[macOS] SkyComputerUseService SIGSEGV in AsyncStream.yield leaves Chrome backend unavailable

What version of the app are you using?

• ChatGPT/Codex Desktop: 26.727.51351 (build 6119)
• Computer Use / SkyComputerUseService: 26.727.1000550 (build 1000550)
• ChatGPT Chrome Extension: 1.2.27236.6274
• Google Chrome: 151.0.7922.72

What platform is your computer?

• macOS 26.6 (25G72)
• Apple Silicon / arm64

What issue are you seeing?

The external Chrome backend intermittently becomes unavailable even though Chrome, the ChatGPT Chrome Extension, and the Native Messaging host all pass the bundled diagnostics.

The browser runtime failed immediately with:

Browser is not available: chrome

A retry after two seconds failed identically. Later, without changing the extension or native-host configuration, the Chrome backend became available again, so the failure is intermittent.

A macOS diagnostic report at 2026-08-03 02:48:46 +0200 shows that SkyComputerUseService crashed with:

Exception: EXC_BAD_ACCESS
Signal: SIGSEGV
Subtype: KERN_INVALID_ADDRESS at 0x0000000000000000
Termination: Segmentation fault: 11
Faulting thread: com.apple.main-thread

The faulting stack begins with:

AsyncStream._Storage.yield(_:)
AsyncStream.Continuation.yield(_:)
<two unsymbolicated SkyComputerUseService frames>
notify_datagram_handler(...)
CGSDatagramReadStream::dispatchMainQueueDatagrams()

This is not the Outlook recursion/stack-exhaustion signature from #20683: there is no SIGBUS, ___chkstk_darwin, or excessive recursion.

Diagnostics

The bundled Chrome diagnostics all passed:

• Google Chrome is running.
• The expected extension is installed and enabled in the selected Default profile.
• The native-host manifest exists and is valid.
• The expected origin chrome-extension://hehggadaopoacecdllhhajmbjkdcmajg/ is present.
• No extension disable reason or enterprise-policy block was reported.

The system was not under memory pressure at the time of follow-up diagnosis.

Three old ChatGPT cua_node kernel processes were also still running as PPID 1 orphans from previous sessions. This may be related to the lifecycle-cleanup family reported in #25744 and #26293, but it does not prove the cause of the SIGSEGV.

Steps to reproduce

The exact trigger is not yet deterministic:

1. Use ChatGPT/Codex Desktop with the Chrome extension and Computer Use enabled.
2. Keep several desktop tasks/browser-control sessions active over time.
3. Attempt to acquire the external Chrome backend.
4. Observe Browser is not available: chrome.
5. Check macOS Diagnostic Reports for a SkyComputerUseService crash.

The failure reproduced twice in the same task, including the documented two-second retry. The backend later recovered without reinstalling the extension or changing the native-host manifest.

Expected behavior

A correctly installed and enabled Chrome extension/native host should remain discoverable. SkyComputerUseService should not crash while dispatching a window/event notification, and a helper crash should be recovered without leaving browser backend state unavailable.

Additional information

No exact public match was found for the AsyncStream._Storage.yield plus null-address signature.

Possibly related, but not duplicates:

• #29389 and #25811 — Chrome backend unavailable despite healthy extension/native-host setup
• #23992 — different SkyComputerUseService EXC_BAD_ACCESS / SIGSEGV
• #25744 and #26293 — Computer Use / Node helper lifecycle leaks and PPID 1 orphans
• #36459 — same helper build 1000550, different noWindowsAvailable failure
• #36461 — same desktop build/macOS version, different Computer Use persistence failure

I can provide the complete sanitized .ips crash report privately if needed.