[macOS 26.6.2][ChatGPT Desktop 26.814.41957 Build 6744] Computer Use worker/thread leak crashes GUI after ~104 seconds

What version of the Codex App are you using (From “About Codex” dialog)?

26.814.41957

What subscription do you have?

gpt pro 20x

What platform is your computer?

Darwin 25.6.0 arm64 arm

What issue are you seeing?

Environment:

• Hardware: Mac16,7
• Architecture: Apple Silicon ARM64
• macOS: 26.6.2 (25G82)
• ChatGPT/Codex Desktop: 26.814.41957 (Build 6744)
• Bundle ID: com.openai.codex
• Crash time: 2026-08-19 11:33:05 +0800
• Launch time: 2026-08-19 11:31:21 +0800

Observed behavior:
The desktop GUI crashes approximately 104 seconds after launch.

Crash signature:

• Triggered by Thread 328: computer-use
• EXC_BREAKPOINT (SIGTRAP)
• Dispatch Thread Soft Limit Reached: 70
• Too many dispatch threads blocked in synchronous operations

Thread evidence:

• 185 threads named "computer-use"
• 70 com.apple.root.user-initiated-qos dispatch threads
• Repeated blocking stack:

_dispatch_group_wait_slow
AESendMessage
NSAppleEventDescriptor sendEventWithOptions:timeout:error:
sky.node
RemoteHostedPIPContentService
sendBootstrapToServiceWithProcessIdentifier:rendezvousPort:attempt:

The computer-use workers appear to be created repeatedly without the
previous workers being cancelled or destroyed. Each worker appears to
create a Node/V8 isolate and then blocks in the synchronous Apple Event
bootstrap path.

This issue previously reproduced on 26.810 builds and still reproduces
on 26.814.41957 and macOS 26.6.2.

Please investigate:

1. Missing single-flight protection around Computer Use bootstrap.
2. Missing cancellation/timeout cleanup for AESendMessage.
3. Worker/V8 isolate lifecycle cleanup after bootstrap failure.
4. Whether Computer Use initializes before config.toml or
   requirements.toml feature gating is applied.
5. Whether RemoteHostedPIPContentService retries create a new worker
   instead of reusing one service instance.

What steps can reproduce the bug?

This is bare‑metal native macOS, NOT Docker / container / VM environment.

Steps: Start ChatGPT desktop app, run multiple Computer‑Use agent tasks via ao cli in local terminal.
After each task finishes, SkyComputerUseService helper becomes orphaned (ppid=1) and keeps accumulating.
Leak is more severe when cache resides on external disk volume.

What is the expected behavior?

No response

Additional information

No response

Start with the Computer Use bootstrap path and RemoteHostedPIPContentService, then reproduce by running multiple Computer-Use agent tasks through the ao cli on bare-metal macOS. Trace worker creation, AESendMessage blocking, and bootstrap failure cleanup; done means repeated tasks no longer accumulate orphaned SkyComputerUseService workers or crash the GUI.