[macOS] SkyComputerUseService build 26.812.1000717 segfaults at launch; unbounded respawn storm forces WindowServer watchdog reboots of the entire Mac #38697

Environment

• ChatGPT desktop app 1.2026.xxx (app bundle dated 2026-07-24), Codex Computer Use helper 26.812.1000717 (auto-downloaded to ~/.codex/computer-use/)
• macOS 26.5.2 (25F84), MacBook Pro M2 Max (Mac14,6), 32 GB RAM

Summary

Starting 2026-08-14, SkyComputerUseService (com.openai.sky.CUAService) segfaults immediately at launch, and the parent ChatGPT process respawns it with no backoff. The result is a fork storm that force-reboots the entire Mac every ~15 minutes:

• 300–650 concurrent SkyComputerUseService instances, 1,400+ total processes
• load average climbing past 900
• WindowServer hits userspace_watchdog_timeout (.spin reports in /Library/Logs/DiagnosticReports/) and macOS force-restarts — 9 forced reboots in one evening, timestamps matching the watchdog reports exactly

Crash signature

From ~/Library/Logs/DiagnosticReports/SkyComputerUseService-*.ips:

EXC_BAD_ACCESS (SIGSEGV), KERN_INVALID_ADDRESS at 0x0000000000000000
Faulting thread:
  _XXMIGPostNotification
  _XPostNotification
  mshMIGPerform
  __CFRUNLOOP_IS_CALLING_OUT_TO_A_SOURCE1_PERFORM_FUNCTION__
  __CFRunLoopDoSource1
  __CFRunLoopRun
  _CFRunLoopRunSpecificWithOptions

The service crashes before doing any work (null-pointer dereference while hooking into Mach notification handling), so this happens even with every "Computer use" toggle disabled in the app's settings.

Reproduction / evidence

1. Launch the ChatGPT desktop app. Within seconds, SkyComputerUseService instances start crashing and multiplying (ps -axo comm | grep -c SkyComputerUseService climbs into the hundreds).
2. Load average exceeds 800–900 within ~10 minutes; WindowServer watchdog then reboots the machine.
3. Renaming ~/.codex/computer-use/Codex Computer Use.app stops the storm — but the app re-downloads the same broken build on next launch and the loop resumes. A freshly downloaded copy crashes identically, so this looks like a bad helper build published around 2026-08-12 (version string 26.812.1000717) rather than local corruption. macOS 26.5.2 and the app itself were both installed weeks earlier with no issues.

Impact

Machine-wide forced reboots with data loss in all running apps. This is effectively a DoS of the whole system by the respawn loop.

Suggested fixes

1. Pull/replace the broken helper build (server-side fix).
2. Add crash-loop backoff to the helper supervisor in the ChatGPT app — a helper that segfaults N times in a row should stop being respawned instead of taking down the host.

Possibly related (different crash site): #36672.

Full .ips crash reports and WindowServer .spin reports available on request.