Windows MSIX: failed auto-update while app is hung uninstalls Claude Desktop (RemoveForAllUsers) - app disappears, daily reinstall required

Claude Desktop (Windows MSIX): failed auto-update while app is hung silently uninstalls the package — app "disappears" and must be reinstalled daily

Environment

• Claude Desktop 1.24012.9 (MSIX, package Claude_pzs8sxrjxfjjc), bundled Claude Code 2.1.219
• Windows 10 Pro 19045, 36 GB RAM
• GPU: AMD Radeon(TM) integrated graphics, driver 31.0.21924.61 (Dec 2025)

Symptom (user-visible)

Shortly after starting a Claude Code session in the desktop app, the app freezes and disappears with no error dialog. Afterwards the app cannot be launched at all — it is no longer installed — and has to be reinstalled from the .msix. This has happened daily (July 25, 27, 28).

Root-cause chain (from logs on the affected machine)

Two independent bugs compound:

1. GPU process crash hangs the whole app

%APPDATA%\Claude\logs\main.log shows 5 occurrences (Jul 25–28) of:

GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }

Each occurs seconds after a Browser/preview pane is created ([Preview] Created browser preview). After the GPU process dies the app does not recover or show any error — main.log goes silent (e.g. today: last entry 09:47:11, nothing until relaunch at 09:54:34) and Windows Error Reporting records MoAppHang (fault bucket 1471911898834861685, hang signature 4185) and terminates the app. gpu-info.json shows Chromium has already blocklisted most features on this GPU (gl=none, skiaBackendType: None, no D3D12/Vulkan).

2. Auto-update during the hang uninstalls the app and doesn't reinstall it

Microsoft-Windows-AppXDeploymentServer/Operational for today (identical pattern Jul 25 and 27):

09:26:39  Add       Claude-3567290728.msix  (ForceApplicationShutdown, ForceUpdateFromAnyVersion)  OK   <- user reinstall
09:27:46  (main.log) GPU process gone -> app hung
09:36:53  Add       Claude-3567290728.msix  FAILED 0x80070020 (sharing violation - hung app still locks files)
09:37:25  Add       retry                   FAILED 0x80070020
09:39:00  Remove    RemoveForAllUsers       OK  -> package moved to WindowsApps\Deleted  ** app is now gone **
09:46:19  Add       Claude-3567290728.msix  OK   <- user reinstalls again
09:47:11  (main.log) GPU process gone -> app hung again
09:49:51  Remove    RemoveForAllUsers       OK  -> ** app gone again ** (finished 09:50:26, same second as WER AppHang)
09:54:22  Add       Claude-3737152193.msix  OK   <- user reinstalls again (current session)

Repair attempts also fail while wedged: RegisterByPackageFullName … RepairAppRegistrationOption → 0x80073D02 (Jul 25 23:04–23:05, Jul 27 09:34 and 14:08).

So the update/installer flow, upon hitting 0x80070020, falls back to a full Remove (RemoveForAllUsers) and then never completes (or delays) the subsequent Add — leaving the user with no installed app and no error message. From the user's perspective the app "crashes and disappears, then cannot be found".

Expected behavior

1. A GPU process crash should not hang the app indefinitely (restart GPU process / fall back to software rendering / show an error).
2. The updater should never leave the machine with the package removed: if the Add fails with 0x80070020, don't Remove the existing package; if a Remove+Add fallback is used, the Remove must be gated on a guaranteed successful staged Add (or rolled back).

Additional notes

• The app was originally a Squirrel install migrated to MSIX ([MSIX] Filesystem not virtualized — … (likely Squirrel upgrade) in main.log).
• Startup also logs EBUSY copying chrome-native-host.exe every launch, consistent with leftover locked files from the hung previous instance.
• The standalone Claude Code CLI (native installer, 2.1.220) on the same machine is unaffected.