claude-code - 💡(How to fix) Fix Cowork: VM connection timeout on Windows ARM64 (Snapdragon X Plus) [2 comments, 3 participants]

Description

Claude Cowork fails to start on Windows ARM64 (Snapdragon X Plus) with "VM connection timeout after 60 seconds." It has never successfully connected — zero successful guest connections across 30+ attempts from March 26 to April 9, 2026.

System Info

• CPU: Snapdragon X Plus (X1P42100), 8 cores
• OS: Windows 11 Home ARM64, Build 26200
• RAM: 32GB
• Disk: 693GB free
• Claude version: 1.1348.0.0 (arm64, Microsoft Store)
• VM config: 4096MB RAM, 2 CPUs allocated

Failure Pattern

The VM boots successfully via HCS, but the Linux guest never establishes an HvSocket connection:

[VM:steps] create_vm_config completed (63ms)
[VM:steps] vm_boot completed (104ms)
[VM:steps] add_plan9_shares completed (3ms)
[VM:start] Still waiting for guest connection... 10130ms elapsed, 21 polls
[VM:start] Still waiting for guest connection... 20293ms elapsed, 41 polls
[VM:start] Still waiting for guest connection... 30474ms elapsed, 61 polls
[VM:start] Still waiting for guest connection... 40645ms elapsed, 81 polls
[VM:start] Still waiting for guest connection... 50849ms elapsed, 101 polls
[VM:start] Connection timeout, last completed step: add_plan9_shares

What I've Tried

• Restarting the computer (multiple times)
• Deleting smol-bin.vhdx and letting it recreate
• Full workspace reinstall (deleted entire vm_bundles directory, fresh re-download) — same timeout
• Stopping vmcompute service, killing vmwp.exe, restarting services
• The auto-reinstall ran once on March 26 and failed with the same timeout

Additional Context

• Hyper-V Host Compute Service (vmcompute) and HvHost are both running correctly
• The VM boots (vmwp.exe starts), but the ARM64 Linux kernel/init inside the VM appears to fail silently before it can start its agent service
• Secondary issue: after a failed attempt, vmwp.exe holds a lock on smol-bin.vhdx, causing EBUSY errors on subsequent attempts until vmcompute is stopped and vmwp is killed
• Cowork was announced for Windows ARM on 2025-03-25; the first attempt on this machine was March 26 and it has never worked since

Relevant Logs

• cowork_vm_node.log: Every attempt ends with Connection timeout, last completed step: add_plan9_shares
• cowork-service.log: HCS initializes correctly, VM session created, network configured — all infrastructure works
• Boot mode: UEFI BootThis+VmbFs (ARM64)
• Kernel cmdline: initrd=\initrd root=/dev/sda1 rootfstype=ext4 rw panic=10