchronously reap any spawned shells BEFORE moving the task. The
+    // task:moved listener fires `abortInFlightTaskWork` (fire-and-forget) so by the
+    // time it runs, the abort needs to already be complete — otherwise a fast
+    // re-dispatch races the live shell and yanks the worktree.
+    await this.awaitAbortInFlightTaskWork(taskId, "pause-before-park").catch((err) => {
+      executorLog.warn(`${taskId}: awaitAbortInFlightTaskWork failed in pause-before-park: ${err}`);
+    });
     if (latestTask.column === "in-progress") {
       await this.store.moveTask(taskId, "todo", { preserveResumeState: true });
     }
@@ -1445,6 +1452,88 @@ export class TaskExecutor {
     this.activeSubagentSessions.delete(taskId);
   }
 
+  /**
+   * FN-5256: synchronously await session disposal so callers (e.g. pause-before-park)
+   * can rely on the worktree-bound shells being reaped before they return. Mirrors
+   * `abortInFlightTaskWork`, but awaits the async `abort()` / `terminateAllSessions()`
+   * calls instead of fire-and-forget.
+   */
+  async awaitAbortInFlightTaskWork(taskId: string, reason: string, options: { userCanceled?: boolean } = {}): Promise<void> {
+    let hadActiveSurface = false;
+
+    if (options.userCanceled) {
+      this.userCanceledTaskIds.add(taskId);
+    }
+    this.pausedAborted.add(taskId);
+    this.options.stuckTaskDetector?.untrackTask(taskId);
+    this.clearWorkflowRerunWatchdog(taskId);
+    this.clearCompletedTaskWatchdog(taskId);
+
+    if (this.activeSessions.has(taskId)) {
+      hadActiveSurface = true;
+      const { session } = this.activeSessions.get(taskId)!;
+      const sessionWithAbort = session as AgentSession & { abort?: () => Promise<void> };
+      if (typeof sessionWithAbort.abort === "function") {
+        await sessionWithAbort.abort().catch((err) => {
+          executorLog.warn(`Failed to abort agent session for ${taskId}: ${err}`);
+        });
+      }
+      try {
+        session.dispose();
+      } catch (err) {
+        executorLog.warn(`Failed to dispose agent session for ${taskId}: ${err}`);
+      }
+      this.deleteActiveSession(taskId);
+    }
+
+    if (this.activeStepExecutors.has(taskId)) {
+      hadActiveSurface = true;
+      const stepExecutor = this.activeStepExecutors.get(taskId)!;
+      const stepExecutorWithAbort = stepExecutor as StepSessionExecutor & { abortAllSessionBash?: () => void };
+      if (typeof stepExecutorWithAbort.abortAllSessionBash === "function") {
+        try {
+          stepExecutorWithAbort.abortAllSessionBash();
+        } catch (err) {
+          executorLog.warn(`Failed to abort step-session bash for ${taskId}: ${err}`);
+        }
+      }
+      await stepExecutor.terminateAllSessions().catch((err) =>
+        executorLog.error(`Failed to terminate step sessions for ${taskId}:`, err),
+      );
+      this.deleteActiveStepExecutor(taskId);
+    }
+
+    if (this.activeWorkflowStepSessions.has(taskId)) {
+      hadActiveSurface = true;
+      const workflowSession = this.activeWorkflowStepSessions.get(taskId)!;
+      const sessionWithAbort = workflowSession as AgentSession & { abort?: () => Promise<void> };
+      if (typeof sessionWithAbort.abort === "function") {
+        await sessionWithAbort.abort().catch((err) => {
+          executorLog.warn(`Failed to abort workflow step session for ${taskId}: ${err}`);
+        });
+      }
+      try {
+        workflowSession.dispose();
+      } catch (err) {
+        executorLog.warn(`Failed to dispose workflow step session for ${taskId}: ${err}`);
+      }
+      this.deleteActiveWorkflowStepSession(taskId);
+    }
+
+    if (this.activeSubagentSessions.has(taskId)) {
+      hadActiveSurface = true;
+      this.disposeSubagentsForTask(taskId, reason);
+    }
+
+    this.loopRecoveryState.delete(taskId);
+    this.spawnedAgents.delete(taskId);
+    this.stuckAborted.delete(taskId);
+
+    if (hadActiveSurface) {
+      executorLog.log(`${taskId}: awaited abort of in-flight work — ${reason}`);
+    }
+  }
+
   private abortInFlightTaskWork(taskId: string, reason: string, options: { userCanceled?: boolean } = {}): void {
     let hadActiveSurface = false;
 
@@ -1595,50 +1684,19 @@ export class TaskExecutor {
     // 5. Each injection is logged to the task for user visibility
     store.on("task:updated", async (task) => {
       try {
-        // Handle pause - terminate the agent session or step sessions
+        // FN-5256: handle pause by synchronously reaping every active session
-        if (task.paused && this.activeSessions.has(task.id)) {
+        // surface in one shot. Awaiting the abort ensures spawned shells are
-          executorLog.log(`Pausing ${task.id} — terminating agent session`);
+        // disposed before any re-dispatch can race the worktree.
-          this.pausedAborted.add(task.id);
+        if (
-          this.options.stuckTaskDetector?.untrackTask(task.id);
+          task.paused
-          const { session } = this.activeSessions.get(task.id)!;
+          && (
-          session.dispose();
+            this.activeSessions.has(task.id)
-          // Also clean up in-memory state to prevent leaks if the task is later unpaused
+            || this.activeStepExecutors.has(task.id)
-          this.loopRecoveryState.delete(task.id);
+            || this.activeWorkflowStepSessions.has(task.id)
-          this.spawnedAgents.delete(task.id);
+          )
-          this.stuckAborted.delete(task.id);
+        ) {
-          this.disposeSubagentsForTask(task.id, "task paused");
+          executorLog.log(`Pausing ${task.id} — awaiting in-flight session disposal`);
-          return;
+          await this.awaitAbortInFlightTaskWork(task.id, "task paused");
-        }
-        if (task.paused && this.activeStepExecutors.has(task.id)) {
-          executorLog.log(`Pausing ${task.id} — terminating step sessions`);
-          this.pausedAborted.add(task.id);
-          this.options.stuckTaskDetector?.untrackTask(task.id);
-          const stepExecutor = this.activeStepExecutors.get(task.id)!;
-          await stepExecutor.terminateAllSessions();
-          // Also clean up in-memory state to prevent leaks if the task is later unpaused
-          this.loopRecoveryState.delete(task.id);
-          this.spawnedAgents.delete(task.id);
-          this.stuckAborted.delete(task.id);
-          this.disposeSubagentsForTask(task.id, "task paused");
-          return;
-        }
-        if (task.paused && this.activeWorkflowStepSessions.has(task.id)) {
-          executorLog.log(`Pausing ${task.id} — terminating workflow step session`);
-          this.pausedAborted.add(task.id);
-          this.options.stuckTaskDetector?.untrackTask(task.id);
-          const workflowSession = this.activeWorkflowStepSessions.get(task.id)!;
-          const sessionWithAbort = workflowSession as AgentSession & { abort?: () => Promise<void> };
-          if (typeof sessionWithAbort.abort === "function") {
-            await sessionWithAbort.abort().catch((err) =>
-              executorLog.warn(`Failed to abort workflow step session for pause ${task.id}: ${err}`),
-            );
-          }
-          workflowSession.dispose();
-          this.deleteActiveWorkflowStepSession(task.id);
-          this.loopRecoveryState.delete(task.id);
-          this.spawnedAgents.delete(task.id);
-          this.stuckAborted.delete(task.id);
-          this.disposeSubagentsForTask(task.id, "task paused");
           return;
         }
 
@@ -9172,12 +9230,33 @@ Backward compat fallback: if JSON is unavailable, you may still begin output wit
       worktreePath,
       taskId,
       (path, ownerTaskId) => this.hasActiveWorktreeBinding(ownerTaskId, path),
+      {
+        processActiveProbe: (probeTaskId) => executingTaskLock.has(probeTaskId),
+      },
     );
     if (outcome.action === "reconciled") {
       executorLog.warn(
         `[FN-5346] ${taskId}: dropped stale self-owned activeSessionRegistry entry before removeWorktree at ${worktreePath}`,
       );
       await this.store.logEntry(taskId, "Cleared stale self-owned active-session entry before remove", worktreePath);
+    } else if (outcome.action === "process-active-refuses") {
+      executorLog.warn(
+        `[FN-5256] refused stale-self-owned reconcile for ${taskId}: process-active=true at ${worktreePath}`,
+      );
+      await this.store.logEntry(
+        taskId,
+        "Refused stale self-owned reconcile — task still actively executing",
+        worktreePath,
+      ).catch(() => undefined);
+    } else if (outcome.action === "too-recent-refuses") {
+      executorLog.warn(
+        `[FN-5256] refused stale-self-owned reconcile for ${taskId}: age=${outcome.ageMs}ms (<${outcome.minIdleMs}ms) at ${worktreePath}`,
+      );
+      await this.store.logEntry(
+        taskId,
+        `Refused stale self-owned reconcile — registration too recent (${outcome.ageMs}ms < ${outcome.minIdleMs}ms)`,
+        worktreePath,
+      ).catch(() => undefined);
     }
   }
 
@@ -9198,6 +9277,9 @@ Backward compat fallback: if JSON is unavailable, you may still begin output wit
       audit: input.audit,
       expectedOwnerTaskId: input.taskId,
       liveOwnerProbe: (path: string, ownerTaskId: string) => this.hasActiveWorktreeBinding(ownerTaskId, path),
+      // FN-5256: route the worktree-backend defensive reconcile through the
+      // hardened gates (process-active + min-idle window).
+      processActiveProbe: (probeTaskId: string) => executingTaskLock.has(probeTaskId),
     } as const;
     try {
       await removeWorktree(removeArgs);
@@ -9207,16 +9289,32 @@ Backward compat fallback: if JSON is unavailable, you may still begin output wit
         && error.details.taskId === input.taskId
         && !this.hasActiveWorktreeBinding(input.taskId, input.worktreePath)
       ) {
-        const reconcileResult = activeSessionRegistry.reconcileStaleSelfOwned(input.worktreePath, input.taskId);
+        // FN-5256: route the post-throw reconcile through the hardened path so
-        if (reconcileResult.reconciled) {
+        // process-active and too-recent signals also gate this leg.
+        const outcome = reconcileSelfOwnedActiveSessionForRemoval(
+          activeSessionRegistry,
+          input.worktreePath,
+          input.taskId,
+          (path, ownerTaskId) => this.hasActiveWorktreeBinding(ownerTaskId, path),
+          {
+            processActiveProbe: (probeTaskId) => executingTaskLock.has(probeTaskId),
+          },
+        );
+        if (outcome.action === "reconciled") {
           await this.store.logEntry(
             input.taskId,
             "Reconciled stale self-owned active-session registration (post-throw)",
             input.worktreePath,
           );
+          await removeWorktree(removeArgs);
+          return;
+        }
+        if (outcome.action === "process-active-refuses" || outcome.action === "too-recent-refuses") {
+          executorLog.warn(
+            `[FN-5256] post-throw reconcile refused for ${input.taskId} at ${input.worktreePath}: action=${outcome.action}`,
+          );
+          // Refused — surface the original error so the caller can decide.
         }
-        await removeWorktree(removeArgs);
-        return;
       }
       throw error;
     }

packages/engine/src/run-audit.ts

@@ -208,6 +208,7 @@ export type DatabaseMutationType =
   | "task:auto-recover-node-unreachable"
   | "task:auto-recover-worktree-metadata-rebound"
   | "task:auto-recover-worktree-metadata-cleared"
+  | "task:auto-recover-worktree-metadata-skipped-active"
   // task:auto-archived-ghost-bug metadata: { findings: Array<{ construct: { kind: string; raw: string; filePath?: string; line?: number }; matched: boolean; probeError?: string; output?: string }>; reason: string }
   // task:auto-archived-duplicate metadata: { siblingTaskIds: string[]; scores: Record<string, number> }
   // task:broad-scope-flagged-at-triage metadata: { score: number; reasons: string[]; signals: { size: "S"|"M"|"L"|null; stepCount: number; fileScopeCount: number; failingFileMentions: number }; thresholds: { stepsHigh: number; fileScopeHigh: number; failingFileMentionsHigh: number; sizeLStepsThreshold: number }; version: number }

packages/engine/src/self-healing.ts

@@ -25,7 +25,7 @@
 
 import { exec, execSync } from "node:child_process";
 import { promisify } from "node:util";
-import { existsSync, mkdirSync, readdirSync, readFileSync, rmSync, statSync, writeFileSync } from "node:fs";
+import { existsSync, mkdirSync, readdirSync, readFileSync, realpathSync, rmSync, statSync, writeFileSync } from "node:fs";
 import { isAbsolute, join, relative, resolve } from "node:path";
 import { IN_REVIEW_STALL_DEADLOCK_LOG_PREFIX, IN_REVIEW_STALL_LOG_PREFIX, countRecentIdenticalStallEntries, detectSelfDefeatingDependency, getInReviewStalledSignal, getInReviewStallReason, getPrimaryPrInfo, getStalePausedReviewSignal, getStalePausedTodoSignal, getTaskHardMergeBlocker, getTaskMergeBlocker, isEphemeralAgent, parseExplicitDuplicateMarker, type AgentStore, type ChatStore, type MessageStore, type TaskStore, type Settings, type Task, type MergeDetails, type TaskPriority, type MergeResult } from "@fusion/core";
 import type { MeshLeaseManager } from "./mesh-lease-manager.js";
@@ -2576,7 +2576,10 @@ export class SelfHealingManager {
 
   private async emitWorktreeMetadataAuditEvent(input: {
     taskId: string;
-    mutationType: "task:auto-recover-worktree-metadata-rebound" | "task:auto-recover-worktree-metadata-cleared";
+    mutationType:
+      | "task:auto-recover-worktree-metadata-rebound"
+      | "task:auto-recover-worktree-metadata-cleared"
+      | "task:auto-recover-worktree-metadata-skipped-active";
     previousWorktree: string | null;
     newWorktree: string | null;
     previousBranch: string | null;
@@ -2809,7 +2812,20 @@ export class SelfHealingManager {
 
       const allTasks = await this.store.listTasks({ slim: true, includeArchived: false });
       const branchMap = await getRegisteredWorktreeBranchMap(this.options.rootDir);
-      const registeredPaths = new Set(branchMap.values());
+      // FN-5256: macOS git surfaces realpath-normalized worktree paths (/private/var/...)
+      // while task.worktree may be persisted as the symlinked path. Compare on realpath
+      // to avoid false-stale flagging that yanks a live worktree.
+      const safeRealpath = (path: string): string => {
+        try {
+          return realpathSync(path);
+        } catch {
+          return path;
+        }
+      };
+      const registeredRealpaths = new Set<string>();
+      for (const path of branchMap.values()) {
+        registeredRealpaths.add(safeRealpath(path));
+      }
       let repaired = 0;
 
       for (const task of allTasks) {
@@ -2823,8 +2839,9 @@ export class SelfHealingManager {
         if (activeSessionRegistry.isPathActive(task.worktree)) continue;
 
         const normalizedBranch = canonicalFusionBranchName(task.id);
-        const canonicalTaskWorktree = resolve(task.worktree);
+        const resolvedTaskWorktree = resolve(task.worktree);
-        const stale = !existsSync(task.worktree) || !registeredPaths.has(canonicalTaskWorktree);
+        const realpathTaskWorktree = safeRealpath(resolvedTaskWorktree);
+        const stale = !existsSync(task.worktree) || !registeredRealpaths.has(realpathTaskWorktree);
         if (!stale) continue;
 
         const previousWorktree = task.worktree;
@@ -2848,6 +2865,25 @@ export class SelfHealingManager {
           continue;
         }
 
+        // FN-5256: never null out worktree/branch metadata for an active task. If a
+        // live task's worktree looks stale here (and we couldn't rebind to a live
+        // fusion/<id>), the executor's own recovery paths will detect and recreate
+        // it. Clearing here yanks the worktree from a still-running shell.
+        if (task.column === "in-progress" || task.column === "in-review") {
+          await this.emitWorktreeMetadataAuditEvent({
+            taskId: task.id,
+            mutationType: "task:auto-recover-worktree-metadata-skipped-active",
+            previousWorktree,
+            newWorktree: previousWorktree,
+            previousBranch,
+            newBranch: previousBranch,
+          });
+          worktreeMetadataReconcileLog.warn(
+            `[FN-5256] skipped clearing worktree metadata for active ${task.column} task ${task.id}: ${previousWorktree} (${previousBranch ?? "<none>"})`,
+          );
+          continue;
+        }
+
         await this.store.updateTask(task.id, { worktree: null, branch: null });
         await this.emitWorktreeMetadataAuditEvent({
           taskId: task.id,

packages/engine/src/worktree-backend.ts

@@ -8,6 +8,7 @@ import {
   activeSessionRegistry,
   reconcileSelfOwnedActiveSessionForRemoval,
   type LiveBindingProbe,
+  type ProcessActiveProbe,
 } from "./active-session-registry.js";
 import type { RunAuditor } from "./run-audit.js";
 import { resolveTaskWorktreePath } from "./worktree-paths.js";
@@ -782,6 +783,8 @@ export async function removeWorktree(input: {
   timeout?: number;
   expectedOwnerTaskId?: string;
   liveOwnerProbe?: LiveBindingProbe;
+  processActiveProbe?: ProcessActiveProbe;
+  reconcileMinIdleMs?: number;
 }): Promise<void> {
   const logger = {
     log: (_message: string): void => {},
@@ -798,6 +801,10 @@ export async function removeWorktree(input: {
       input.worktreePath,
       input.expectedOwnerTaskId,
       input.liveOwnerProbe,
+      {
+        processActiveProbe: input.processActiveProbe,
+        minIdleMs: input.reconcileMinIdleMs,
+      },
     );
     if (reconciled.action === "reconciled") {
       await input.audit?.git({