[BUG] Workflow launch + multi-tab focus switching crashes GPU process, cascades into CoworkVMService restart loop requiring OS reboot

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Cross-referenced logs from two separate Claude Code sessions (this app's main.log and a Windows service log) point to a reproducible crash chain:

1. WarmLifecycle:preview fires for multiple tabs nearly simultaneously (switching focus between tabs).
2. Around the same moment, a multi-agent Workflow is launched in one of the tabs (happened in two different tabs within ~10 seconds of each other).
3. A second renderer process spawns (process count 10→11) and the GPU process's memory usage doubles (142MB → 304MB).
4. The GPU process crashes shortly after (reason: crashed, exitCode: 101457950).
5. Within the same minute, a separate Windows service (CoworkVMService, display name "Claude", part of the Cowork VM sandbox feature) enters a rapid stop/start loop — restarting every 5-15 seconds for about 3 minutes (Windows System log Event ID 7034 "terminated unexpectedly" x4, and the service's own cowork-service.log shows clean stop/start cycles rather than an internal fault).
6. The user manually rebooted the machine to recover (Event ID 1074, reason "Other (Unplanned)").
7. Post-reboot, Windows also logged high I/O pressure (Volsnap Event ID 25: shadow copies of volume C: were deleted because the shadow copy storage could not grow in time).
8. Once back up, a process named Pulsar.exe crashed once with STATUS_HEAP_CORRUPTION (0xc0000374) — not confirmed whether this is part of the same component tree.

This same pattern reportedly occurred once before, in July, under similar conditions (multiple tabs + concurrent workflow launch).

What Should Happen?

Launching a Workflow while switching focus between multiple Claude Code tabs should not be able to crash the renderer/GPU process, and should especially not cascade into destabilizing the unrelated Cowork VM service to the point a full OS reboot is needed to recover.

Error Messages/Logs

Event ID 7034 (System log, x4 in ~3 min):
The Claude service terminated unexpectedly. It has done this 1 time(s).

Event ID 1074 (System log, restart trigger):
The process C:\WINDOWS\SystemApps\Microsoft.Windows.StartMenuExperienceHost_cw5n1h2txyewy\StartMenuExperienceHost.exe (W) has initiated the restart of computer W on behalf of user W\wapel for the following reason: Other (Unplanned)
Reason Code: 0x0
Shutdown Type: restart

Event ID 25, Volsnap (System log):
The shadow copies of volume C: were deleted because the shadow copy storage could not grow in time. Consider reducing the IO load on the system...

Event ID 1000, Application Error (post-reboot):
Faulting application name: Pulsar.exe, version: 0.0.0.0
Faulting module name: ntdll.dll, version: 10.0.26100.8875
Exception code: 0xc0000374
Fault offset: 0x0000000000112165

cowork-service.log (repeating every 5-15s during the loop):
Service stop requested
Shutdown signal received...
[Server] Stopping server...
[Server] Server stopped
Service stopped.
Starting CoworkVMService as Windows service
Claude VM Service starting...
Waiting for configuration from app via 'configure' method...
[HCS] Initializing HCS DLLs...
Warning: failed to open service for recovery config: Access is denied.

Steps to Reproduce

1. Open Claude Code desktop app with 2+ project sessions open as separate tabs.
2. In one tab, launch a multi-agent Workflow.
3. Within roughly the same 10-30 seconds, switch focus between 2 or more other tabs (this triggers WarmLifecycle:preview for each tab you land on).
4. Observed result: a second renderer process spawns and the GPU process's memory usage roughly doubles, followed by the GPU process crashing (exitCode 101457950).
5. Separately/shortly after: the CoworkVMService Windows service enters a rapid stop/start loop (Event ID 7034 in Windows Event Viewer, System log; also visible in C:\ProgramData\Claude\Logs\cowork-service.log).

Note: this is the sequence that preceded the crash, observed across two concurrent sessions' logs — not yet confirmed as a guaranteed, deterministic repro recipe.

Claude Model

None

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.210 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

This crash was cross-referenced live across two concurrent Claude Code sessions on the same machine, each reading a different log source (this app's own main.log for the renderer/GPU side, Windows Event Viewer + cowork-service.log for the service/OS side) — that's how the two halves of this timeline were connected.

Context at the time: the machine was also running several remote PowerShell/WinRM sessions to 4 other machines and a CPU-heavy multi-agent Workflow had just been launched in two tabs within ~10 seconds of each other, so overall system load was high. Not confirmed whether that contributed or is incidental.