Cowork 'Failed to create bridge sockets' on macOS 26.2 Tahoe — regression after #19503 fix

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?

Claude Desktop Cowork fails on every task (new and resumed) with:

Error: Failed to create bridge sockets after 5 attempts

Exit code 1, process dies within ~750-900ms of spawning every time.

This was previously reported and fixed in #19503, but the fix does not work on macOS 26.2 (Tahoe). The issue reproduces on both claude-opus-4-7 and claude-haiku-4-5-20251001 model selections. Every cowork agent spawned fails identically.

Environment

• macOS: 26.2 (Build 25C56)
• Kernel: Darwin 25.2.0 (xnu-12377.61.12~1/RELEASE_ARM64_T8142)
• Architecture: arm64 (Apple Silicon)
• Claude Desktop SDK: 2.1.111
• Claude Code CLI: 2.1.92 (works perfectly — only cowork VM is affected)
• Node.js: v22.22.2 (also tested with v24.13.1 — same result)

Debugging Performed (extensive — ~1 hour, significant token usage)

We performed thorough debugging via Claude Code CLI to diagnose this issue:

1. Verified CLI works fine: claude --print "hello" succeeds, auth is valid (Max subscription, OAuth via claude.ai)
2. Checked all cowork logs:

• ~/Library/Logs/Claude/cowork_vm_node.log — every spawn fails with Failed to create bridge sockets after 5 attempts
• ~/Library/Logs/Claude/cowork_vm_swift.log — VM boots and connects successfully
• ~/Library/Logs/Claude/coworkd.log — VM infrastructure (systemd, networking, user recovery) all healthy

1. VM boots fine: Network connects (CONNECTED), API is REACHABLE, SDK installs successfully, OAuth tokens approved — the failure happens after the Claude Code process spawns inside the VM
2. Deleted session disk (sessiondata.img) — recreated on boot, same error
3. Deleted all VM state files (efivars.fd, machineIdentifier, vmIP, gvisorMacAddress, sessiondata.img) — same error
4. Full cache/state purge — deleted all of:

• ~/Library/Application Support/Claude/vm_bundles/
• ~/Library/Application Support/Claude/ (all caches, blobs, IndexedDB, sessions, etc.)
• ~/Library/Caches/Claude/
• ~/Library/Logs/Claude/
• ~/Library/Saved Application State/com.anthropic.claudedesktop.savedState/
• ~/Library/WebKit/com.anthropic.claudedesktop/

1. Complete uninstall and fresh reinstall of Claude Desktop — same error
2. Switched Node.js versions (v24 → v22 LTS) — same error
3. Also saw secondary error on retry: RPC error: ensure user: user gallant-sleepy-lovelace already exists unexpectedly: uid=1023 gid=1023 — stale VM users persisting across session disk resets

Log Evidence

From cowork_vm_node.log — consistent pattern across every attempt:

[Spawn:create] id=073f0d50 name=bold-epic-goodall cmd=/usr/local/bin/claude args=--output-format stream-json ...
[Process:073f0d50] Created, name=bold-epic-goodall, total active=1
[Spawn:vm] id=073f0d50 OAuth token approved with MITM proxy
[Process:073f0d50] Spawn confirmed, flushing 2 buffered stdin chunks
[Spawn:vm] id=073f0d50 Spawn succeeded in 40ms
[vm-stderr 073f0d50] Error: Failed to create bridge sockets after 5 attempts
[Process:073f0d50] Exited, code=1, signal=null, duration=760ms, oom=false

This pattern repeats for every single cowork process spawned — both Opus and Haiku models, new tasks only (no resumes).

Suspected Root Cause

macOS 26 (Tahoe) includes changes to Apple's Virtualization.framework. The VM boots and connects fine, but the bridge socket creation (socat bridges from Unix sockets to TCP proxy ports) fails inside the guest VM. This suggests either:

1. A Virtualization.framework behavioral change in macOS 26 affecting Unix socket creation within guest VMs
2. A TMPDIR or filesystem permission change in the guest under the new hypervisor
3. The gvisor user-mode networking layer not being compatible with macOS 26's network stack

Token & Time Cost

Significant debugging effort was spent on this issue — approximately 1 hour of Claude Code CLI time using Opus model with extensive log analysis, multiple file operations, and systematic troubleshooting. We'd appreciate token credits for the time spent diagnosing what turned out to be a platform bug.

What Should Happen?

Starting a new Cowork task should launch successfully, as it does on macOS 15.x (Sequoia).

Steps to Reproduce

1. Use macOS 26.2 (Tahoe) on Apple Silicon
2. Install Claude Desktop (latest)
3. Open Cowork
4. Start any new task
5. Error appears immediately — process exits within ~750ms

Claude Model

claude-opus-4-7 and claude-haiku-4-5-20251001 (both fail)

Is this a regression?

Yes — #19503 was reported fixed in late February 2026, but the fix does not cover macOS 26 (Tahoe)

Last Working Version

Unknown — Cowork has never worked on this macOS 26.2 machine

Claude Code Version

Claude Desktop SDK: 2.1.111 / Claude Code CLI: 2.1.92

Platform

Anthropic API (Max subscription, OAuth)

Operating System

macOS 26.2 Tahoe (Darwin 25.2.0)

Terminal/Shell

zsh

Additional Information

Related issues: #19503, #23746, #11079

cc @MarshallOfSound (closed #19503 as fixed — this is a resurgence on macOS 26)