Skip to content

Windows: GPU process crashes and app restarts whenever a preview pane renders (Claude Desktop 1.40609.0) #91273

Open
Open
Windows: GPU process crashes and app restarts whenever a preview pane renders (Claude Desktop 1.40609.0)#91273
Labels
invalidIssue doesn't seem to be related to Claude Code

Description

@phillipcruz01-rgb
phillipcruz01-rgb
opened on Sep 1, 2026

Claude Desktop: GPU process crashes when a preview pane renders, app restarts

Summary

On Windows 11 with an NVIDIA RTX 3070, the Claude Desktop GPU process crashes
and the app relaunches itself. It reproduces when a preview pane is created,
which happens whenever I open a subagent that carries the browser toolset. It
has happened four times today.

Persists after a clean NVIDIA driver install, including the driver whose release
notes list a Claude Desktop fix.

Environment
Claude Desktop 1.40609.0 (MSIX, Claude_1.40609.0.0_x64__pzs8sxrjxfjjc)
OS Windows 11 Pro 10.0.26200
GPU NVIDIA GeForce RTX 3070
Driver before 32.0.15.9186 (dated 2026-01-19)
Driver after 32.0.16.1656 / Studio 616.56 (dated 2026-08-26)
Display 2560x1440

What happens

The GPU process exits with reason: 'crashed', the window goes down, and the
app restarts on its own. No Crashpad minidump is written and there is no
claude.exe entry in the Windows Application event log, so the crash leaves no
artifact outside the app's own log.

Occurrences, all 2026-09-01

Time exitCode Driver at the time
09:38:50 101457950 32.0.15.9186
09:46:09 101457950 32.0.15.9186
10:03:30 34 during driver install
11:48:39 101457950 32.0.16.1656

Correlation: a preview pane renders 1 to 4 seconds before every crash

Instance at 09:46:09:

09:46:05 [info] [Preview] Created session preview context {
  previewId: 'preview-local_afce282b-5c41-4e66-9231-b6543f6a08ea',
  totalContexts: 1
}
09:46:05 [info] [Preview] Created browser preview { serverId: 'browser-preview-1788273965306-0' }
09:46:07 [info] [oauth] using cached token
09:46:09 [info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }
09:47:05 [info] Starting app { appVersion: '1.40609.0' }

Instance at 11:48:39, after the driver update:

11:48:38 [warn] [PreviewContext] Blocked subresource to private-resolving host { resourceType: 'xhr' }
11:48:39 [info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }
11:50:18 [info] Starting app { appVersion: '1.40609.0' }

There is also a [PreviewContext] Max retries reached, giving up at 09:32:20,
so the preview surface appears to be unhealthy before it takes the GPU process
down.

Ruled out

  • Memory. At 11:48:39 the process tree was 4,738 MB with 10.6 GB system free.
    At 10:03:30 it was 6,694 MB with 10,614 MB free. No resource-exhaustion events.
  • Graphics driver. Reproduced on 32.0.16.1656 after a clean install with
    "Perform a clean installation" checked. Notably, the 616.56 release notes list
    a fix for "Claude Desktop: Driver incorrectly detects application as a 3D app
    [6468260]", which did not resolve this.
  • OS-level display fault. No TDR events (System log id 4101) at any point.
    Only the app's GPU process dies; the display driver itself never resets.

Separate issue seen in the same logs

On 2026-08-31 at 15:59:59 the whole child-process tree was killed at once, with a
different signature. Including it in case it is related.

15:59:59 [System]  The Claude service terminated unexpectedly (event 7034)
15:59:59 [error]   Network Service process gone (reason=killed, exitCode=1) outside app quit
15:59:59 [error]   Claude Code process exited with code 1   x4 sessions
15:59:59 [warn]    pty-host worker exited (1) with 4 live PTY(s)
16:00:25 [AppX]    Deployment Register operation on Package Claude_1.40609.0.0 finished successfully

Windows re-registered the MSIX package 26 seconds after killing the service. The
AppXDeploymentServer log shows the Claude package being re-registered ten times
in one day, and 5,326 deployment events over four days, driven by a retry loop
on an unrelated stale package that Windows cannot delete. If Claude Desktop can
survive package servicing without losing its service and every child process,
that would be worth hardening.

Recurring error, possibly related to the MSIX packaging

This one repeats several times an hour, every hour, going back days:

[error] [Chrome Extension MCP] Failed to copy native host binary:
Error: EBUSY: resource busy or locked, copyfile
'C:\Program Files\WindowsApps\Claude_1.40609.0.0_x64__pzs8sxrjxfjjc\app\resources\chrome-nat...'

Ask

A way to disable hardware acceleration would let me work around this today. I
looked through the app config and found no such key, and the app is MSIX
packaged so passing --disable-gpu on the command line is not straightforward.

Log location

%LOCALAPPDATA%\Claude\Logs\main.log

Happy to send the full log or run any diagnostic.

Activity

  1. github-actions
    added
    invalidIssue doesn't seem to be related to Claude Code
    on Sep 1, 2026

  2. phillipcruz01-rgb commented on Sep 1, 2026

    @phillipcruz01-rgb
    phillipcruz01-rgb
    on Sep 1, 2026
    Author

    Correction to the original report, plus the surface it happens on

    Two updates after re-reading the full day's log.

    1. I overstated the trigger. Preview creation is not it.

    The original report said a preview pane renders 1 to 4 seconds before every
    crash. That does not hold up. Browser previews were created 11 times on
    2026-09-01, and only one of the four crashes followed a creation within seconds.

    Crash exitCode Nearest preceding preview event
    09:38:50 101457950 Created browser preview at 09:35:03, 3m47s earlier
    09:46:09 101457950 Created browser preview at 09:46:05, 4s earlier
    10:03:30 34 Created session preview context at 09:59:05, 4m25s earlier
    11:48:39 101457950 [PreviewContext] Blocked subresource at 11:48:38, 1s earlier; last creation 11:01:40

    What is accurate: at least one browser preview pane was open and active during
    every crash.     That is much weaker than a creation-time trigger and I should not
    have framed it as one. Creation itself is clearly survivable, 10 out of 11 times.

    Preview creations that did not crash: 07:27:56, 07:28:44, 08:56:13, 09:17:29,
    09:35:03, 09:49:55, 10:22:52, 11:01:40, 11:51:11, 11:57:22.

    2. Which surface: Claude Code in the desktop app, not Cowork.

    Every crashing session is a local_* session handled by LocalSessions and
    logged under the [CCD] prefix. The preview object is:

    [Launch] Updating active servers store {
  servers: '[{"serverId":"browser-preview-1788273965306-0",
              "name":"Browser",
              "sessionId":"local_afce282b-5c41-4e66-9231-b6543f6a08ea", ...

    Cowork appears in the log only at app startup, initialized and destroyed inside
    one second, and was not running at any crash:

    11:50:20 [WarmLifecycle:cowork] Initialized (arm=always)
11:50:21 [CoworkScheduledTasksApi] getAllScheduledTasks: Scheduled tasks not initialized
11:50:21 [WarmLifecycle:cowork] Destroyed

    So: Claude Code desktop sessions with the Browser pane, not Claude Cowork.

    What still holds from the original report

    • GPU process dies with reason: 'crashed', exitCode 101457950 three times and
      34 once. App relaunches itself each time.
    • No Crashpad minidump, no claude.exe entry in the Windows Application event
      log. The app's own main.log is the only record.
    • Not memory. 4,738 MB in use with 10.6 GB free at the 11:48:39 crash.
    • Not the graphics driver. Reproduced on a clean install of Studio 616.56
      (32.0.16.1656), the release whose notes list "Claude Desktop: Driver
      incorrectly detects application as a 3D app [6468260]".
    • No TDR events (System log id 4101). The display driver never resets; only the
      app's GPU process dies.

    I do not have a confirmed trigger. Happy to run whatever instrumentation helps
    narrow it, or to send the full log.

  3. jasonbanks-tx commented on Sep 2, 2026

    @jasonbanks-tx
    jasonbanks-tx
    on Sep 2, 2026

    Another data point for this, from a different machine.

    Reproduced on: Claude Desktop 1.40609.1.0 (MSIX), Windows 11 build 10.0.26200.8655 (25H2), NVIDIA GeForce RTX 4080 Laptop GPU, driver 32.0.16.1088 (nvidia-smi reports 610.88).

    Sequence from %LOCALAPPDATA%\Claude\Logs\main.log:

    • 10:47:55 [info] [Preview] Created browser preview { serverId: 'browser-preview-...' }
    • 10:48:03 [info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }

    Same exit code as the original report, 8 seconds after the preview pane was created. Every window of the app went dark at once (one GPU process serves all windows). No nvlddmkm / TDR event in the System log - the display driver did not reset; only Electron's GPU process died.

    Not deterministic here: an earlier preview pane opened at 10:13 the same morning ran for about 34 minutes without incident. The second one crashed the GPU process in 8 seconds.

    This driver version sits between the two the original reporter tested (32.0.15.9186 and 32.0.16.1656), so it does not look driver-version-specific.

    One side effect worth knowing for anyone whose AppX deployment stack is already unhealthy: after the GPU crash, Windows attempted an automatic package re-register, and on this machine that failed on a pre-existing, unrelated AppX defect, which escalated to 0x80073CFC NEEDS_REMEDIATION and forced the Settings > Repair / Reset flow. Repair failed with 0x80073D02 while any Claude.exe process was still alive; Reset succeeded. That part is machine-specific - the GPU crash itself is not.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    invalidIssue doesn't seem to be related to Claude Code

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions