[BUG] Claude Desktop 1.28929.0 (Windows/MSIX): GPU process crashes on session preview — regression from 1.26832.0

Summary

After the auto-update from 1.26832.0 → 1.28929.0, the Electron GPU process crashes
whenever a session preview context is created. The previous version ran the same preview
code path thousands of times with zero crashes.

After the crash Windows marks the AppX as Modified, NeedsRemediation and offers
"repair", and orphaned node.exe processes running MCP browser servers are left behind.

Environment

• Claude Desktop 1.28929.0.0, x64, MSIX/Store install ([CCD-autoupdate] Disabled: MSIX install)
• Windows 11
• Electron nodeVersion 24.18.0
• GPU: NVIDIA RTX 4090, driver 32.0.16.1062, installed 2026-06-30
• Also present: two virtual display adapters (Meta Virtual Monitor, SuperDisplay Virtual Adapter),

both with Kernel-PnP 219 / WUDFRd errors in the System log. These predate the regression.

The regression is in the app version, not the driver

| date | event | source |
|---|---|---|
| 2026-06-30 01:54 | NVIDIA driver 32.0.16.1062 installed | C:\Windows\INF\oem194.inf |
| … → 2026-08-11 20:13 | running 1.26832.0 — zero reason: 'crashed' GPU events | main1.log, main2.log |
| 2026-08-12 01:13 | AppX Claude_1.28929.0.0_x64 written to disk | WindowsApps CreationTime |
| 2026-08-12 06:34 | [updater] Version changed since last launch: 1.26832.0 → 1.28929.0 | main.log |
| 2026-08-12 23:02 | first reason: 'crashed' GPU event | main.log |
| 2026-08-13 08:21 / 08:53 / 09:13 | three more | main.log |

Six weeks on the same GPU driver with no crashes; the first crash arrives ~17 hours after
the app update. Every GPU process gone event in the older logs is
reason: 'killed', exitCode: 1 — i.e. clean shutdown, not a fault.

The preview feature is not new — only the crash is

Occurrences of the preview code path per log file:

| log | version | WarmLifecycle:preview | PreviewContext | crashes |
|---|---|---|---|---|
| main1.log | 1.26832.0 | 3319 | 28 | 0 |
| main.log | 1.28929.0 | 465 | 26 | 4 |

Repro / signature

All four crashes have the same immediate predecessor, within 0–1 second:

2026-08-12 23:02:38 [info] [Preview] Created session preview context {
  previewId: 'preview-local_<redacted>',
  totalContexts: 1
}
2026-08-12 23:02:45 [warn] [PreviewContext] Blocked subresource to private-resolving host { resourceType: 'xhr' }
2026-08-12 23:02:46 [info] GPU process gone: {
  type: 'GPU',
  reason: 'crashed',
  exitCode: 101457950,
  serviceName: 'GPU'
}

Three of the four follow [PreviewContext] Blocked subresource to private-resolving host;
the fourth follows [WarmLifecycle:preview] Warming up session.

Exit codes are not stable across crashes — 101457950 (0x60C201E) three times and 34
once — which is consistent with a genuine GPU-process fault rather than a clean teardown.

Preceding renderer memory at the time of the crash was ~640–700 MB
([process-memory] electron_renderer:...:701MB), so this does not look like OOM.

The renderer also reports WebGL: CONTEXT_LOST_WEBGL: loseContext: context lost, which is
the expected downstream effect of the shared GPU process dying.

Impact

• The app becomes unusable until restarted, and Windows flags the package as

Modified, NeedsRemediation ("This app can't open… select repair").

• Orphaned node.exe MCP browser-server processes survive the crash.
• Because the trigger is the session preview, this is not avoidable by changing

MCP configuration — it fires during normal use.

Notes

• Rollback is not available locally: C:\Program Files\WindowsApps contains no prior

version directory, and the MSIX install disables the in-app updater.

• Happy to supply the raw main.log privately if useful.