fix(desktop): heal launch-mode/shell split-brain that hung local runtime on Windows

Desktop startup used two independent persisted sources of truth that could
disagree: desktop-launch-mode.json decided whether main STARTS the embedded
local runtime, while shell-connections.json (desktopMode) decided whether the
renderer launch gate WAITS for it. shell:setDesktopMode persists shell settings
before the fallible startLocalRuntimeOnce()/saveDesktopLaunchMode(), so a first
"local" selection whose runtime start failed or was interrupted left
shell=local / launch-mode=choose permanently. Every subsequent launch then sat
at "Starting local Fusion runtime…" polling a runtime nobody started, timing out
after 30s.

Fix (defense in depth):
- initializeApp reconciles: a completed shell "local" selection is authoritative;
  it heals the launch-mode file and starts the runtime.
- onDesktopModeChange/onDesktopLaunchModeChange persist launch-mode BEFORE the
  fallible start so an interrupted start cannot re-create the desync.
- DesktopLaunchGate no longer assumes main started the runtime; if it is not
  running/starting it actively (re)starts via setDesktopMode("local") before polling.
- Add env-gated startup trace (FUSION_STARTUP_TRACE) so packaged builds, which
  otherwise log nothing, can diagnose this class of stall.

Regression tests assert the invariant (split-brain -> runtime starts + file heals;
agreement-on-choose -> no start). flushPromises now drains via a macrotask so
run()-based tests observe a fully-initialized app regardless of async chain length.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

packages/dashboard/app/components/DesktopLaunchGate.tsx

@@ -98,6 +98,32 @@ export function DesktopLaunchGate({ children }: PropsWithChildren) {
             return;
           }
           setPhase({ kind: "starting-local", message: t("desktop.startingLocalRuntime", "Starting local Fusion runtime…") });
+          /*
+           * FNXC:DesktopLaunchGate 2026-07-02-14:35:
+           * Self-healing start. Do NOT assume main already started the embedded runtime.
+           * The gate decides to WAIT from shell `desktopMode:"local"`, but main decides to
+           * START from a separate launch-mode file; when those desync (a first local
+           * selection whose runtime start failed/was interrupted), main never starts the
+           * runtime and this branch would poll a permanently "stopped" runtime until the 30s
+           * timeout — the "hangs at Starting local runtime" bug. If the runtime is not already
+           * running or starting, actively (re)start it via setDesktopMode("local") — idempotent
+           * and awaits startup — before polling, so the gate can never wait for a runtime nobody
+           * launched.
+           */
+          const rt = state.localRuntime;
+          if (rt?.state !== "running" && rt?.state !== "starting") {
+            try {
+              await shell.setDesktopMode("local");
+            } catch (startError) {
+              if (cancelled) return;
+              setPhase({
+                kind: "local-error",
+                message: startError instanceof Error ? startError.message : String(startError),
+              });
+              return;
+            }
+            if (cancelled) return;
+          }
           const { baseUrl } = await waitForLocalRuntime(shell);
           if (cancelled) return;
           applyServerBaseUrl(baseUrl);

packages/desktop/src/__tests__/main-integration.test.ts

@@ -231,8 +231,13 @@ async function importMainModule() {
 }
 
 async function flushPromises() {
-  await Promise.resolve();
+  // Drain ALL pending microtasks via a macrotask boundary rather than counting a
-  await Promise.resolve();
+  // fixed number of ticks. initializeApp()'s async chain length is an implementation
+  // detail (e.g. the launch-mode/shell split-brain reconciliation adds a readShellSettings
+  // await); a hardcoded tick count silently under-flushes when that chain grows and the
+  // run()-based tests then observe a half-initialized app. setTimeout(0) waits for the
+  // whole microtask queue to settle, so these tests assert on a fully-initialized app.
+  await new Promise((resolve) => setTimeout(resolve, 0));
 }
 
 describe("main integration", () => {

packages/desktop/src/__tests__/main-local-mode.test.ts

@@ -49,7 +49,19 @@ const mocks = vi.hoisted(() => {
     return localRuntimeManager;
   });
 
-  return { app, appHandlers, BrowserWindow, Tray, browserWindow, localRuntimeManager, LocalRuntimeManager, screen };
+  const DEFAULT_SHELL_SETTINGS = {
+    desktopMode: null as "local" | "remote" | null,
+    hasCompletedModeSelection: false,
+    activeProfileId: null,
+    profiles: [],
+  };
+  const readShellSettings = vi.fn(async () => ({ ...DEFAULT_SHELL_SETTINGS }));
+  const writeShellSettings = vi.fn(async () => undefined);
+
+  const loadDesktopLaunchMode = vi.fn(async () => "choose" as "choose" | "local" | "remote");
+  const saveDesktopLaunchMode = vi.fn(async () => undefined);
+
+  return { app, appHandlers, BrowserWindow, Tray, browserWindow, localRuntimeManager, LocalRuntimeManager, screen, readShellSettings, writeShellSettings, DEFAULT_SHELL_SETTINGS, loadDesktopLaunchMode, saveDesktopLaunchMode };
 });
 
 vi.mock("electron", () => ({
@@ -67,8 +79,8 @@ vi.mock("../ipc.js", () => ({ registerIpcHandlers: vi.fn() }));
 vi.mock("../native.js", () => ({
   DEFAULT_WINDOW_STATE: { width: 1000, height: 800 },
   loadWindowState: vi.fn(async () => null),
-  loadDesktopLaunchMode: vi.fn(async () => "choose"),
+  loadDesktopLaunchMode: mocks.loadDesktopLaunchMode,
-  saveDesktopLaunchMode: vi.fn(async () => undefined),
+  saveDesktopLaunchMode: mocks.saveDesktopLaunchMode,
   saveWindowState: vi.fn(),
   setupAutoUpdater: vi.fn(),
   startUpdateCheckInterval: vi.fn(() => vi.fn()),
@@ -76,12 +88,18 @@ vi.mock("../native.js", () => ({
 }));
 vi.mock("../deep-link.js", () => ({ registerDeepLinkProtocol: vi.fn(), setupDeepLinkHandler: vi.fn() }));
 vi.mock("../local-runtime.js", () => ({ LocalRuntimeManager: mocks.LocalRuntimeManager }));
+vi.mock("../shell-settings.js", () => ({
+  readShellSettings: mocks.readShellSettings,
+  writeShellSettings: mocks.writeShellSettings,
+}));
 
 describe("main local mode", () => {
   beforeEach(() => {
     vi.resetModules();
     vi.clearAllMocks();
     mocks.appHandlers.clear();
+    mocks.readShellSettings.mockResolvedValue({ ...mocks.DEFAULT_SHELL_SETTINGS });
+    mocks.loadDesktopLaunchMode.mockResolvedValue("choose");
     delete process.env.FUSION_DESKTOP_MODE;
   });
 
@@ -95,4 +113,41 @@ describe("main local mode", () => {
     expect(mocks.localRuntimeManager.startLocal).toHaveBeenCalled();
     delete process.env.FUSION_DESKTOP_MODE;
   });
+
+  /*
+   * FNXC:DesktopRuntimeMode 2026-07-02-14:35 — regression for the "hangs at Starting
+   * local Fusion runtime" bug. Split-brain persisted state: the launch-mode file says
+   * "choose" (so main would not start the runtime) while shell-connections.json records a
+   * completed "local" selection (so the renderer gate waits for the runtime). Before the
+   * fix the runtime was never started and the gate polled until a 30s timeout on EVERY
+   * launch. initializeApp must reconcile: treat the completed shell "local" as authoritative,
+   * heal the launch-mode file, and start the runtime.
+   */
+  it("reconciles a launch-mode/shell split-brain and starts the runtime (no FUSION_DESKTOP_MODE)", async () => {
+    mocks.loadDesktopLaunchMode.mockResolvedValue("choose");
+    mocks.readShellSettings.mockResolvedValue({
+      desktopMode: "local",
+      hasCompletedModeSelection: true,
+      activeProfileId: null,
+      profiles: [],
+    });
+
+    const { initializeApp } = await import("../main.ts");
+    await initializeApp();
+
+    // Runtime is started despite launch-mode being "choose" ...
+    expect(mocks.localRuntimeManager.startLocal).toHaveBeenCalled();
+    // ... and the launch-mode file is healed to "local" so both sources agree next launch.
+    expect(mocks.saveDesktopLaunchMode).toHaveBeenCalledWith("local");
+  });
+
+  it("does NOT start the runtime when both sources agree on choose (first run / no selection)", async () => {
+    mocks.loadDesktopLaunchMode.mockResolvedValue("choose");
+    mocks.readShellSettings.mockResolvedValue({ ...mocks.DEFAULT_SHELL_SETTINGS });
+
+    const { initializeApp } = await import("../main.ts");
+    await initializeApp();
+
+    expect(mocks.localRuntimeManager.startLocal).not.toHaveBeenCalled();
+  });
 });

packages/desktop/src/local-runtime.ts

@@ -1,9 +1,30 @@
 import { once } from "node:events";
+import { appendFileSync } from "node:fs";
 import type { Server } from "node:http";
 import type { AddressInfo } from "node:net";
 
 import { ensureDesktopRuntimeProject } from "./engine-runtime.js";
 
+/*
+ * FNXC:DesktopRuntime 2026-07-02-14:35:
+ * Env-gated startup trace. Packaged desktop builds have no file logging, so a stalled
+ * or failed embedded-runtime start is invisible to operators (the symptom is only a
+ * spinner that times out). Setting FUSION_STARTUP_TRACE=<path> appends a timestamped
+ * step-by-step trace of startLocal()/startEmbedded()/createDashboardServer() to that
+ * file — the diagnostic that pinpointed the launch-mode split-brain hang. Zero cost
+ * when unset; keep it so this class of hang is diagnosable in the field.
+ */
+const STARTUP_TRACE_FILE = process.env.FUSION_STARTUP_TRACE;
+const __traceStart = Date.now();
+function strace(msg: string): void {
+  if (!STARTUP_TRACE_FILE) return;
+  try {
+    appendFileSync(STARTUP_TRACE_FILE, `[+${((Date.now() - __traceStart) / 1000).toFixed(2)}s] ${msg}\n`);
+  } catch {
+    // best-effort
+  }
+}
+
 export type RuntimeSource = "embedded-local" | "external-cli" | "none";
 export type RuntimeState = "stopped" | "starting" | "running" | "error";
 
@@ -60,11 +81,17 @@ async function createDashboardServerDefault(store: TaskStoreLike, rootDir: strin
   };
 
   try {
+    strace("createDashboardServer: centralCore.init");
     await centralCore.init();
+    strace("createDashboardServer: ensureDesktopRuntimeProject");
     const rootProject = await ensureDesktopRuntimeProject(centralCore, rootDir);
+    strace(`createDashboardServer: startAll (rootProject=${rootProject.id})`);
     await engineManager.startAll();
+    strace("createDashboardServer: startAll DONE; startReconciliation");
     engineManager.startReconciliation();
+    strace("createDashboardServer: ensureEngine(rootProject)");
     const primaryEngine = await engineManager.ensureEngine(rootProject.id);
+    strace("createDashboardServer: createServer");
     const app = createServer(store as never, {
       engine: primaryEngine,
       engineManager,
@@ -72,11 +99,15 @@ async function createDashboardServerDefault(store: TaskStoreLike, rootDir: strin
       onProjectFirstAccessed: (projectId: string) => engineManager.onProjectAccessed(projectId),
     });
 
+    strace("createDashboardServer: app.listen(0)");
+    const server = app.listen(0);
+    strace("createDashboardServer: returning server object");
     return {
-      server: app.listen(0),
+      server,
       cleanup,
     };
   } catch (error) {
+    strace(`createDashboardServer: THREW ${error instanceof Error ? error.stack : String(error)}`);
     await cleanup();
     throw error;
   }
@@ -148,8 +179,10 @@ export class LocalRuntimeManager {
   }
 
   async startLocal(): Promise<DesktopRuntimeStatus> {
+    strace("startLocal: ENTER");
     const externalPort = this.getExternalPort();
     if (externalPort) {
+      strace(`startLocal: external-cli branch (FUSION_SERVER_PORT=${externalPort}) — NOT starting embedded`);
       this.status = {
         source: "external-cli",
         state: "running",
@@ -183,13 +216,18 @@ export class LocalRuntimeManager {
     let cleanup: RuntimeCleanup | undefined;
 
     try {
+      strace(`startEmbedded: BEGIN rootDir=${this.options.rootDir}`);
       store = await this.createStore(this.options.rootDir);
+      strace("startEmbedded: store.init");
       await store.init();
+      strace("startEmbedded: store.watch");
       await store.watch();
+      strace("startEmbedded: createDashboardServer()");
 
       const dashboardServer = await this.createDashboardServer(store, this.options.rootDir);
       cleanup = "server" in dashboardServer ? dashboardServer.cleanup : undefined;
       server = "server" in dashboardServer ? dashboardServer.server : dashboardServer;
+      strace("startEmbedded: awaiting server 'listening' | 'error'");
       await Promise.race([
         once(server, "listening"),
         once(server, "error").then(([error]) => {
@@ -201,6 +239,7 @@ export class LocalRuntimeManager {
       const baseUrl = `http://127.0.0.1:${port}`;
       this.runtime = { store, server, port, baseUrl, cleanup };
       this.status = { source: "embedded-local", state: "running", port, baseUrl };
+      strace(`startEmbedded: RUNNING port=${port}`);
       return this.status;
     } catch (error) {
       if (server) {
@@ -218,6 +257,7 @@ export class LocalRuntimeManager {
         state: "error",
         error: error instanceof Error ? error.message : String(error),
       };
+      strace(`startEmbedded: CATCH/ERROR ${error instanceof Error ? error.stack : String(error)}`);
       throw error;
     }
   }

packages/desktop/src/main.ts

@@ -181,7 +181,33 @@ export function resolveLocalRuntimeRoot(): string {
 
 export async function initializeApp(): Promise<void> {
   const state = await loadWindowState();
-  const rememberedLaunchMode = await loadDesktopLaunchMode();
+  let rememberedLaunchMode = await loadDesktopLaunchMode();
+
+  /*
+   * FNXC:DesktopRuntimeMode 2026-07-02-14:35:
+   * Split-brain reconciliation. Desktop startup has TWO persisted sources of truth
+   * that must agree: `desktop-launch-mode.json` (loadDesktopLaunchMode) decides whether
+   * THIS function STARTS the embedded local runtime, while `shell-connections.json`
+   * (readShellSettings().desktopMode) decides whether the renderer launch gate WAITS
+   * for it. They desync because shell:setDesktopMode persists shell settings BEFORE the
+   * fallible startLocalRuntimeOnce()/saveDesktopLaunchMode() — so a first local selection
+   * whose runtime start throws or is interrupted leaves shell=`local` but launch-mode=`choose`.
+   * The gate then shows "Starting local Fusion runtime…" and polls forever for a runtime
+   * nobody started, timing out after 30s on EVERY launch. Treat a completed shell "local"
+   * selection as authoritative and heal the launch-mode file so both halves agree.
+   */
+  const reconcileShellSettings = await readShellSettings();
+  if (
+    rememberedLaunchMode !== "local" &&
+    reconcileShellSettings.desktopMode === "local" &&
+    reconcileShellSettings.hasCompletedModeSelection === true
+  ) {
+    console.warn(
+      `[desktop/main] Healing launch-mode split-brain: shell desktopMode="local" but launch-mode="${rememberedLaunchMode}"; adopting "local"`,
+    );
+    rememberedLaunchMode = "local";
+    await saveDesktopLaunchMode("local");
+  }
 
   localRuntimeManager = new LocalRuntimeManager({ rootDir: resolveLocalRuntimeRoot() });
   currentDesktopLaunchMode = rememberedLaunchMode;
@@ -274,13 +300,23 @@ export async function initializeApp(): Promise<void> {
       if (mode === "local") {
         currentRemoteLaunch = null;
         localRuntimeStartupAttempted = false;
+        /*
+         * FNXC:DesktopRuntimeMode 2026-07-02-14:35:
+         * Persist launch-mode BEFORE the fallible startLocalRuntimeOnce(). shell:setDesktopMode
+         * already wrote shell-connections.json `desktopMode:"local"` before invoking this callback;
+         * if the runtime start throws/is interrupted and we saved launch-mode only afterward, the two
+         * files desync (shell=local, launch-mode=choose) and every future launch hangs at "Starting
+         * local runtime". Saving first keeps both sources in agreement so a failed start simply retries
+         * on next launch instead of deadlocking.
+         */
+        await saveDesktopLaunchMode(mode);
         await startLocalRuntimeOnce();
-      } else {
+        return;
-        localRuntimeStartupAttempted = false;
-        await localRuntimeManager.stopLocal();
-        const shellSettings = await readShellSettings();
-        currentRemoteLaunch = normalizeDesktopRemoteLaunch({ ...shellSettings, desktopMode: "remote" });
       }
+      localRuntimeStartupAttempted = false;
+      await localRuntimeManager.stopLocal();
+      const shellSettings = await readShellSettings();
+      currentRemoteLaunch = normalizeDesktopRemoteLaunch({ ...shellSettings, desktopMode: "remote" });
       await saveDesktopLaunchMode(mode);
     },
     onDesktopLaunchModeChange: async (mode) => {
@@ -291,12 +327,14 @@ export async function initializeApp(): Promise<void> {
       localRuntimeStartupAttempted = false;
       if (mode === "local") {
         currentRemoteLaunch = null;
+        // FNXC:DesktopRuntimeMode 2026-07-02-14:35: persist before the fallible start (see onDesktopModeChange).
+        await saveDesktopLaunchMode(mode);
         await startLocalRuntimeOnce();
-      } else {
+        return;
-        await localRuntimeManager.stopLocal();
-        const shellSettings = await readShellSettings();
-        currentRemoteLaunch = normalizeDesktopRemoteLaunch({ ...shellSettings, desktopMode: "remote" });
       }
+      await localRuntimeManager.stopLocal();
+      const shellSettings = await readShellSettings();
+      currentRemoteLaunch = normalizeDesktopRemoteLaunch({ ...shellSettings, desktopMode: "remote" });
       await saveDesktopLaunchMode(mode);
     },
     getRuntimeStatus: () => localRuntimeManager?.getStatus() ?? { source: "none", state: "stopped" },