Claude Desktop browser pane crash 0x060C201E: a missing MSIX catalog, not a GPU fault

Claude Desktop browser pane crash 0x060C201E

Solved and verified 29 August 2026 on Claude desktop app 1.40609.0, Windows 11.

Symptom

Loading a WebGL-probing page (a site behind a Cloudflare Turnstile check) in the in-app Browser pane killed the whole app within 1 to 35 seconds. The log in %LOCALAPPDATA%\Claude\Logs\main.log showed:

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

101457950 is 0x060C201E. There were seven crashes over a week. Sometimes the MSIX package was left unlaunchable; Settings, Apps, Claude, Repair plus killing the leftover process recovered it.

Root cause

It is not a graphics problem.

• The page's WebGL probe makes Chromium's GPU process lazy-load vk_swiftshader.dll.
• The GPU process runs with Code Integrity Guard, which only accepts Microsoft or store-signed code. Package DLLs normally qualify through the MSIX catalog.
• On the affected install, the whole AppxMetadata directory, including CodeIntegrity.cat, was missing on disk (AppxBlockMap, signature and manifest were present).
• The CodeIntegrity event log showed event 3033 "vk_swiftshader.dll did not meet the Microsoft signing level requirements" and event 3010 "unable to load ...AppxMetadata\CodeIntegrity.cat" with 0xC000003A.
• The loader then raised the non-continuable 0x060C201E through ntdll!LdrAppxHandleIntegrityFailure, which ends the whole app and marks the package NeedsRemediation.

This explains why the usual mitigations did nothing: pinning the discrete GPU, disabling hardware acceleration (WARP crashed within about a second), and driver or DNS changes. Light pages never load vk_swiftshader.dll, so they never crash. The DLLs are validly Authenticode-signed, which is not enough under Code Integrity Guard. Crashpad and Windows Error Reporting never fire on loader kills.

Fix

A clean reinstall so Windows re-stages the package metadata:

winget uninstall Anthropic.Claude
winget install Anthropic.Claude

In the observed case the reinstall replaced the MSIX package with a standard per-user install (under %LOCALAPPDATA%, Get-AppxPackage *Claude* returned nothing). With no package identity the AppX integrity-kill path cannot fire. A test page with the Turnstile check stayed up for 40+ seconds with no new GPU process gone line, and a WebGL test page rendered correctly. Because the package is gone, Test-Path on WindowsApps\Claude_*\AppxMetadata\CodeIntegrity.cat returning False after the reinstall is expected.

Related gotcha: MSIX virtualises AppData writes

While the app was MSIX, shells spawned from inside it had AppData writes virtualised into the package's LocalCache. A crash-dump folder that appeared to live in %LOCALAPPDATA% actually lived in the package and was deleted on uninstall, and other MSIX apps such as WinDbg could not see the files. MSIX uninstall also wipes LocalCache, so back up anything needed first.

Diagnostic method that found it

• procdump64 -ma -t -e <gpu pid> attached before the repro catches loader kills that Crashpad and WER miss.
• WER LocalDumps armed through HKLM (this needed a UAC-elevated script file; Start-Process -Verb RunAs with an inline -Command mangles nested quotes).
• WinDbg from winget (Microsoft.WinDbg) can run headless-ish: WinDbgX -z <dump> -c '$$<C:\path\cmds.txt' with .logopen and .logclose. The $$>< form joins script lines into one command and breaks .sympath; use $$<. The MSIX WinDbg could not open a dump under %LOCALAPPDATA% (Win32 error 3); copy it to a plain directory first.
• ELECTRON_ENABLE_LOGGING=file with ELECTRON_LOG_FILE left the log empty for the GPU child process and was not useful.
• Full dumps contain process memory and should stay local.