[BUG] Claude Desktop (Windows/MSIX): main process self-exits with code -1 seconds after launch - survives reinstall, empty profile and reboot

## Summary

Since a crash that happened while previewing a localhost dev server from inside the app, Claude Desktop no longer stays running on this machine. Every launch boots normally — sometimes far enough to render the real UI at claude.ai/epitaxy — and then the main process terminates itself with exit code -1 (0xFFFFFFFF) between 1.5 and 10 seconds after start. The child processes shut down cleanly (exit code 0) after the main process is gone.

This is not a crash: with WER LocalDumps enabled no dump is produced, Chromium verbose logging contains no FATAL/ERROR, Sentry records no exception event, and the app's own main.log shows no exception — the log simply stops mid-line.

Environment

| | |
|---|---|
| OS | Windows 11 Home, 10.0.26200 Build 26200.9168 (x64) |
| App version | reproduced on 1.30096.0.0 and 1.28929.0.0 |
| Install type | MSIX, Claude_pzs8sxrjxfjjc, SignatureKind: Developer |
| Electron/Node | Node 24.18.0 |
| CPU / RAM | Intel Core i7-9700F, 16 GB |
| GPU | NVIDIA GeForce RTX 2070 SUPER, driver 32.0.15.9186 (single GPU, no hybrid graphics) |
| Displays | 3 × 1920×1080 |
| Virtualization | VirtualMachinePlatform enabled; HypervisorPlatform and WSL disabled |
| Antivirus | Windows Defender only; no ASR rules, Controlled Folder Access off, Smart App Control off |

Timeline

| Time (local) | Event |
|---|---|
| 22:25 | App running normally, working in a Claude Code session (project: an Electron+Vite game) |
| ~22:40 | User previews the project's localhost dev server from inside the app → app crashes |
| 22:42 | First renderer error: Electron sandboxed_renderer.bundle.js script failed to run / TypeError: Cannot destructure property 'preloadScripts' of 'binding.startupData' as it is null. |
| 22:42 → | No successful launch since. |
| 23:13 | Machine rebooted — no change |
| 23:14 | App auto-updated 1.28929.0.0 → 1.30096.0.0 — no change |
| 23:53 | Reinstalled from the official installer (downgrade to 1.28929.0.0) — no change |

Key evidence: kernel process trace

Captured via Microsoft-Windows-Kernel-Process/Analytic:

00:21:44.145  START  PID 16216                    <- main process (returned by ActivateApplication)
00:21:44.363  START  PID 1740
00:21:44.821  START  PID 15068
00:21:44.846  START  PID 16240
00:21:45.055  START  PID 1100
00:21:45.065  START  PID 6808
00:21:45.206  START  PID 8764
00:21:45.254  STOP   PID 6808   ExitCode 0
00:21:45.601  STOP   PID 16216  ExitCode 4294967295   <- main process, -1
00:21:45.607  STOP   PID 16240  ExitCode 0
00:21:45.611  STOP   PID 1100   ExitCode 0
00:21:45.616  STOP   PID 15068  ExitCode 0
00:21:45.618  STOP   PID 8764   ExitCode 0
00:21:45.642  STOP   PID 1740   ExitCode 0

The main process exits first, with -1; everything else follows with a clean 0. The Desktop AppX container is created and destroyed normally on every attempt (AppModel-Runtime IDs 210/211/201 then 217) — there is no orphaned silo/job object, so this is a different failure mode from issue #53247.

Where main.log stops

The last logged line varies between attempts, which suggests the exit is asynchronous rather than tied to one code path:

[browser-live-preview] initialized (passive tap)
mainView backgroundThrottling disabled (... boot: in progress)
[watch-record] failed to subscribe to swift events Cannot find package '@ant/claude-swift'
LocalAgentModeSessions.getDefaultWorkspaceFolders: 0 folders
[RemotePluginManager] Sync complete: 0 downloaded, 0 removed, 0 orphans cleaned

In one run with an empty profile the app got all the way to mainView backgroundThrottling restored (... boot: done) and a visible window titled "Claude" showing claude.ai/login, then exited one second later.

A healthy run (22:25, before the failure) continues past that point with:

[PluginBridgeMcp] no account/org yet — skipping plugin scan
[account] IPC did not provide accountUuid, falling back to bootstrap API
claude.ai account active and logged in
[WarmLifecycle:cowork] Destroyed

[WarmLifecycle:cowork] Destroyed never appears in any failing run.

Diagnostics that came back empty

• WER LocalDumps enabled for Claude.exe (DumpType 2) → no dump written
• Chromium verbose log (--enable-logging=file --v=1, 129 KB) → no FATAL, no ERROR; log ends mid-stream while the renderer is fetching the web app's JS modules
• Sentry scope_v3.json → no exception, no breadcrumbs
• Windows Application/System event logs → no Application Error, no Application Hang
• Crashpad reports/ → empty (last modified July)
• app.asar contains no literal exit(-1) call site

Ruled out

Each of these was tested individually and did not change the behaviour:

• Both app versions (1.30096.0.0 and 1.28929.0.0)
• Full reinstall from the official installer
• Add-AppxPackage -Register (re-registration)
• Completely empty user profile (LocalCache\Roaming\Claude renamed away)
• Different profile path (--user-data-dir=…)
• Resetting the UWP container state (Settings, AC, LocalState, RoamingState, TempState, SystemAppData)
• Deleting the Claude Code session and transcripts associated with the original crash
• All 25 transcript .jsonl files validated — none corrupt
• --disable-gpu, --no-sandbox, --allow-multiple-instances
• Stopping the CoworkVMService (Claude VM Service) entirely
• Providing the requested Claude Code bundle 2.1.229 locally (the app re-downloads it on every start and never finishes; the partial stalls at 67 MB of ~279 MB)
• Reboot
• Windows Defender, ASR, Controlled Folder Access, Smart App Control, Code Integrity, parental controls, IFEO debugger entries — all clean/absent
• Bluetooth (the [buddy] BLE bridge): machine has no Bluetooth adapter at all
• Disk space (250 GB free), environment variables (no ELECTRON_RUN_AS_NODE, no NODE_OPTIONS)
• Other Electron apps (Notion) run for hours without issue on the same machine

Reproduction

Not reliably reproducible from scratch — but on this machine it is 100% reproducible: every launch fails identically, and has done so across ~40 attempts, two app versions, a reinstall and a reboot.

The triggering action was previewing a localhost dev server (an Electron+Vite project) from inside a Claude Code session in the desktop app.

Notes

• The app re-downloads Claude Code bundle 2.1.229 on every start into claude-code\2.1.229\download.<hash>.zst.partial and never completes it, because the process exits first. 2.1.227 is present and verified.
• main-window.log has been 0 bytes since before the failure; window output goes to unknown-window.log instead.
• Possibly related, but different mechanisms: #46390 (worked, then crashed, reinstalled, won't open), #53247 (orphaned silo — ruled out here).

Logs available on request

main.log, unknown-window.log, claude.ai-web.log, cowork_vm_node.log, ssh.log, the Chromium verbose log, and the kernel process trace.