fix(FN-1572): stabilize fusion agent execution

2026-04-17 12:13:39 -07:00
parent 0515882da1
commit ef696f4564
29 changed files with 657 additions and 283 deletions
--- a/packages/cli/CHANGELOG.md
+++ b/packages/cli/CHANGELOG.md
@@ -4,7 +4,7 @@

 ### Minor Changes

- 2d13b82: Add pi extension. Installing `@gsxdsm/fusion` via `pi install` now provides native tools (`kb_task_create`, `kb_task_list`, `kb_task_show`, `kb_task_attach`, `kb_task_pause`, `kb_task_unpause`) and a `/kb` command to start the dashboard and AI engine from within a pi session.
+- 2d13b82: Add pi extension. Installing `@gsxdsm/fusion` via `pi install` now provides native tools (`fn_task_create`, `fn_task_list`, `fn_task_show`, `fn_task_attach`, `fn_task_pause`, `fn_task_unpause`) and a `/fn` command to start the dashboard and AI engine from within a pi session.
 - 494de14: Changed `autoMerge` to default to `true` for new boards.
 - 50821fc: Add global pause button to stop all automated agents and scheduling
 - cac10af: Split engine control into Pause (soft) and Stop (hard). The dashboard Header now shows two buttons: "Pause AI engine" stops new work from being dispatched while letting in-flight agents finish gracefully, and "Stop AI engine" (previously the only Pause button) immediately kills all active agent sessions. A new `enginePaused` setting field controls the soft-pause state alongside the existing `globalPause` hard-stop.
--- a/packages/cli/skill/fusion/references/extension-tools.md
+++ b/packages/cli/skill/fusion/references/extension-tools.md
@@ -11,7 +11,7 @@ Create a new task on the Fusion board. Enters triage for AI specification.
 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
 | `description` | string | ✓ | What needs to be done — be descriptive |
-| `depends` | string[] | — | Task IDs this depends on (e.g., ["KB-001"]) |
+| `depends` | string[] | — | Task IDs this depends on (e.g., ["FN-001"]) |

 Returns: task ID, column, dependencies, path

@@ -21,7 +21,7 @@ Update fields on an existing task (title, description, dependencies).

 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
-| `id` | string | ✓ | Task ID (e.g., KB-001) |
+| `id` | string | ✓ | Task ID (e.g., FN-001) |
 | `title` | string | — | New task title |
 | `description` | string | — | New task description |
 | `depends` | string[] | — | New dependency list — replaces existing |
@@ -45,7 +45,7 @@ Show full task details including steps, progress, prompt preview, and log.

 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
-| `id` | string | ✓ | Task ID (e.g., KB-001) |
+| `id` | string | ✓ | Task ID (e.g., FN-001) |

 Returns: task details with steps, prompt preview (500 chars), last 5 log entries

@@ -244,7 +244,7 @@ Link a feature to a kb task. Updates feature status to triaged.
 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
 | `featureId` | string | ✓ | Feature ID (e.g., F-001) |
-| `taskId` | string | ✓ | Task ID (e.g., KB-001) |
+| `taskId` | string | ✓ | Task ID (e.g., FN-001) |

 ### fn_agent_stop

--- a/packages/cli/src/extension.ts
+++ b/packages/cli/src/extension.ts
@@ -136,7 +136,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_create",
-    label: "KB: Create Task",
+    label: "fn: Create Task",
    description:
      "Create a new task on the Fusion task board. The task enters the triage column " +
      "where the AI triage agent will specify it into a full prompt with steps, " +
@@ -150,7 +150,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      description: Type.String({ description: "What needs to be done — be descriptive" }),
      depends: Type.Optional(
        Type.Array(Type.String(), {
-          description: "Task IDs this depends on (e.g. ['KB-001', 'KB-002'])",
+          description: "Task IDs this depends on (e.g. ['FN-001', 'FN-002'])",
        }),
      ),
      agentId: Type.Optional(
@@ -203,7 +203,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_update",
-    label: "KB: Update Task",
+    label: "fn: Update Task",
    description:
      "Update fields on an existing task. Supports modifying the title, " +
      "description, dependencies, and assigned agent after task creation.",
@@ -213,12 +213,12 @@ export default function kbExtension(pi: ExtensionAPI) {
      "At least one field must be provided to update.",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID (e.g. FN-001)" }),
      title: Type.Optional(Type.String({ description: "New task title" })),
      description: Type.Optional(Type.String({ description: "New task description" })),
      depends: Type.Optional(
        Type.Array(Type.String(), {
-          description: "New dependency list — replaces existing dependencies (e.g. ['KB-001', 'KB-002'])",
+          description: "New dependency list — replaces existing dependencies (e.g. ['FN-001', 'FN-002'])",
        }),
      ),
      agentId: Type.Optional(
@@ -293,7 +293,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_list",
-    label: "KB: List Tasks",
+    label: "fn: List Tasks",
    description: "List all tasks on the Fusion board, grouped by column.",
    promptSnippet: "List all tasks on the Fusion board grouped by column",
    parameters: Type.Object({
@@ -351,11 +351,11 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_show",
-    label: "KB: Show Task",
+    label: "fn: Show Task",
    description: "Show full details for a task including steps, progress, and log entries.",
    promptSnippet: "Show full details for a Fusion task",
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID (e.g. FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -430,13 +430,13 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_attach",
-    label: "KB: Attach File",
+    label: "fn: Attach File",
    description:
      "Attach a file to a task. Supports images (png, jpg, gif, webp) and " +
      "text files (txt, log, json, yaml, yml, toml, csv, xml).",
    promptSnippet: "Attach a file to a Fusion task",
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID (e.g. FN-001)" }),
      path: Type.String({ description: "Path to the file to attach" }),
    }),

@@ -481,12 +481,12 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_pause",
-    label: "KB: Pause Task",
+    label: "fn: Pause Task",
    description:
      "Pause a task — stops all automated agent and scheduler interaction for this task.",
    promptSnippet: "Pause a Fusion task (stops automation)",
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID (e.g. FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -504,12 +504,12 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_unpause",
-    label: "KB: Unpause Task",
+    label: "fn: Unpause Task",
    description:
      "Unpause a task — resumes automated agent and scheduler interaction.",
    promptSnippet: "Unpause a Fusion task (resumes automation)",
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID (e.g. FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -527,7 +527,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_retry",
-    label: "KB: Retry Task",
+    label: "fn: Retry Task",
    description:
      "Retry a failed task — clears the error state and moves it back to the todo column for re-execution.",
    promptSnippet: "Retry a failed Fusion task (clears error, moves to todo)",
@@ -537,7 +537,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "The task will be moved to the todo column with error state cleared",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID to retry (e.g. KB-001). Must be in 'failed' state." }),
+      id: Type.String({ description: "Task ID to retry (e.g. FN-001). Must be in 'failed' state." }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -584,7 +584,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_duplicate",
-    label: "KB: Duplicate Task",
+    label: "fn: Duplicate Task",
    description:
      "Duplicate an existing task, creating a fresh copy in triage. " +
      "Copies the title and description but resets all execution state. " +
@@ -596,7 +596,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "Dependencies, attachments, and execution state are NOT copied",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Source task ID to duplicate (e.g. KB-001)" }),
+      id: Type.String({ description: "Source task ID to duplicate (e.g. FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -614,7 +614,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_refine",
-    label: "KB: Refine Task",
+    label: "fn: Refine Task",
    description:
      "Request a refinement of a completed or in-review task. " +
      "Creates a new follow-up task in triage that references the original task as a dependency. " +
@@ -627,7 +627,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "Provide clear feedback about what needs to be refined or improved",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID to refine (e.g. KB-001). Must be in 'done' or 'in-review' column." }),
+      id: Type.String({ description: "Task ID to refine (e.g. FN-001). Must be in 'done' or 'in-review' column." }),
      feedback: Type.String({ 
        description: "Description of what needs to be refined or improved",
        minLength: 1,
@@ -652,7 +652,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_archive",
-    label: "KB: Archive Task",
+    label: "fn: Archive Task",
    description:
      "Archive a done task (move from done → archived). " +
      "Archived tasks are preserved for historical reference but moved out of the main board view.",
@@ -663,7 +663,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "Archived tasks can be unarchived later if needed",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID to archive (e.g. KB-001). Must be in 'done' column." }),
+      id: Type.String({ description: "Task ID to archive (e.g. FN-001). Must be in 'done' column." }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -681,7 +681,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_unarchive",
-    label: "KB: Unarchive Task",
+    label: "fn: Unarchive Task",
    description:
      "Unarchive an archived task (move from archived → done). " +
      "Restores the task to the done column.",
@@ -691,7 +691,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "Only tasks in the 'archived' column can be unarchived",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID to unarchive (e.g. KB-001). Must be in 'archived' column." }),
+      id: Type.String({ description: "Task ID to unarchive (e.g. FN-001). Must be in 'archived' column." }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -709,7 +709,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_delete",
-    label: "KB: Delete Task",
+    label: "fn: Delete Task",
    description:
      "Permanently delete a task from the Fusion board. " +
      "Tasks are deleted immediately and cannot be recovered.",
@@ -720,7 +720,7 @@ export default function kbExtension(pi: ExtensionAPI) {
      "Consider archiving instead of deleting for completed work you may need to reference later",
    ],
    parameters: Type.Object({
-      id: Type.String({ description: "Task ID to delete (e.g. KB-001)" }),
+      id: Type.String({ description: "Task ID to delete (e.g. FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -738,7 +738,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_import_github",
-    label: "KB: Import GitHub Issues",
+    label: "fn: Import GitHub Issues",
    description:
      "Import GitHub issues as Fusion tasks. Fetches open issues from a repository " +
      "and creates tasks in the triage column. Each task includes the issue title " +
@@ -829,7 +829,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_import_github_issue",
-    label: "KB: Import GitHub Issue",
+    label: "fn: Import GitHub Issue",
    description:
      "Import a specific GitHub issue as a Fusion task. Fetches the issue by number " +
      "and creates a single task in the triage column with the issue title and body.",
@@ -910,7 +910,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_browse_github_issues",
-    label: "KB: Browse GitHub Issues",
+    label: "fn: Browse GitHub Issues",
    description:
      "List open GitHub issues from a repository to browse before importing. " +
      "Returns issue numbers, titles, and URLs for selection. Use with fn_task_import_github_issue " +
@@ -1002,7 +1002,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_task_plan",
-    label: "KB: Plan Task",
+    label: "fn: Plan Task",
    description:
      "Create a task via AI-guided planning mode — interactive conversation to refine your idea into a well-specified task.",
    promptSnippet: "Create a task via AI-guided planning mode",
@@ -1050,8 +1050,8 @@ export default function kbExtension(pi: ExtensionAPI) {
      }

      // Parse created task ID from logs
-      const createdMatch = logs.find((l) => l.match(/Created (KB-\d+):/));
-      const taskId = createdMatch ? createdMatch.match(/Created (KB-\d+):/)?.[1] : undefined;
+      const createdMatch = logs.find((l) => l.match(/Created (FN-\d+):/));
+      const taskId = createdMatch ? createdMatch.match(/Created (FN-\d+):/)?.[1] : undefined;

      // Get summary line
      const summaryLine = logs.find((l) => l.includes("✓ Created")) || "Task created";
@@ -1075,7 +1075,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_mission_create",
-    label: "KB: Create Mission",
+    label: "fn: Create Mission",
    description:
      "Create a new mission — a high-level objective that can span multiple milestones. " +
      "Missions contain milestones that break down work into phases.",
@@ -1131,7 +1131,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_mission_list",
-    label: "KB: List Missions",
+    label: "fn: List Missions",
    description: "List all missions with their current status.",
    promptSnippet: "List all missions",
    promptGuidelines: [
@@ -1185,7 +1185,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_mission_show",
-    label: "KB: Show Mission",
+    label: "fn: Show Mission",
    description: "Show mission details with full hierarchy: milestones → slices → features.",
    promptSnippet: "Show mission details with hierarchy",
    promptGuidelines: [
@@ -1250,7 +1250,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_mission_delete",
-    label: "KB: Delete Mission",
+    label: "fn: Delete Mission",
    description: "Delete a mission and all its milestones, slices, and features. Cannot be undone.",
    promptSnippet: "Delete a mission and all its contents",
    promptGuidelines: [
@@ -1288,7 +1288,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_milestone_add",
-    label: "KB: Add Milestone",
+    label: "fn: Add Milestone",
    description: "Add a milestone to a mission. Milestones represent phases of work.",
    promptSnippet: "Add a milestone to a mission",
    promptGuidelines: [
@@ -1332,7 +1332,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_slice_add",
-    label: "KB: Add Slice",
+    label: "fn: Add Slice",
    description: "Add a slice to a milestone. Slices are work units that can be activated for implementation.",
    promptSnippet: "Add a work slice to a milestone",
    promptGuidelines: [
@@ -1377,7 +1377,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_feature_add",
-    label: "KB: Add Feature",
+    label: "fn: Add Feature",
    description: "Add a feature to a slice. Features are deliverables that can be linked to tasks.",
    promptSnippet: "Add a feature to a slice",
    promptGuidelines: [
@@ -1426,7 +1426,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_slice_activate",
-    label: "KB: Activate Slice",
+    label: "fn: Activate Slice",
    description:
      "Activate a pending slice for implementation. " +
      "Sets status to 'active' and enables task linking for its features.",
@@ -1479,7 +1479,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_feature_link_task",
-    label: "KB: Link Feature to Task",
+    label: "fn: Link Feature to Task",
    description:
      "Link a feature to a fn task for implementation. " +
      "Updates the feature status to 'triaged' and associates it with the task.",
@@ -1492,7 +1492,7 @@ export default function kbExtension(pi: ExtensionAPI) {
    ],
    parameters: Type.Object({
      featureId: Type.String({ description: "Feature ID to link (e.g., F-001)" }),
-      taskId: Type.String({ description: "Task ID to link to (e.g., KB-001)" }),
+      taskId: Type.String({ description: "Task ID to link to (e.g., FN-001)" }),
    }),

    async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
@@ -1538,7 +1538,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_agent_stop",
-    label: "KB: Stop Agent",
+    label: "fn: Stop Agent",
    description:
      "Stop a running agent — pauses its execution. " +
      "Transitions the agent from running/active to paused state.",
@@ -1601,7 +1601,7 @@ export default function kbExtension(pi: ExtensionAPI) {

  pi.registerTool({
    name: "fn_agent_start",
-    label: "KB: Start Agent",
+    label: "fn: Start Agent",
    description:
      "Start a stopped agent — resumes its execution. " +
      "Transitions the agent from paused to active state.",
--- a/packages/core/src/index.ts
+++ b/packages/core/src/index.ts
@@ -69,6 +69,8 @@ export { AUTOMATION_PRESETS, MAX_RUN_HISTORY } from "./automation.js";
 export type { ScheduleType, ScheduledTask, ScheduledTaskCreateInput, ScheduledTaskUpdateInput, AutomationRunResult, AutomationStepType, AutomationStep, AutomationStepResult } from "./automation.js";
 export { AutomationStore } from "./automation-store.js";
 export type { AutomationStoreEvents } from "./automation-store.js";
+export { runCommandAsync } from "./run-command.js";
+export type { RunCommandOptions, RunCommandResult } from "./run-command.js";

 // ── Routine System ───────────────────────────────────────────────────
 export {
--- a/packages/core/src/memory-backend.test.ts
+++ b/packages/core/src/memory-backend.test.ts
@@ -27,6 +27,7 @@ import {
  ensureQmdInstalled,
  qmdMemoryCollectionName,
  QMD_REFRESH_INTERVAL_MS,
+  shouldSkipBackgroundQmdRefresh,
 } from "./memory-backend.js";
 import type { MemoryBackend } from "./memory-backend.js";

@@ -481,6 +482,10 @@ describe("memory-backend", () => {
        }
      });

+      it("skips background qmd refreshes under Vitest by default", () => {
+        expect(shouldSkipBackgroundQmdRefresh()).toBe(true);
+      });
+
      it("uses the OpenClaw qmd package install command", () => {
        expect(QMD_INSTALL_COMMAND).toBe("bun install -g @tobilu/qmd");
      });
--- a/packages/core/src/memory-backend.ts
+++ b/packages/core/src/memory-backend.ts
@@ -36,6 +36,11 @@ type ExecFileAsync = (
 const qmdRefreshState = new Map<string, { lastStartedAt: number; inFlight?: Promise<void> }>();
 let qmdInstallPromise: Promise<boolean> | null = null;

+export function shouldSkipBackgroundQmdRefresh(): boolean {
+  return (process.env.VITEST === "true" || process.env.NODE_ENV === "test")
+    && process.env.FUSION_ENABLE_QMD_REFRESH_IN_TESTS !== "1";
+}
+
 // ── Type Definitions ────────────────────────────────────────────────

 /**
@@ -966,6 +971,10 @@ export async function refreshQmdProjectMemoryIndex(
 }

 export function scheduleQmdProjectMemoryRefresh(rootDir: string): void {
+  if (shouldSkipBackgroundQmdRefresh()) {
+    return;
+  }
+
  void refreshQmdProjectMemoryIndex(rootDir).catch(() => {
    // qmd is optional. Search falls back to local file scanning when refresh fails.
  });
@@ -1029,6 +1038,10 @@ export async function ensureQmdInstalledAndRefresh(rootDir: string): Promise<voi
 }

 export function scheduleQmdInstallAndRefresh(rootDir: string): void {
+  if (shouldSkipBackgroundQmdRefresh()) {
+    return;
+  }
+
  void ensureQmdInstalledAndRefresh(rootDir).catch(() => {
    // qmd remains optional at runtime. Search falls back to local file scanning.
  });
--- a/packages/core/src/run-command.test.ts
+++ b/packages/core/src/run-command.test.ts
@@ -0,0 +1,46 @@
+import { describe, expect, it } from "vitest";
+import { runCommandAsync } from "./run-command.js";
+
+function isProcessAlive(pid: number): boolean {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+describe("runCommandAsync", () => {
+  it("terminates background children left in the command process group", async () => {
+    if (process.platform === "win32") {
+      return;
+    }
+
+    const childScript = "setInterval(() => {}, 1000)";
+    const parentScript = [
+      "const { spawn } = require('node:child_process');",
+      `const child = spawn(process.execPath, ['-e', ${JSON.stringify(childScript)}], { stdio: 'ignore' });`,
+      "console.log(child.pid);",
+      "child.unref();",
+    ].join(" ");
+
+    const result = await runCommandAsync(
+      `${process.execPath} -e ${JSON.stringify(parentScript)}`,
+      { timeoutMs: 5_000 },
+    );
+
+    expect(result.exitCode).toBe(0);
+    const leakedPid = Number.parseInt(result.stdout.trim(), 10);
+    expect(Number.isFinite(leakedPid)).toBe(true);
+
+    for (let i = 0; i < 10 && isProcessAlive(leakedPid); i++) {
+      await sleep(100);
+    }
+
+    expect(isProcessAlive(leakedPid)).toBe(false);
+  });
+});
--- a/packages/core/src/run-command.ts
+++ b/packages/core/src/run-command.ts
@@ -23,6 +23,8 @@ export interface RunCommandResult {
 }

 const DEFAULT_MAX_BUFFER = 10 * 1024 * 1024;
+const FORCE_KILL_DELAY_MS = 5_000;
+const NORMAL_CLEANUP_FORCE_KILL_DELAY_MS = 500;

 /**
 * Run a shell command without blocking the Node.js event loop.
@@ -45,14 +47,38 @@ export function runCommandAsync(
    let stderr = "";
    let bufferExceeded = false;
    let timedOut = false;
+    let forceKillTimer: NodeJS.Timeout | null = null;
+    const useProcessGroup = process.platform !== "win32";

    const child = spawn(command, {
      cwd: options.cwd,
      env: options.env,
+      detached: useProcessGroup,
      shell: true,
      stdio: ["ignore", "pipe", "pipe"],
    });

+    const signalProcessGroup = (signal: NodeJS.Signals): void => {
+      if (!child.pid) return;
+      try {
+        if (useProcessGroup) {
+          process.kill(-child.pid, signal);
+        } else {
+          child.kill(signal);
+        }
+      } catch {
+        // The command may already have exited and cleaned up its process group.
+      }
+    };
+
+    const scheduleForceKill = (delayMs: number): void => {
+      if (forceKillTimer) return;
+      forceKillTimer = setTimeout(() => {
+        signalProcessGroup("SIGKILL");
+      }, delayMs);
+      forceKillTimer.unref();
+    };
+
    const append = (current: string, chunk: Buffer): string => {
      const s = chunk.toString("utf-8");
      if (current.length + s.length > maxBuffer) {
@@ -73,17 +99,17 @@ export function runCommandAsync(
    const timer = options.timeoutMs
      ? setTimeout(() => {
          timedOut = true;
-          child.kill("SIGTERM");
-          setTimeout(() => {
-            if (child.exitCode === null && child.signalCode === null) {
-              child.kill("SIGKILL");
-            }
-          }, 5_000).unref();
+          signalProcessGroup("SIGTERM");
+          scheduleForceKill(FORCE_KILL_DELAY_MS);
        }, options.timeoutMs)
      : null;

    child.on("error", (err) => {
      if (timer) clearTimeout(timer);
+      if (forceKillTimer) {
+        clearTimeout(forceKillTimer);
+        forceKillTimer = null;
+      }
      resolve({
        stdout,
        stderr,
@@ -97,6 +123,16 @@ export function runCommandAsync(

    child.on("close", (code, signal) => {
      if (timer) clearTimeout(timer);
+      if (forceKillTimer) {
+        clearTimeout(forceKillTimer);
+        forceKillTimer = null;
+      }
+      // A shell command can exit successfully while leaving background children
+      // in its process group (for example test runners, qmd indexers, or dev
+      // servers launched with `&`). Clean the group after every run so Fusion
+      // agents do not leak processes beyond the command lifecycle.
+      signalProcessGroup("SIGTERM");
+      scheduleForceKill(NORMAL_CLEANUP_FORCE_KILL_DELAY_MS);
      resolve({
        stdout,
        stderr,
--- a/packages/dashboard/app/components/ModelSelectorTab.tsx
+++ b/packages/dashboard/app/components/ModelSelectorTab.tsx
@@ -237,26 +237,19 @@ export function ModelSelectorTab({ task, addToast, onTaskUpdated, settings }: Mo
      setSavingTarget(target);

      try {
-        const updatedTask = await updateTask(requestTaskId, {
-          modelProvider: target === "executor"
-            ? nextSelection.provider ?? null
-            : previousSavedExecutor.provider ?? null,
-          modelId: target === "executor"
-            ? nextSelection.modelId ?? null
-            : previousSavedExecutor.modelId ?? null,
-          validatorModelProvider: target === "validator"
-            ? nextSelection.provider ?? null
-            : previousSavedValidator.provider ?? null,
-          validatorModelId: target === "validator"
-            ? nextSelection.modelId ?? null
-            : previousSavedValidator.modelId ?? null,
-          planningModelProvider: target === "planning"
-            ? nextSelection.provider ?? null
-            : previousSavedPlanning.provider ?? null,
-          planningModelId: target === "planning"
-            ? nextSelection.modelId ?? null
-            : previousSavedPlanning.modelId ?? null,
-        });
+        const updates: Parameters<typeof updateTask>[1] = {};
+        if (target === "executor") {
+          updates.modelProvider = nextSelection.provider ?? null;
+          updates.modelId = nextSelection.modelId ?? null;
+        } else if (target === "validator") {
+          updates.validatorModelProvider = nextSelection.provider ?? null;
+          updates.validatorModelId = nextSelection.modelId ?? null;
+        } else {
+          updates.planningModelProvider = nextSelection.provider ?? null;
+          updates.planningModelId = nextSelection.modelId ?? null;
+        }
+
+        const updatedTask = await updateTask(requestTaskId, updates);

        if (activeTaskIdRef.current !== requestTaskId) {
          return;
--- a/packages/dashboard/app/components/TaskDetailModal.tsx
+++ b/packages/dashboard/app/components/TaskDetailModal.tsx
@@ -193,6 +193,19 @@ function truncate(s: string, max: number): string {
  return s.length > max ? s.slice(0, max) + "…" : s;
 }

+function sameStringArray(a: string[] = [], b: string[] = []): boolean {
+  return a.length === b.length && a.every((value, index) => value === b[index]);
+}
+
+function splitModelSelection(value: string): { provider: string; modelId: string } | null {
+  const slashIdx = value.indexOf("/");
+  if (!value || slashIdx === -1) return null;
+  return {
+    provider: value.slice(0, slashIdx),
+    modelId: value.slice(slashIdx + 1),
+  };
+}
+
 const DESCRIPTION_TRUNCATE_LENGTH = 200;

 const EDITABLE_COLUMNS: Set<Column> = new Set(["triage", "todo"]);
@@ -529,26 +542,54 @@ export function TaskDetailModal({
  const handleSave = useCallback(async () => {
    setIsSaving(true);
    try {
-      // Build update payload with all changed fields
-      const executorSlashIdx = editExecutorModel.indexOf("/");
-      const validatorSlashIdx = editValidatorModel.indexOf("/");
-      const planningSlashIdx = editPlanningModel.indexOf("/");
+      const updates: Parameters<typeof updateTask>[1] = {};
+      const trimmedTitle = editTitle.trim();
+      const trimmedDescription = editDescription.trim();

-      const updates: Parameters<typeof updateTask>[1] = {
-        title: editTitle.trim() || undefined,
-        description: editDescription.trim() || undefined,
-        dependencies: editDependencies,
-        enabledWorkflowSteps: editSelectedWorkflowSteps,
-        modelProvider: editExecutorModel && executorSlashIdx !== -1 ? editExecutorModel.slice(0, executorSlashIdx) : null,
-        modelId: editExecutorModel && executorSlashIdx !== -1 ? editExecutorModel.slice(executorSlashIdx + 1) : null,
-        validatorModelProvider: editValidatorModel && validatorSlashIdx !== -1 ? editValidatorModel.slice(0, validatorSlashIdx) : null,
-        validatorModelId: editValidatorModel && validatorSlashIdx !== -1 ? editValidatorModel.slice(validatorSlashIdx + 1) : null,
-        planningModelProvider: editPlanningModel && planningSlashIdx !== -1 ? editPlanningModel.slice(0, planningSlashIdx) : null,
-        planningModelId: editPlanningModel && planningSlashIdx !== -1 ? editPlanningModel.slice(planningSlashIdx + 1) : null,
-        thinkingLevel: editThinkingLevel !== "" ? (editThinkingLevel as "minimal" | "low" | "medium" | "high") : null,
-      };
+      if (trimmedTitle && trimmedTitle !== (task.title ?? "")) {
+        updates.title = trimmedTitle;
+      }
+      if (trimmedDescription && trimmedDescription !== (task.description ?? "")) {
+        updates.description = trimmedDescription;
+      }
+      if (!sameStringArray(editDependencies, task.dependencies ?? [])) {
+        updates.dependencies = editDependencies;
+      }
+      if (!sameStringArray(editSelectedWorkflowSteps, task.enabledWorkflowSteps ?? [])) {
+        updates.enabledWorkflowSteps = editSelectedWorkflowSteps;
+      }

-      await updateTask(task.id, updates, projectId);
+      const executorSelection = splitModelSelection(editExecutorModel);
+      const currentExecutorModel = task.modelProvider && task.modelId ? `${task.modelProvider}/${task.modelId}` : "";
+      if (editExecutorModel !== currentExecutorModel) {
+        updates.modelProvider = executorSelection?.provider ?? null;
+        updates.modelId = executorSelection?.modelId ?? null;
+      }
+
+      const validatorSelection = splitModelSelection(editValidatorModel);
+      const currentValidatorModel = task.validatorModelProvider && task.validatorModelId ? `${task.validatorModelProvider}/${task.validatorModelId}` : "";
+      if (editValidatorModel !== currentValidatorModel) {
+        updates.validatorModelProvider = validatorSelection?.provider ?? null;
+        updates.validatorModelId = validatorSelection?.modelId ?? null;
+      }
+
+      const planningSelection = splitModelSelection(editPlanningModel);
+      const currentPlanningModel = task.planningModelProvider && task.planningModelId ? `${task.planningModelProvider}/${task.planningModelId}` : "";
+      if (editPlanningModel !== currentPlanningModel) {
+        updates.planningModelProvider = planningSelection?.provider ?? null;
+        updates.planningModelId = planningSelection?.modelId ?? null;
+      }
+
+      const currentThinkingLevel = task.thinkingLevel ?? "";
+      if (editThinkingLevel !== currentThinkingLevel) {
+        updates.thinkingLevel = editThinkingLevel !== "" ? (editThinkingLevel as "minimal" | "low" | "medium" | "high") : null;
+      }
+
+      const hasTaskUpdates = Object.keys(updates).length > 0;
+      if (hasTaskUpdates) {
+        const updatedTask = await updateTask(task.id, updates, projectId);
+        onTaskUpdated?.(updatedTask);
+      }

      // Upload pending images as attachments
      if (editPendingImages.length > 0) {
@@ -578,7 +619,7 @@ export function TaskDetailModal({
        setIsSaving(false);
      }
    }
-  }, [task.id, editTitle, editDescription, editDependencies, editExecutorModel, editValidatorModel, editPlanningModel, editThinkingLevel, editSelectedWorkflowSteps, editPendingImages, addToast, projectId]);
+  }, [task, editTitle, editDescription, editDependencies, editExecutorModel, editValidatorModel, editPlanningModel, editThinkingLevel, editSelectedWorkflowSteps, editPendingImages, addToast, projectId, onTaskUpdated]);

  const handleAutoSaveDescription = useCallback(async (description: string) => {
    try {
--- a/packages/dashboard/app/components/tests/TaskDetailModal.test.tsx
+++ b/packages/dashboard/app/components/tests/TaskDetailModal.test.tsx
@@ -4134,6 +4134,9 @@ describe("TaskDetailModal", () => {
      // Enter edit mode
      fireEvent.click(container.querySelector(".modal-edit-btn")!);

+      const titleInput = container.querySelector("#task-form-title") as HTMLInputElement;
+      fireEvent.change(titleInput, { target: { value: "Changed title" } });
+
      // Click Save
      fireEvent.click(screen.getByText("Save"));

@@ -4163,6 +4166,9 @@ describe("TaskDetailModal", () => {
      // Enter edit mode
      fireEvent.click(container.querySelector(".modal-edit-btn")!);

+      const titleInput = container.querySelector("#task-form-title") as HTMLInputElement;
+      fireEvent.change(titleInput, { target: { value: "Changed title" } });
+
      // Click Save
      fireEvent.click(screen.getByText("Save"));

@@ -4196,6 +4202,9 @@ describe("TaskDetailModal", () => {
      // Enter edit mode
      fireEvent.click(container.querySelector(".modal-edit-btn")!);

+      const titleInput = container.querySelector("#task-form-title") as HTMLInputElement;
+      fireEvent.change(titleInput, { target: { value: "Changed title" } });
+
      // Click Save
      fireEvent.click(screen.getByText("Save"));

@@ -4276,7 +4285,7 @@ describe("TaskDetailModal", () => {
      expect(screen.getByText(/Workflow Steps/i)).toBeTruthy();
    });

-    it("save sends all changed fields via updateTask", async () => {
+    it("save sends only changed fields via updateTask", async () => {
      const { updateTask } = await import("../../api");
      const mockUpdate = vi.mocked(updateTask);
      mockUpdate.mockResolvedValueOnce({ id: "FN-001" } as Task);
@@ -4296,16 +4305,16 @@ describe("TaskDetailModal", () => {
      // Enter edit mode
      fireEvent.click(container.querySelector(".modal-edit-btn")!);

+      const descTextarea = container.querySelector("#task-form-description") as HTMLTextAreaElement;
+      fireEvent.change(descTextarea, { target: { value: "Updated desc" } });
+
      // Click Save
      fireEvent.click(screen.getByText("Save"));

      await waitFor(() => {
-        expect(mockUpdate).toHaveBeenCalledWith("FN-001", expect.objectContaining({
-          title: "Test",
-          description: "Desc",
-          dependencies: ["FN-002"],
-          enabledWorkflowSteps: [],
-        }), undefined);
+        expect(mockUpdate).toHaveBeenCalledWith("FN-001", {
+          description: "Updated desc",
+        }, undefined);
      });
    });

@@ -4408,12 +4417,10 @@ describe("TaskDetailModal", () => {
        expect(mockUpdateTask).toHaveBeenNthCalledWith(
          2,
          "FN-001",
-          expect.objectContaining({
-            modelProvider: "anthropic",
-            modelId: "claude-sonnet-4-5",
+          {
            validatorModelProvider: "openai",
            validatorModelId: "gpt-4o",
-          }),
+          },
        );
      });

--- a/packages/dashboard/src/tests/chat-manager.test.ts
+++ b/packages/dashboard/src/tests/chat-manager.test.ts
@@ -303,6 +303,27 @@ describe("ChatManager.sendMessage", () => {
    expect(assistantCall?.[1].content).toBe("Hello world!");
  });

+  it("creates chat agents with the full coding toolset", async () => {
+    let createOptions: any;
+    __setCreateKbAgent(async (options: any) => {
+      createOptions = options;
+      return {
+        session: {
+          prompt: vi.fn().mockResolvedValue(undefined),
+          dispose: vi.fn(),
+          state: {
+            messages: [{ role: "assistant", content: "Done" }],
+          },
+        },
+      };
+    });
+
+    const chatManager = createChatManager();
+    await chatManager.sendMessage("chat-001", "Hello");
+
+    expect(createOptions.tools).toBe("coding");
+  });
+
  it("accumulates thinking output separately from text", async () => {
    let onThinkingCb: ((delta: string) => void) | undefined;
    let onTextCb: ((delta: string) => void) | undefined;
--- a/packages/dashboard/src/chat.ts
+++ b/packages/dashboard/src/chat.ts
@@ -537,7 +537,7 @@ export class ChatManager {
      agentResult = await createKbAgent({
        cwd: this.rootDir,
        systemPrompt,
-        tools: "readonly",
+        tools: "coding",
        ...(effectiveModelProvider && effectiveModelId
          ? {
              defaultProvider: effectiveModelProvider,
--- a/packages/dashboard/src/routes.test.ts
+++ b/packages/dashboard/src/routes.test.ts
@@ -2296,23 +2296,30 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
      dependencies: ["FN-002"],
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
    expect(res.body.dependencies).toEqual(["FN-002"]);
  });

+  it("does not clear model or assignee fields when they are omitted", async () => {
+    (store.updateTask as ReturnType<typeof vi.fn>).mockResolvedValue({ ...FAKE_TASK_DETAIL, title: "New" });
+
+    const res = await REQUEST(buildApp(), "PATCH", "/api/tasks/KB-001", JSON.stringify({ title: "New" }), {
+      "Content-Type": "application/json",
+    });
+
+    expect(res.status).toBe(200);
+    expect(store.updateTask).toHaveBeenCalledWith("KB-001", { title: "New" });
+    expect(store.updateTask).not.toHaveBeenCalledWith(
+      "KB-001",
+      expect.objectContaining({
+        modelProvider: null,
+        modelId: null,
+        assigneeUserId: null,
+      }),
+    );
+  });
+
  it("forwards title and description without dependencies", async () => {
    (store.updateTask as ReturnType<typeof vi.fn>).mockResolvedValue({ ...FAKE_TASK_DETAIL, title: "New" });

@@ -2323,18 +2330,6 @@ describe("PATCH /tasks/:id", () => {
    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
      title: "New",
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2358,19 +2353,10 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
      modelProvider: "anthropic",
      modelId: "claude-sonnet-4-5",
      validatorModelProvider: "openai",
      validatorModelId: "gpt-4o",
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2388,18 +2374,6 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
      assigneeUserId: "requesting-user",
    });
  });
@@ -2442,19 +2416,8 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
      modelProvider: null,
      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2476,19 +2439,8 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
      planningModelProvider: "google",
      planningModelId: "gemini-2.5-pro",
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2530,19 +2482,8 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
      planningModelProvider: null,
      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2560,19 +2501,7 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
      enabledWorkflowSteps: ["browser-verification"],
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
-      thinkingLevel: undefined,
-      assigneeUserId: null,
    });
  });

@@ -2601,19 +2530,7 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
      thinkingLevel: "high",
-      assigneeUserId: null,
    });
  });

@@ -2632,19 +2549,7 @@ describe("PATCH /tasks/:id", () => {

    expect(res.status).toBe(200);
    expect(store.updateTask).toHaveBeenCalledWith("KB-001", {
-      title: undefined,
-      description: undefined,
-      prompt: undefined,
-      dependencies: undefined,
-      enabledWorkflowSteps: undefined,
-      modelProvider: null,
-      modelId: null,
-      validatorModelProvider: null,
-      validatorModelId: null,
-      planningModelProvider: null,
-      planningModelId: null,
      thinkingLevel: null,
-      assigneeUserId: null,
    });
  });

--- a/packages/dashboard/src/routes.ts
+++ b/packages/dashboard/src/routes.ts
@@ -4618,10 +4618,12 @@ export function createApiRoutes(store: TaskStore, options?: ServerOptions): Rout
    try {
      const { store: scopedStore } = await getProjectContext(req);
      const { title, description, prompt, dependencies, enabledWorkflowSteps, modelProvider, modelId, validatorModelProvider, validatorModelId, planningModelProvider, planningModelId, thinkingLevel, assigneeUserId } = req.body;
+      const hasBodyField = (field: string) => Object.prototype.hasOwnProperty.call(req.body, field);

      // Validate model fields are strings or undefined/null
      const validateModelField = (value: unknown, name: string): string | null | undefined => {
-        if (value === undefined || value === null) return null;
+        if (value === undefined) return undefined;
+        if (value === null) return null;
        if (typeof value !== "string") {
          throw new Error(`${name} must be a string`);
        }
@@ -4648,21 +4650,22 @@ export function createApiRoutes(store: TaskStore, options?: ServerOptions): Rout
        }
      }

-      const task = await scopedStore.updateTask(req.params.id, {
-        title,
-        description,
-        prompt,
-        dependencies,
-        enabledWorkflowSteps,
-        modelProvider: validatedModelProvider,
-        modelId: validatedModelId,
-        validatorModelProvider: validatedValidatorModelProvider,
-        validatorModelId: validatedValidatorModelId,
-        planningModelProvider: validatedPlanningModelProvider,
-        planningModelId: validatedPlanningModelId,
-        thinkingLevel: thinkingLevel === null ? null : thinkingLevel,
-        assigneeUserId: validatedAssigneeUserId,
-      });
+      const updates: Parameters<typeof scopedStore.updateTask>[1] = {};
+      if (title !== undefined) updates.title = title;
+      if (description !== undefined) updates.description = description;
+      if (prompt !== undefined) updates.prompt = prompt;
+      if (dependencies !== undefined) updates.dependencies = dependencies;
+      if (enabledWorkflowSteps !== undefined) updates.enabledWorkflowSteps = enabledWorkflowSteps;
+      if (hasBodyField("modelProvider")) updates.modelProvider = validatedModelProvider;
+      if (hasBodyField("modelId")) updates.modelId = validatedModelId;
+      if (hasBodyField("validatorModelProvider")) updates.validatorModelProvider = validatedValidatorModelProvider;
+      if (hasBodyField("validatorModelId")) updates.validatorModelId = validatedValidatorModelId;
+      if (hasBodyField("planningModelProvider")) updates.planningModelProvider = validatedPlanningModelProvider;
+      if (hasBodyField("planningModelId")) updates.planningModelId = validatedPlanningModelId;
+      if (hasBodyField("thinkingLevel")) updates.thinkingLevel = thinkingLevel === null ? null : thinkingLevel;
+      if (hasBodyField("assigneeUserId")) updates.assigneeUserId = validatedAssigneeUserId;
+
+      const task = await scopedStore.updateTask(req.params.id, updates);
      res.json(task);
    } catch (err: unknown) {
      if (err instanceof ApiError) {
--- a/packages/engine/src/executor.test.ts
+++ b/packages/engine/src/executor.test.ts
@@ -64,7 +64,35 @@ vi.mock("./worktree-names.js", async () => {
 // promisify(exec) in executor.ts.
 vi.mock("node:child_process", async () => {
  const { promisify } = await import("node:util");
+  const { EventEmitter } = await import("node:events");
  const execSyncFn = vi.fn();
+  const spawnFn = vi.fn((cmd: string, opts?: any) => {
+    const child = new EventEmitter() as any;
+    child.stdout = new EventEmitter();
+    child.stderr = new EventEmitter();
+    child.pid = 12345;
+    child.exitCode = null;
+    child.signalCode = null;
+    child.kill = vi.fn();
+    queueMicrotask(() => {
+      try {
+        const out = execSyncFn(cmd, opts);
+        const stdout = out === undefined ? "" : out.toString();
+        if (stdout) child.stdout.emit("data", Buffer.from(stdout));
+        child.exitCode = 0;
+        child.emit("close", 0, null);
+      } catch (err) {
+        const error = err as { stdout?: string; stderr?: string; status?: number; code?: number };
+        const stdout = error?.stdout?.toString?.() ?? "";
+        const stderr = error?.stderr?.toString?.() ?? "";
+        if (stdout) child.stdout.emit("data", Buffer.from(stdout));
+        if (stderr) child.stderr.emit("data", Buffer.from(stderr));
+        child.exitCode = error.status ?? error.code ?? 1;
+        child.emit("close", child.exitCode, null);
+      }
+    });
+    return child;
+  });
   
  const execFn: any = vi.fn((cmd: string, opts: any, cb: any) => {
    const callback = typeof opts === "function" ? opts : cb;
@@ -95,7 +123,7 @@ vi.mock("node:child_process", async () => {
        }
      });
    });
-  return { execSync: execSyncFn, exec: execFn };
+  return { execSync: execSyncFn, exec: execFn, spawn: spawnFn };
 });
 vi.mock("node:fs", () => ({
  existsSync: vi.fn().mockReturnValue(true),
--- a/packages/engine/src/executor.ts
+++ b/packages/engine/src/executor.ts
@@ -6,7 +6,7 @@ import { isAbsolute, join } from "node:path";
 import { existsSync } from "node:fs";
 import { readFile, writeFile } from "node:fs/promises";
 import type { TaskStore, Task, TaskDetail, StepStatus, Settings, WorkflowStep, MissionStore, Slice, AgentState, AgentCapability, RunMutationContext } from "@fusion/core";
-import { buildExecutionMemoryInstructions, getTaskMergeBlocker, resolveAgentPrompt } from "@fusion/core";
+import { buildExecutionMemoryInstructions, getTaskMergeBlocker, resolveAgentPrompt, runCommandAsync, type RunCommandResult } from "@fusion/core";
 import { findWorktreeUser } from "./merger.js";
 import { generateWorktreeName, slugify } from "./worktree-names.js";
 import { Type, type Static } from "@mariozechner/pi-ai";
@@ -78,6 +78,31 @@ function truncateWorkflowScriptOutput(output: string): string {
  return `... output truncated to last ${WORKFLOW_SCRIPT_OUTPUT_MAX_CHARS} characters ...\n${output.slice(-WORKFLOW_SCRIPT_OUTPUT_MAX_CHARS)}`;
 }

+function configuredCommandErrorMessage(result: RunCommandResult): string {
+  if (result.spawnError) return result.spawnError.message;
+  const parts: string[] = [];
+  if (result.timedOut) parts.push("Timed out");
+  if (result.exitCode !== null) parts.push(`Exit code: ${result.exitCode}`);
+  if (result.signal) parts.push(`Signal: ${result.signal}`);
+  const stdout = result.stdout.trim();
+  const stderr = result.stderr.trim();
+  if (stdout) parts.push(`stdout: ${truncateWorkflowScriptOutput(stdout)}`);
+  if (stderr) parts.push(`stderr: ${truncateWorkflowScriptOutput(stderr)}`);
+  return parts.length ? parts.join("\n") : "Command failed";
+}
+
+async function runConfiguredCommand(
+  command: string,
+  cwd: string,
+  timeoutMs: number,
+): Promise<RunCommandResult> {
+  return runCommandAsync(command, {
+    cwd,
+    timeoutMs,
+    maxBuffer: 10 * 1024 * 1024,
+  });
+}
+
 // ── Tool parameter schemas (module-level for reuse in ToolDefinition generics) ──

 const taskUpdateParams = Type.Object({
@@ -1072,10 +1097,10 @@ export class TaskExecutor {
          // while the user-configured command (e.g. `pnpm install`) executes.
          if (settings.worktreeInitCommand) {
            try {
-              await execAsync(settings.worktreeInitCommand, {
-                cwd: worktreePath,
-                timeout: 120_000,
-              });
+              const initResult = await runConfiguredCommand(settings.worktreeInitCommand, worktreePath, 120_000);
+              if (initResult.spawnError || initResult.timedOut || initResult.exitCode !== 0) {
+                throw new Error(configuredCommandErrorMessage(initResult));
+              }
              await this.store.logEntry(task.id, "Worktree init command completed", settings.worktreeInitCommand, this.currentRunContext);
            } catch (err: unknown) {
              const execError = err instanceof Error ? err : new Error(String(err));
@@ -1091,10 +1116,10 @@ export class TaskExecutor {
            const scriptCommand = settings.scripts?.[settings.setupScript];
            if (scriptCommand) {
              try {
-                await execAsync(scriptCommand, {
-                  cwd: worktreePath,
-                  timeout: 120_000,
-                });
+                const setupResult = await runConfiguredCommand(scriptCommand, worktreePath, 120_000);
+                if (setupResult.spawnError || setupResult.timedOut || setupResult.exitCode !== 0) {
+                  throw new Error(configuredCommandErrorMessage(setupResult));
+                }
                await this.store.logEntry(task.id, `Setup script '${settings.setupScript}' completed`, scriptCommand, this.currentRunContext);
              } catch (err: unknown) {
                const execError = err instanceof Error ? err : new Error(String(err));
@@ -3126,12 +3151,10 @@ ${failureFeedback}
    await this.store.logEntry(task.id, `Workflow step '${workflowStep.name}' executing script '${scriptName}': ${scriptCommand}`);

    try {
-      // Non-blocking: async exec so the executor event loop keeps running
-      // while the user-configured workflow script executes.
-      await execAsync(scriptCommand, {
-        cwd: worktreePath,
-        timeout: 120_000,
-      });
+      const scriptResult = await runConfiguredCommand(scriptCommand, worktreePath, 120_000);
+      if (scriptResult.spawnError || scriptResult.timedOut || scriptResult.exitCode !== 0) {
+        return { success: false, error: configuredCommandErrorMessage(scriptResult) };
+      }
      return { success: true, output: `Script '${scriptName}' completed successfully` };
    } catch (err: unknown) {
      const execError = err instanceof Error ? err : new Error(String(err));
--- a/packages/engine/src/merger.test.ts
+++ b/packages/engine/src/merger.test.ts
@@ -19,7 +19,35 @@ vi.mock("./pi.js", () => ({
 // resolves/rejects based on the callback wired here.
 vi.mock("node:child_process", async () => {
  const { promisify } = await import("node:util");
+  const { EventEmitter } = await import("node:events");
  const execSyncFn = vi.fn();
+  const spawnFn = vi.fn((cmd: string, opts?: any) => {
+    const child = new EventEmitter() as any;
+    child.stdout = new EventEmitter();
+    child.stderr = new EventEmitter();
+    child.pid = 12345;
+    child.exitCode = null;
+    child.signalCode = null;
+    child.kill = vi.fn();
+    queueMicrotask(() => {
+      try {
+        const out = execSyncFn(cmd, opts);
+        const stdout = out === undefined ? "" : out.toString();
+        if (stdout) child.stdout.emit("data", Buffer.from(stdout));
+        child.exitCode = 0;
+        child.emit("close", 0, null);
+      } catch (err) {
+        const error = err as { stdout?: string; stderr?: string; status?: number; code?: number };
+        const stdout = error?.stdout?.toString?.() ?? "";
+        const stderr = error?.stderr?.toString?.() ?? "";
+        if (stdout) child.stdout.emit("data", Buffer.from(stdout));
+        if (stderr) child.stderr.emit("data", Buffer.from(stderr));
+        child.exitCode = error.status ?? error.code ?? 1;
+        child.emit("close", child.exitCode, null);
+      }
+    });
+    return child;
+  });
  const execFn: any = vi.fn((cmd: any, opts: any, cb: any) => {
    const callback = typeof opts === "function" ? opts : cb;
    try {
@@ -45,7 +73,7 @@ vi.mock("node:child_process", async () => {
        }
      });
    });
-  return { execSync: execSyncFn, exec: execFn };
+  return { execSync: execSyncFn, exec: execFn, spawn: spawnFn };
 });

 vi.mock("node:fs", () => ({
--- a/packages/engine/src/merger.ts
+++ b/packages/engine/src/merger.ts
@@ -5,7 +5,7 @@ import { promisify } from "node:util";
 const execAsync = promisify(exec);
 import { existsSync } from "node:fs";
 import { join } from "node:path";
-import { getTaskMergeBlocker, type TaskStore, type MergeResult, type MergeDetails, type WorkflowStep, type WorkflowStepResult, type Settings, type AgentPromptsConfig } from "@fusion/core";
+import { getTaskMergeBlocker, runCommandAsync, type TaskStore, type MergeResult, type MergeDetails, type WorkflowStep, type WorkflowStepResult, type Settings, type AgentPromptsConfig } from "@fusion/core";
 import { resolveAgentPrompt } from "@fusion/core";
 import { createKbAgent, describeModel, promptWithFallback, compactSessionContext } from "./pi.js";
 import { buildSessionSkillContext } from "./session-skill-context.js";
@@ -570,20 +570,36 @@ async function runVerificationCommand(
  };

  try {
-    // Execute the command with timeout (non-blocking: uses async exec so the
-    // engine event loop keeps running while the child process executes)
-    const { stdout, stderr } = await execAsync(command, {
+    const commandResult = await runCommandAsync(command, {
      cwd: rootDir,
-      encoding: "utf-8",
+      timeoutMs: 300_000,
      maxBuffer: VERIFICATION_COMMAND_MAX_BUFFER,
-      timeout: 300_000, // 5 minute timeout for verification commands
    });
-    result.stdout = stdout;
-    result.stderr = stderr;
-    result.exitCode = 0;
-    result.success = true;
-    mergerLog.log(`${taskId}: ${type} command succeeded`);
-    await store.logEntry(taskId, `[verification] ${type} command succeeded (exit 0)`);
+    result.stdout = commandResult.stdout;
+    result.stderr = commandResult.stderr;
+    result.exitCode = commandResult.exitCode;
+    result.success = !commandResult.spawnError
+      && !commandResult.timedOut
+      && commandResult.exitCode === 0;
+
+    if (result.success) {
+      const bufferNote = commandResult.bufferExceeded ? ", output exceeded buffer" : "";
+      mergerLog.log(`${taskId}: ${type} command succeeded`);
+      await store.logEntry(taskId, `[verification] ${type} command succeeded (exit 0${bufferNote})`);
+      return result;
+    }
+
+    const failureText = commandResult.spawnError?.message
+      || (commandResult.timedOut ? "Command timed out" : "")
+      || commandResult.stderr
+      || commandResult.stdout
+      || `Command exited with ${commandResult.exitCode ?? commandResult.signal ?? "unknown status"}`;
+    throw Object.assign(new Error(failureText), {
+      stdout: commandResult.stdout,
+      stderr: commandResult.stderr,
+      status: commandResult.exitCode,
+      code: commandResult.timedOut ? "ETIMEDOUT" : undefined,
+    });
  } catch (error: any) {
    result.stdout = error.stdout?.toString() || "";
    result.stderr = error.stderr?.toString() || "";
--- a/packages/engine/src/peer-exchange-service.test.ts
+++ b/packages/engine/src/peer-exchange-service.test.ts
@@ -111,6 +111,17 @@ describe("PeerExchangeService", () => {
      expect(service).toBeDefined();
    });

+    it("should use default sync interval of 120 seconds", () => {
+      mockListNodes.mockResolvedValue([]);
+      const setIntervalSpy = vi.spyOn(globalThis, "setInterval");
+      const service = new PeerExchangeService(mockCentralCore);
+
+      service.start();
+
+      expect(setIntervalSpy).toHaveBeenCalledWith(expect.any(Function), 120_000);
+      service.stop();
+    });
+
    it("should default settingsSyncEnabled to false", async () => {
      const service = new PeerExchangeService(mockCentralCore);
      setupSuccessfulSync();
--- a/packages/engine/src/peer-exchange-service.ts
+++ b/packages/engine/src/peer-exchange-service.ts
@@ -3,7 +3,7 @@ import type { NodeConfig, PeerSyncRequest, PeerSyncResponse } from "@fusion/core
 import { peerExchangeLog } from "./logger.js";

 export interface PeerExchangeServiceOptions {
-  /** Interval between peer sync cycles in milliseconds. Default: 60000 (1 minute) */
+  /** Interval between peer sync cycles in milliseconds. Default: 120000 (2 minutes) */
  syncIntervalMs?: number;
  /** When true, include settings and model auth data in peer sync exchanges. Default: false. */
  settingsSyncEnabled?: boolean;
@@ -70,7 +70,7 @@ export class PeerExchangeService {
   */
  constructor(centralCore: CentralCore, options: PeerExchangeServiceOptions = {}) {
    this.centralCore = centralCore;
-    this.syncIntervalMs = options.syncIntervalMs ?? 60_000; // 1 minute default
+    this.syncIntervalMs = options.syncIntervalMs ?? 120_000; // 2 minute default
    this.settingsSyncEnabled = options.settingsSyncEnabled ?? false;
    this.settingsSyncThrottleMs = options.settingsSyncThrottleMs ?? 300_000; // 5 minutes default
    this.globalSettings = options.globalSettings;
--- a/packages/engine/src/pi-create-kb-agent.test.ts
+++ b/packages/engine/src/pi-create-kb-agent.test.ts
@@ -555,6 +555,53 @@ describe("createKbAgent", () => {
    }));
  });

+  it("falls back during prompt when the primary model rejects temperature settings", async () => {
+    const primaryPrompt = vi.fn().mockRejectedValue(
+      new Error("400 invalid temperature: only 0.6 is allowed for this model"),
+    );
+    const fallbackPrompt = vi.fn().mockResolvedValue(undefined);
+    const primaryDispose = vi.fn();
+
+    createAgentSessionMock
+      .mockResolvedValueOnce({
+        session: {
+          prompt: primaryPrompt,
+          subscribe: vi.fn(),
+          dispose: primaryDispose,
+          setThinkingLevel: vi.fn(),
+        },
+      })
+      .mockResolvedValueOnce({
+        session: {
+          prompt: fallbackPrompt,
+          subscribe: vi.fn(),
+          dispose: vi.fn(),
+          setThinkingLevel: vi.fn(),
+        },
+      });
+
+    const { createKbAgent } = await import("./pi.js");
+
+    const { session } = await createKbAgent({
+      cwd: "/tmp",
+      systemPrompt: "test",
+      tools: "readonly",
+      defaultProvider: "kimi-coding",
+      defaultModelId: "kimi-k2.6-preview",
+      fallbackProvider: "zai",
+      fallbackModelId: "glm-5.1",
+    });
+
+    await (session as any).promptWithFallback("review this spec");
+
+    expect(primaryPrompt).toHaveBeenCalledWith("review this spec");
+    expect(primaryDispose).toHaveBeenCalled();
+    expect(fallbackPrompt).toHaveBeenCalledWith("review this spec");
+    expect(createAgentSessionMock).toHaveBeenNthCalledWith(2, expect.objectContaining({
+      model: { provider: "zai", id: "glm-5.1" },
+    }));
+  });
+
  it("enables auto-compaction to prevent context-window overflow", async () => {
    const { createKbAgent } = await import("./pi.js");

--- a/packages/engine/src/pi.ts
+++ b/packages/engine/src/pi.ts
@@ -381,7 +381,8 @@ function isRetryableModelSelectionError(message: string): boolean {
    || normalized.includes("overloaded")
    || normalized.includes("quota")
    || normalized.includes("capacity")
-    || normalized.includes("temporarily unavailable");
+    || normalized.includes("temporarily unavailable")
+    || normalized.includes("invalid temperature");
 }

 interface PackageManagerSettingsView {
--- a/packages/engine/src/self-healing.test.ts
+++ b/packages/engine/src/self-healing.test.ts
@@ -1120,6 +1120,72 @@ describe("SelfHealingManager", () => {
      managerWithRecovery.stop();
    });

+    it("moves stale in-review tasks with incomplete steps back to todo for retry", async () => {
+      const managerWithRecovery = new SelfHealingManager(store, {
+        rootDir: "/tmp/test-project",
+      });
+      (store.getSettings as ReturnType<typeof vi.fn>).mockResolvedValue({
+        taskStuckTimeoutMs: 1_000,
+      });
+      (store.listTasks as ReturnType<typeof vi.fn>).mockResolvedValue([
+        {
+          id: "FN-1572",
+          column: "in-review",
+          paused: false,
+          status: null,
+          error: null,
+          worktree: "/tmp/test-project/.worktrees/fn-1572",
+          updatedAt: new Date(Date.now() - 5_000).toISOString(),
+          steps: [
+            { name: "Preflight", status: "done" },
+            { name: "Testing & Verification", status: "in-progress" },
+          ],
+          workflowStepResults: [],
+          mergeDetails: undefined,
+          log: [],
+        },
+      ]);
+
+      const result = await managerWithRecovery.recoverStaleIncompleteReviewTasks();
+
+      expect(result).toBe(1);
+      expect(store.logEntry).toHaveBeenCalledWith(
+        "FN-1572",
+        expect.stringContaining("in-review task still had incomplete steps"),
+      );
+      expect(store.moveTask).toHaveBeenCalledWith("FN-1572", "todo");
+
+      managerWithRecovery.stop();
+    });
+
+    it("does not move fresh in-review tasks with incomplete steps", async () => {
+      const managerWithRecovery = new SelfHealingManager(store, {
+        rootDir: "/tmp/test-project",
+      });
+      (store.getSettings as ReturnType<typeof vi.fn>).mockResolvedValue({
+        taskStuckTimeoutMs: 60_000,
+      });
+      (store.listTasks as ReturnType<typeof vi.fn>).mockResolvedValue([
+        {
+          id: "FN-1573",
+          column: "in-review",
+          paused: false,
+          status: null,
+          updatedAt: new Date().toISOString(),
+          steps: [{ name: "Testing", status: "in-progress" }],
+          workflowStepResults: [],
+          log: [],
+        },
+      ]);
+
+      const result = await managerWithRecovery.recoverStaleIncompleteReviewTasks();
+
+      expect(result).toBe(0);
+      expect(store.moveTask).not.toHaveBeenCalled();
+
+      managerWithRecovery.stop();
+    });
+
    it("moves merged in-review tasks to done and clears transient merge state", async () => {
      const managerWithRecovery = new SelfHealingManager(store, {
        rootDir: "/tmp/test-project",
--- a/packages/engine/src/self-healing.ts
+++ b/packages/engine/src/self-healing.ts
@@ -64,6 +64,7 @@ export interface SelfHealingOptions {
 const APPROVED_TRIAGE_RECOVERY_GRACE_MS = 60_000;
 const ORPHANED_EXECUTION_RECOVERY_GRACE_MS = 60_000;
 const ACTIVE_MERGE_STATUSES = new Set(["merging", "merging-pr"]);
+const NON_TERMINAL_STEP_STATUSES = new Set(["pending", "in-progress"]);
 /**
 * Longer grace period for tasks that still have a worktree on disk.
 * This avoids racing with `executor.resumeOrphaned()` which runs on
@@ -140,6 +141,7 @@ export class SelfHealingManager {
  async runStartupRecovery(): Promise<void> {
    await this.recoverNoProgressNoTaskDoneFailures();
    await this.recoverCompletedTasks();
+    await this.recoverStaleIncompleteReviewTasks();
    await this.recoverInterruptedMergingTasks();
    await this.recoverMisclassifiedFailures();
    await this.recoverOrphanedExecutions();
@@ -456,6 +458,7 @@ export class SelfHealingManager {
      this.checkpointWal();
      await this.enforceWorktreeCap();
      await this.recoverCompletedTasks();
+      await this.recoverStaleIncompleteReviewTasks();
      await this.recoverInterruptedMergingTasks();
      await this.recoverMergeableReviewTasks();
      await this.recoverMergedReviewTasks();
@@ -632,6 +635,58 @@ export class SelfHealingManager {
    }
  }

+  /**
+   * Recover tasks that reached `in-review` while a task step was still marked
+   * pending/in-progress. These tasks are not tracked by StuckTaskDetector
+   * anymore because the executor session is gone, and they are not mergeable
+   * because `getTaskMergeBlocker()` correctly blocks incomplete steps.
+   *
+   * Moving them back to `todo` lets the normal scheduler/executor resume the
+   * incomplete step instead of leaving the task stranded in review.
+   */
+  async recoverStaleIncompleteReviewTasks(): Promise<number> {
+    try {
+      const settings = await this.store.getSettings();
+      const timeoutMs = settings.taskStuckTimeoutMs;
+      if (!timeoutMs || timeoutMs <= 0) return 0;
+
+      const now = Date.now();
+      const tasks = await this.store.listTasks({ column: "in-review" });
+      const staleIncomplete = tasks.filter((task) =>
+        task.column === "in-review" &&
+        !task.paused &&
+        !task.status &&
+        task.steps.length > 0 &&
+        task.steps.some((step) => NON_TERMINAL_STEP_STATUSES.has(step.status)) &&
+        now - new Date(task.updatedAt).getTime() >= timeoutMs
+      );
+
+      if (staleIncomplete.length === 0) return 0;
+
+      log.warn(`Found ${staleIncomplete.length} stale in-review task(s) with incomplete steps`);
+
+      let recovered = 0;
+      for (const task of staleIncomplete) {
+        try {
+          await this.store.logEntry(
+            task.id,
+            "Auto-recovered: in-review task still had incomplete steps — moved back to todo for retry",
+          );
+          await this.store.moveTask(task.id, "todo");
+          log.log(`Recovered stale incomplete review task ${task.id}: moved back to todo`);
+          recovered++;
+        } catch (err: unknown) { const errorMessage = err instanceof Error ? err.message : String(err);
+          log.error(`Failed to recover stale incomplete review task ${task.id}: ${errorMessage}`);
+        }
+      }
+
+      return recovered;
+    } catch (err: unknown) { const errorMessage = err instanceof Error ? err.message : String(err);
+      log.error(`Stale incomplete review recovery failed: ${errorMessage}`);
+      return 0;
+    }
+  }
+
  /**
   * Recover stale `in-review` tasks left in a transient merge status.
   *
--- a/packages/engine/src/triage.test.ts
+++ b/packages/engine/src/triage.test.ts
@@ -1386,7 +1386,7 @@ describe("taskCreate tool model inheritance", () => {
  });

  describe("bounded recovery retries for triage", () => {
-    it("marks triage failed when the agent exits without calling review_spec", async () => {
+    it("requeues triage with backoff when the agent exits without calling review_spec", async () => {
      const task = {
        id: "FN-202",
        description: "Test triage task",
@@ -1424,14 +1424,14 @@ describe("taskCreate tool model inheritance", () => {
      await processor.specifyTask(task);

      expect(store.updateTask).toHaveBeenCalledWith("FN-202", expect.objectContaining({
-        status: "failed",
-        error: expect.stringContaining("review_spec was never called"),
-        recoveryRetryCount: null,
-        nextRecoveryAt: null,
+        status: null,
+        error: null,
+        recoveryRetryCount: 1,
+        nextRecoveryAt: expect.any(String),
      }));
      expect(store.logEntry).toHaveBeenCalledWith(
        "FN-202",
-        expect.stringContaining("Specification failed: spec review not approved"),
+        expect.stringContaining("Spec review not approved (review_spec was never called)"),
      );
    });

--- a/packages/engine/src/triage.ts
+++ b/packages/engine/src/triage.ts
@@ -902,19 +902,41 @@ export class TriageProcessor {
          }

          // Post-session APPROVE gate: only advance to todo when the spec
-          // reviewer explicitly approved.  Any other verdict (REVISE,
-          // RETHINK, UNAVAILABLE) or a missing review (null) keeps the task
-          // in triage so unreviewed / rejected specs never reach execution.
+          // reviewer explicitly approved. Any other verdict (REVISE,
+          // RETHINK, UNAVAILABLE) or a missing review (null) stays in triage
+          // and is retried with bounded backoff instead of immediately failing.
          if (specReviewVerdictRef.current !== "APPROVE") {
            const verdictDesc =
              specReviewVerdictRef.current === null
                ? "review_spec was never called"
                : `verdict was ${specReviewVerdictRef.current}`;
+            const decision = computeRecoveryDecision({
+              recoveryRetryCount: task.recoveryRetryCount,
+              nextRecoveryAt: task.nextRecoveryAt,
+            });
+
+            if (decision.shouldRetry) {
+              const attempt = decision.nextState.recoveryRetryCount;
+              const delay = formatDelay(decision.delayMs);
+              const retryMessage =
+                `Spec review not approved (${verdictDesc}) — retry ${attempt}/${MAX_RECOVERY_RETRIES} in ${delay}.`;
+              triageLog.warn(`${task.id} ${retryMessage}`);
+              await this.store.logEntry(task.id, retryMessage);
+              const restoreStatus = task.status === "needs-respecify" ? "needs-respecify" : null;
+              await this.store.updateTask(task.id, {
+                status: restoreStatus,
+                error: null,
+                recoveryRetryCount: decision.nextState.recoveryRetryCount,
+                nextRecoveryAt: decision.nextState.nextRecoveryAt,
+              });
+              return;
+            }
+
            const failureMessage =
-              `Specification failed: spec review not approved (${verdictDesc}). ` +
+              `Specification failed after ${MAX_RECOVERY_RETRIES} unapproved spec reviews (${verdictDesc}). ` +
              "Retry after adjusting the task prompt or model.";
            triageLog.log(
-              `${task.id} spec review not approved (${verdictDesc}) — marking specification failed`,
+              `${task.id} spec review not approved (${verdictDesc}) — retry budget exhausted`,
            );
            await this.store.logEntry(
              task.id,