GPU process crashes repeatedly with constant exitCode 101457950 on a hybrid-graphics laptop

Summary

The Electron GPU process dies with reason: 'crashed', exitCode: 101457950 — 16 times over 16
days, 1–4 per day. The exit code is byte-identical every single time. The app window survives some
of these and disappears entirely on others; after one crash it would not reopen at all until a
reinstall.

The constancy of the exit code across two GPU vendors, two graphics drivers, three app versions and
two antivirus configurations is the reason I'm filing: it looks like one deterministic code path
rather than random driver memory corruption.

Environment

| | |
|---|---|
| App | Claude Code desktop, MSIX package Claude_pzs8sxrjxfjjc |
| Versions affected | 1.24012.9, 1.25927.0, 1.26832.0 (all three) |
| OS | Windows 11 Home 25H2, build 26200 |
| Laptop | Lenovo Legion 5 16IRX9 (83DG) |
| CPU | Intel Core i9-14900HX, 24 cores |
| GPUs | Hybrid/Optimus: NVIDIA RTX 4070 Laptop + Intel UHD Graphics, both adapters active |
| RAM | 31.7 GiB |

The log signature

Identical every time, in %APPDATA%\Claude\logs\main.log:

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

101457950 = 0x60C201E. Not a documented NTSTATUS as far as I can tell. The app writes this line
twice per event, so a naive grep -c doubles the count.

All 16 occurrences (local time, PT)

2026-07-25 17:28:03   v1.24012.9
2026-08-05 16:20:31   v1.25927.0
2026-08-05 16:26:56   v1.25927.0
2026-08-05 16:33:03   v1.25927.0
2026-08-05 16:42:18   v1.25927.0
2026-08-06 09:21:50   v1.25927.0
2026-08-06 09:51:39   v1.25927.0
2026-08-07 08:05:57   v1.26832.0
2026-08-07 16:53:11   v1.26832.0
2026-08-07 16:55:53   v1.26832.0
2026-08-08 11:36:26   v1.26832.0
2026-08-09 07:32:43   v1.26832.0
2026-08-10 17:31:41   v1.26832.0
2026-08-10 17:42:44   v1.26832.0
2026-08-10 19:24:04   v1.26832.0
2026-08-10 19:26:27   v1.26832.0

Under sustained load the interval compresses to minutes (four crashes in 22 minutes on 08-05; two
2.4 minutes apart on 08-10). Idle periods go many hours without one.

What I ruled out, and how

Each of these was tested, not assumed:

• Memory pressure. 9–15 GB free at every sampled crash window; app tree ~2.3 GB.
• Windows display-driver reset (TDR). No Event ID 4101/4102 in the System log in any crash

window, and no Application Error entry for Claude at all — consistent with Chromium's GPU process
isolation working as designed.

• Corrupted shader cache. Deleted GPUCache, DawnGraphiteCache, DawnWebGPUCache; crashes

continued.

• App version. Crashes on 1.24012.9, 1.25927.0 and 1.26832.0. The oldest predates the

others by 11 days.

• Stale graphics drivers. Both were badly out of date and both were updated —

NVIDIA 572.70 → 610.88, Intel 32.0.101.6790 → 32.0.101.7088. Crashes continued at the same daily
rate afterward.

• Which GPU renders. The app was originally rendering on the Intel iGPU (verified by the loaded

module list in the GPU process: igd10iumd64.dll, igc64.dll, etc., zero NVIDIA modules). I
pinned it to the discrete GPU via HKCU\SOFTWARE\Microsoft\DirectX\UserGpuPreferences and
confirmed the switch two ways (6 NVIDIA modules incl. nvwgf2umx.dll; nvidia-smi listing the
GPU-process PID on the 4070). It still crashes on NVIDIA, with the same exit code.

• Third-party AV interference. A third-party suite (360安全卫士) was installed for most of this

period and has since been fully uninstalled, with Microsoft Defender restored to Normal.
Crashes occur both before and after.

• Injected DLLs. The GPU process loads 72 modules; zero are non-Microsoft, non-Claude.

Possible trigger, stated honestly as weak

Two of the sixteen crashes landed 3 seconds after a browser:open_site permission grant — i.e.
immediately after a Browser-pane page load:

2026-08-05 16:20:28  Received permission response ... (tool: browser:open_site)
2026-08-05 16:20:31  GPU process gone

2026-08-10 19:24:01  Received permission response ... (tool: browser:open_site)
2026-08-10 19:24:04  GPU process gone

That's striking, but I checked all 16 against every browser-load event in the logs and only 3 of
16 fall within 60 seconds of one. So I do not think Browser-pane rendering is the general cause —
though the two 3-second cases may still be worth a look, since they're the only crashes with an
obvious immediately-preceding action.

What would help

I can supply full main.log excerpts, the loaded-module list of the GPU process on either adapter,
or run any diagnostic you'd like. Happy to test a build with extra GPU-process logging.

If there's a way to make the GPU process emit a crash dump or a more descriptive exit reason on
Windows, that would probably resolve this faster than anything I can do from outside.