Environment
- Windows 10 Pro 19045, x64
- Claude Desktop 1.24012.9 (MSIX, direct download, not Store)
- Claude Code 2.1.219 native at crash time (2.1.220 after recovery)
- claude.ai subscription auth (Max), Remote Control active with two concurrent Claude Code sessions bridged to the desktop app
Summary
Claude Desktop hard-crashed mid-stream. The app could not relaunch ("This app can't open - There's a problem with Claude. Reinstall the application"), the taskbar shortcut disappeared, and reinstall failed repeatedly until a manual Remove-AppxPackage plus reboot. The wedge chain is fully visible in the AppXDeploymentServer/Operational event log plus ClaudeSetup.log: zombie app processes and the CoworkVMService held file locks, the installer's own cleanup could not remove that service ("Access is denied"), elevation failed, and Windows refused the data-preserving removal path. The eventual manual package removal destroyed all package-local app data.
Timeline (2026-07-25, UTC-7, from event log + ClaudeSetup.log)
- 14:52:04 - precursor: 15x event 8761 "AppExecutionAlias directory missing" (0x8007010B), ~2.5 h before the crash (auto-update staging?)
- 17:31:46 - last activity timestamps in both live CC session transcripts (crash moment; = 2026-07-26T00:31:46Z)
- 17:32:04 onward - repeated register/repair attempts fail 0x80073D02 "Unable to install because the following apps need to be closed" (zombie processes still alive)
- 17:35-17:46 - installer attempts (ClaudeSetup.log excerpt below): service removal denied, data-preserving removal rejected (0x80073CFA, requires developer mode), AddPackage 0x80073CF9 with inner 0x80070020 sharing violation creating
...\app\resources\cowork-svc.exe, then "Elevation failed: elevated installer exited with code 1" while the UI showed "Administrator access is required to install Claude with full features" to an already-elevated user
- 18:03-18:04 - after manual
Get-AppxPackage *Claude* | Remove-AppxPackage (Windows PowerShell 5.1 admin; the Appx module fails in pwsh 7 with 0x80131539): AddPackage now fails 0x80073CF6 with inner 0x80073D05 "An error occurred while deleting the package's previously existing application data" while registering windows.stateExtension
- reboot flushes the locks
- 18:13 - clean install succeeds, fresh package state
- Aftermath: fresh OAuth login required (401 on
claude rc); a globally set ANTHROPIC_API_KEY forced API-key auth so Remote Control refused to start, but the desktop banner only says "check your terminal" while claude doctor knows the actual cause; the manual /login URL flow was nearly unusable in default Win10 conhost because clipboard paste was broken and OAuth codes kept expiring
ClaudeSetup.log excerpt
17:35:15.925682 WARNING: failed to remove conflicting service: could not open CoworkVMService: Access is denied.
17:35:15.946275 Windows rejected data-preserving removal for Claude_1.24012.9.0_x64__pzs8sxrjxfjjc (0x80073CFA, requires developer mode); relying on in-place update
17:35:17.230496 MSIX installation failed: AddPackage failed: AddPackage failed with HRESULT 0x80073CF9
17:35:21.256976 Elevation failed: elevated installer exited with code 1
17:35:21.256976 ERROR dialog (with log path): Administrator access is required to install Claude with full features. You can try again or install without Cowork.
17:46:00.414773 (repeat cycle, same failure)
18:04:37.236124 MSIX installation failed: AddPackage failed with HRESULT 0x80073CF6
Full 29 KB log available on request.
Data lost
Recovery required package removal, which destroyed package-local state:
- Code-tab custom group assignments and pins. These live in renderer localStorage (
dframe-store / dframe-starred-code, origin https://claude.ai) inside the package profile, so they die with the package while the session transcripts under ~/.claude survive. Post-recovery investigation confirmed the desktop app is the sync owner for this state (ccd-sync-owner protocol): assignments written in the claude.ai web app do not propagate back to Desktop, so there is no user-accessible way to restore them - they had to be reconstructed by hand-patching the leveldb WAL offline.
- Crashpad/sentry evidence for the original fault (see below).
- Device registration, window state, and app settings.
Original crash cause
Unrecoverable locally, and the reason is circular: no Event 1000/WER record exists (Crashpad captured the fault), but the Crashpad/sentry data lived in the package profile that had to be destroyed to un-wedge the app. Server-side crash telemetry for account 3b7ebde4-1173-48bb-a3f7-05f7ae380469 (org d65be200-e7c3-4a7e-b86d-789666e120a2) around 2026-07-26T00:31:50Z should hold the dump; happy to correlate.
Suggested fixes
- Installer: detect and terminate stale app processes and stop/remove CoworkVMService before AddPackage; retry on 0x80070020. The cleanup exists but fails ("Access is denied") even when elevated - service ACL issue?
- Fix the elevation failure (elevated installer exit code 1) and the misleading "Administrator access is required" dialog when already elevated.
- Provide a data-preserving reinstall that does not require developer mode (0x80073CFA), or move critical state out of package-local storage.
- Persist Code-tab group assignments and pins next to session data (e.g. under
~/.claude) rather than in renderer localStorage - it is exactly the state users lose in any reinstall, while their transcripts survive. Being the sync owner makes it unrecoverable even from claude.ai.
- The Remote Control disconnected banner should surface "ANTHROPIC_API_KEY is forcing API-key auth" when that is the cause (
claude doctor already knows).
- Login fallback: a short user-typeable device-code flow for environments where clipboard interop is broken (default Win10 conhost).
Possibly related
#49917 (same 0x80073CF9 AddPackage failure, open), #47877 (MSIX stuck permanently), #40682 (installer fails, Canary), #71159 (0x80073CF9 orphaned package)
Environment
Summary
Claude Desktop hard-crashed mid-stream. The app could not relaunch ("This app can't open - There's a problem with Claude. Reinstall the application"), the taskbar shortcut disappeared, and reinstall failed repeatedly until a manual
Remove-AppxPackageplus reboot. The wedge chain is fully visible in the AppXDeploymentServer/Operational event log plus ClaudeSetup.log: zombie app processes and the CoworkVMService held file locks, the installer's own cleanup could not remove that service ("Access is denied"), elevation failed, and Windows refused the data-preserving removal path. The eventual manual package removal destroyed all package-local app data.Timeline (2026-07-25, UTC-7, from event log + ClaudeSetup.log)
...\app\resources\cowork-svc.exe, then "Elevation failed: elevated installer exited with code 1" while the UI showed "Administrator access is required to install Claude with full features" to an already-elevated userGet-AppxPackage *Claude* | Remove-AppxPackage(Windows PowerShell 5.1 admin; the Appx module fails in pwsh 7 with 0x80131539): AddPackage now fails 0x80073CF6 with inner 0x80073D05 "An error occurred while deleting the package's previously existing application data" while registering windows.stateExtensionclaude rc); a globally set ANTHROPIC_API_KEY forced API-key auth so Remote Control refused to start, but the desktop banner only says "check your terminal" whileclaude doctorknows the actual cause; the manual /login URL flow was nearly unusable in default Win10 conhost because clipboard paste was broken and OAuth codes kept expiringClaudeSetup.log excerpt
Full 29 KB log available on request.
Data lost
Recovery required package removal, which destroyed package-local state:
dframe-store/dframe-starred-code, origin https://claude.ai) inside the package profile, so they die with the package while the session transcripts under~/.claudesurvive. Post-recovery investigation confirmed the desktop app is the sync owner for this state (ccd-sync-ownerprotocol): assignments written in the claude.ai web app do not propagate back to Desktop, so there is no user-accessible way to restore them - they had to be reconstructed by hand-patching the leveldb WAL offline.Original crash cause
Unrecoverable locally, and the reason is circular: no Event 1000/WER record exists (Crashpad captured the fault), but the Crashpad/sentry data lived in the package profile that had to be destroyed to un-wedge the app. Server-side crash telemetry for account 3b7ebde4-1173-48bb-a3f7-05f7ae380469 (org d65be200-e7c3-4a7e-b86d-789666e120a2) around 2026-07-26T00:31:50Z should hold the dump; happy to correlate.
Suggested fixes
~/.claude) rather than in renderer localStorage - it is exactly the state users lose in any reinstall, while their transcripts survive. Being the sync owner makes it unrecoverable even from claude.ai.claude doctoralready knows).Possibly related
#49917 (same 0x80073CF9 AddPackage failure, open), #47877 (MSIX stuck permanently), #40682 (installer fails, Canary), #71159 (0x80073CF9 orphaned package)