[BUG] Claude Desktop (Windows/MSIX): fatal GPU-process crash (exit 101457950 / 0x060C201E) on preview surfaces — --disable-gpu-compositing does NOT mitigate; package left Modified/NeedsRemediation

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

Claude Desktop disappears without warning while working in the Code tab. Reproduced on two independent machines, including a brand-new laptop with a clean install performed the previous evening (2026-08-22 20:39). Crashpad captured the crash: the embedded Node.js utility process performs a deliberate abort.

The most reliable trigger observed so far: resuming a heavy Code session and starting a review with 3 parallel agents — the app dies within ~5 minutes. Light sessions run noticeably longer. (Caveat: Code is the only tab in regular use here, so I cannot confirm whether Chat alone is stable.)

After each crash:

• Get-AppxPackage Claude flips from Ok to Modified, NeedsRemediation without any deployment action by the user.


• On next launch the app re-initializes (first_launch_at is reset) and the session index is emptied; transcripts under ~/.claude survive.


• No Application Error (Event ID 1000) is written to the Windows event log — consistent with Crashpad intercepting the exception before WER.

On machine A this cascaded into the known reinstall spiral: 0x80073CF9 with inner 0x80070020 ERROR_SHARING_VIOLATION on app\resources\cowork-svc.exe (held by CoworkVMService, whose ACL also blocks sc stop/delete even elevated), 0x80073CFA (requires developer mode), and 0x80073D02. Full logs available.

CoworkVMService is ruled out as the crash cause: on machine A a crash occurred while the service was Stopped + Disabled (verified before and after). System memory exhaustion is also ruled out for the latest crash (no Resource-Exhaustion-Detector events; 16 GB RAM).

Crash evidence (machine B, 2026-08-23)

Two Crashpad reports written one second apart (two processes involved):

• 7b8a748a-15b3-4368-b5e2-4595133c8e3d — 15:19:10 local (13:19:10 UTC)


• 9b76bbed-b37f-4f17-aabe-c9e53e4e33e2 — 15:19:11 local

Sentry session files were updated at crash time, so these reports may already be on your backend under those IDs.

Minidump analysis of 9b76bbed… (no public symbols, so frames are Claude.exe+offset):

• Crashed process command line: "…\Claude_1.34493.1.0_x64__pzs8sxrjxfjjc\app\Claude.exe" --type=utility --utility-sub-type=node.mojom.NodeService --lang=es …


• Crashpad annotations: ptype=utility, process_type=utility, platform=win32, osarch=x86_64, pid=11064


• Exception: 0x80000003 EXCEPTION_BREAKPOINT, raised in KERNELBASE.dll (deliberate abort / fatal path), crashing thread 536; stack resolves into Claude.exe (statically linked Electron/Node)


• Not a per-process OOM: only ~0.37 GB committed at crash time (~3.5 TB reserved, i.e. normal V8 cage reservations); 24 threads incl. V8Worker; trivial CPU time (3 s user)


• No OOM/fatal message strings ("Reached heap limit", "FatalProcessOutOfMemory", …) present in the captured memory


• Process start 14:44:44 local → ~34 min uptime, crash within ~5 min of resuming the heavy session


• Roaming\Claude\git-worktrees.json was written inside the crash window (agents/worktrees being set up)

Sentry breadcrumb timeline (browser process — last seconds before the crash)

sentry/scope_v3.json (event.process: browser) holds 100 breadcrumbs ending at 13:19:09 UTC — one second before the Crashpad dumps (13:19:10/11 UTC):

• 13:18:49 / 13:18:59 / 13:19:09 — powerMonitor.speed-limit-change ×3 in 20 s (CPU speed-limit changes: machine throttling under load)


• 13:19:06 — app.session-created, app.web-contents-created, several renderer.dom-ready (new sessions/webcontents spinning up — the agent review starting)


• 13:19:06 — POST https://api.anthropic.com/v1/oauth/{uuid}/authorize → 200; POST …/v1/oauth/token → 200


• 13:19:07–09 — url_hash_check / domain_info (browser_extension) → 200


• trail ends; both dumps written at 13:19:10–11

sentry/session.json was written at process start (12:44:41.7 UTC, matching app_start_time 12:44:41.597Z) and was never finalized — still init: true, status: "ok", errors: 0, duration ≈ 0 — i.e. no graceful shutdown path ran, and no handled error preceded the abort. The second Crashpad report (7b8a748a…, written 1 s before 9b76bbed…) is therefore most likely the browser process going down together with the Node utility.

Runtime (from Sentry contexts): Electron 42.9.2, Chromium 148.0.7778.280, Node 24.18.1, release Claude@1.34493.1, build_type: windows-store.

Environment

Steps to Reproduce

1. Clean install of Claude Desktop 1.34493.1.0 on Windows 11 (10.0.26200) — installer completes with no errors.


2. Sign in, work in the Code tab. Light sessions: no crash for a while.


3. Resume a heavier session; start a review using 3 parallel agents.


4. Within ~5 minutes the app window disappears (no dialog). Crashpad writes two .dmp reports; Get-AppxPackage now shows Modified, NeedsRemediation; next launch re-initializes the app.

Reproduced 5+ times across the two machines over two days (2026-08-22/23).

### What Should Happen?

A Code-tab multi-agent review should not abort the Node utility process; and a child-process crash should not take down the whole app, wipe the session index, or leave the MSIX package flagged Modified, NeedsRemediation.

### Error Messages/Logs

```shell

```

### Steps to Reproduce

Clean install of Claude Desktop 1.34493.1.0 on Windows 11 (10.0.26200) — installer completes with no errors.
Sign in, work in the Code tab. Light sessions: no crash for a while.
Resume a heavier session; start a review using 3 parallel agents.
Within ~5 minutes the app window disappears (no dialog). Crashpad writes two .dmp reports; Get-AppxPackage now shows Modified, NeedsRemediation; next launch re-initializes the app.

Reproduced 5+ times across the two machines over two days (2026-08-22/23).

### Claude Model

Not sure / Multiple models

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

2.1.241 (Claude Code)

### Platform

Anthropic API

### Operating System

Windows

### Terminal/Shell

PowerShell

### Additional Information

[informe-github-claude-desktop.md](https://github.com/user-attachments/files/31350912/informe-github-claude-desktop.md)

[eventos-appx-maquinaA.txt](https://github.com/user-attachments/files/31350973/eventos-appx-maquinaA.txt)

[ClaudeSetup-portatil.txt](https://github.com/user-attachments/files/31350978/ClaudeSetup-portatil.txt)

[ClaudeSetup-maquinaA.txt](https://github.com/user-attachments/files/31350995/ClaudeSetup-maquinaA.txt)

Start with the two Crashpad reports, sentry/scope_v3.json, sentry/session.json, and the attached Windows/MSIX logs, comparing the Node utility abort with the browser-process exit and package state changes. Review the git-worktrees.json write in the crash window and define done as preventing the abort and preserving the app session and MSIX package state during the reproduced three-agent review.

Escrito por el modelo de indexación a partir del texto del issue.