fix(desktop): recover failed update installs

[leoisadev1]

Packaged Akeru inherited an ambient T3CODE_PORT, which could send the renderer to another local server while its own backend failed to bind. A failed update install could then leave the old desktop process alive after its backends and windows had stopped.

Packaged builds now ignore the development-only port override and use the existing free-port selection. If the updater reports an install failure, the desktop clears its quitting state and restarts every registered backend so the main window can recover. Focused tests cover both failure paths.

Model: gpt-5.6-sol
Harness: Codex harness in T3 Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
akeru-bot-landing	Skipped		Sep 2, 2026 11:50pm UTC

[greptile-apps]

Greptile Summary

This change prevents packaged desktop builds from inheriting the development backend port and restores backends when updater error events interrupt installation. A direct quitAndInstall rejection still leaves previously stopped backends offline, so a user whose update installation fails can be left with a running desktop process but no usable backend or window.

Confidence Score: 4/5

Not safe to merge until the direct installation-rejection handler restores the backends it stopped.

A focused desktop test directly exercised the typed installer-rejection path and observed that backend recovery was never invoked after shutdown.

Files Needing Attention: apps/desktop/src/updates/DesktopUpdates.ts needs recovery logic in the ElectronUpdaterQuitAndInstallError catch path, with a regression test in DesktopUpdates.test.ts.

T-Rex Logs

What T-Rex did

• T-Rex produced proof for the posted P1 finding.
• T-Rex produced proof for a second posted P1 finding.
• T-Rex performed general contract validation on the ElectronUpdaterQuitAndInstallError path and ran the test command, noting that the recovery expectation fails while the current implementation passes (59 files and 583 tests).

_{Ran code and verified through T-Rex}

Comments Outside Diff (2)

1. apps/desktop/src/updates/DesktopUpdates.ts, line 507-520 (link)

   Installer rejection leaves backends stopped

   The installation flow stops all pooled backends before closing windows. When quitAndInstall rejects with ElectronUpdaterQuitAndInstallError, this handler clears the update action and quitting state but does not restart those stopped instances. The process remains alive after the failed installation without a usable backend or window. Restore every instance from pool.list here, matching the recovery already performed for installation-related updater error events.

   Artifacts

   
   

   Focused DesktopUpdates rejection-branch test source

   • The authored focused mock harness extension rejects quitAndInstall after backend shutdown and observes the backend start counter, providing the executable reproduction.

   Recovery expectation failure before the fix

   • Running the focused desktop test with the required restart expectation fails because the observed backend start count is zero, proving the missing recovery.

   Current behavior assertion pass

   • Running the same focused rejection path with the observed zero-start assertion passes all desktop tests, confirming the current behavior.

   

   _{Ran code and verified through T-Rex}

   

2. General comment

   quitAndInstall rejection does not restore stopped backends

   • Bug
     • The install flow stops all pooled backends before closing windows. If the typed ElectronUpdaterQuitAndInstallError is caught, the handler clears the action and quitting flag and reports install failure, but never calls each instance's start; the exercised mock observed zero starts.
   • Cause
     • The dedicated catch at apps/desktop/src/updates/DesktopUpdates.ts:507-520 omits the backend-restart sequence implemented for install-associated updater error events at lines 635-642.
   • Fix
     • After resetInstallAction in the typed catch, obtain pool.list and run instance.start for all instances with unbounded concurrency before returning the install failure state; retain a focused regression test that expects the restart count to be one.

   _{Ran code and verified through T-Rex}

_{Reviews (1): Last reviewed commit: "fix(desktop): recover failed update inst..." | Re-trigger Greptile}