Claude Desktop (Windows): Electron GPU process crashes with exitCode 101457950 on browser-view creation, forcing repeated app Repair

Summary

Claude Desktop on Windows 11 repeatedly kills its Electron GPU process with exitCode 101457950 (0x60C201E) whenever a browser view is created (opening the Browser pane / preview, or taking a screenshot in it). The app then enters the state where Windows offers "Repair", and I have had to run Repair several times in a single working day.

This looks related to #85199 but the signature is different, so I am filing separately rather than adding noise there. I do not have that issue's log truncation: my logs rotate normally (main1..main4.log exist) and main.log was still writing at 5.8 MB.

Crash signature

Identical every time, nine occurrences:

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

101457950 = 0x60C201E. Notably this is not a Windows exception code — it is not 0xC0000005, 0xC0000409 or any NTSTATUS value — which suggests Chromium is terminating its own GPU process rather than being killed by a native fault.

Occurrences, across three app versions

| When | App version |
|---|---|
| 2026-06-17 15:24:18 (x4) | 1.13576.0 |
| 2026-08-11 14:51:02 (x2) | 1.26832.0 |
| 2026-08-12 12:34:15 | 1.28929.0 |
| 2026-08-12 12:40:13 | 1.28929.0 |
| 2026-08-12 12:59:31 | 1.28929.0 |

Six clean weeks in between (18 Jun – 10 Aug, zero crashes) — that period simply involved no browser-pane work.

Trigger

Every crash on 11–12 Aug happened while a browser view was being created or driven:

• Two occurred immediately after opening the Browser pane (preview_start).
• Two were preceded by a burst of Screenshot timed out after 5s: the Browser pane is not displayed, so the page is not compositing frames — 3 and 6 of them respectively in the seconds beforehand.
• The 11 Aug pair occurred during an ordinary claude.ai page load (Cloudflare Turnstile iframe + redirect), with no browser tool involved.

I logged 86 of those screenshot timeouts against only 4 crashes, so the timeouts are a symptom of the GPU process being unavailable to composite, not the cause.

What I ruled out

• Not file corruption. It survives every Windows "Repair" and predates them by two months.
• Not a single bad build. Three versions affected: 1.13576.0, 1.26832.0, 1.28929.0.
• Not the GPU driver. Zero TDR events (Event ID 4101) and zero nvlddmkm errors in 7 days — Windows never records a display-driver reset. Driver is current.
• Not a second display adapter. I had a Parsec Virtual Display Adapter (driver 0.45.0.0, dated Jan 2024) throwing Kernel-PnP 219 / \Driver\WUDFRd failed to load / 0xC0000365 on attach. I fully uninstalled Parsec and both its drivers. The crash recurred afterwards, so this was not the cause.
• Not memory pressure. At crash time: ~1.6–2.0 GB total across the Electron tree, 15 GB of 32 GB system memory free.
• Not GPU scheduling / TDR config. HwSchMode, TdrDelay, TdrDdiDelay, TdrLevel all at defaults.

Environment

• Claude Desktop 1.28929.0.0 (also 1.26832.0.0, 1.13576.0)
• Windows 11 Pro, build 10.0.26200
• NVIDIA GeForce RTX 3060, driver 32.0.15.9186 (dated 2026-01-20) — sole display adapter
• 32 GB RAM
• Installed from the Store/MSIX (C:\Program Files\WindowsApps\Claude_1.28929.0.0_x64__…)

Impact

Each crash drops the session and requires Repair via Settings → Apps → Claude. Because the trigger is browser-view creation, any workflow using the Browser pane or previews is effectively unusable — I have had to move all page rendering and screenshotting to a separate, out-of-process Playwright install to keep working.

What would help

There is currently no user-facing way to disable hardware acceleration in Claude Desktop. If the GPU process is going to be this fragile on some hardware, an option to run with --disable-gpu / software compositing (a setting, or an env var the MSIX launcher honours) would be a complete workaround. At present the only remedy is to avoid the feature entirely.

Also seen on every startup (probably unrelated, reporting for completeness)

[warn] [watch-record] failed to subscribe to swift events
Cannot find package '@ant/claude-swift' imported from
  …\app.asar\.vite\build\index.js { code: 'ERR_MODULE_NOT_FOUND' }