[macOS] ChatGPT account switching after a long session can hang login and configd, disabling system networking

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

Codex Desktop 26.903.71938 (build 8576), with bundled codex-cli / app-server 0.153.4.

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.6.0 arm64 arm; macOS 26.6.2 (25G83); MacBook Pro Mac14,7; Apple M2; 8 GB RAM; USB Ethernet on interface en8.

What issue are you seeing?

After Codex Desktop has been used continuously for several hours, signing out of one ChatGPT account and signing into another can leave the new login/app launch frozen. At the same time, system networking becomes unusable and does not recover until I restart the Mac.

I have experienced the same symptom pattern three times. I have a confirmed macOS configd watchdog report for the most recent occurrence.

The most recent timeline was:

1. 07:44:12 local time: the long-lived Codex process shut down its app-server transport. During teardown, its IPC router logged write EPIPE.
2. 07:44:25: a new Codex launch started.
3. 07:44:32: the bundled app-server completed initialization, but the log then stopped before the normal main-window/account lookup events seen in a healthy launch.
4. 07:57:15: macOS generated a bug_type 409 watchdog report and terminated configd after its watchdog request handling code had likely been hung for 120 seconds.
5. Networking remained unavailable until I restarted the Mac.
6. 08:02:23: Codex launched normally after the restart.

The crash report is for macOS configd, not a direct ChatGPT/Codex application crash. However, the stackshot captured these relevant facts at the same time:

• All four captured configd threads were in uninterruptible waits on the same kernel read/write lock.
• A Codex (Service) process had been alive for approximately 17 hours.
• kernel_task turnstile data identified that service's Chrome_ChildIOThread as the owner of a kernel mutex blocking two kernel threads, including SOFLOW_GC.
• Another thread in the same Codex service was blocked sending a Mach message to com.apple.SystemConfiguration.configd.
• Thermal pressure was nominal, so the report does not show a thermal event.

This establishes a strong timing and lock-contention correlation, but I am not claiming that the available evidence proves Codex caused the kernel lock condition.

What steps can reproduce the bug?

The failure is intermittent rather than a deterministic minimal reproduction:

1. Use Codex Desktop continuously for roughly three to four hours without intentionally putting the Mac to sleep.
2. In Codex Desktop, sign out of the active ChatGPT account.
3. Sign into a different ChatGPT account.
4. Observe that the new login/app launch can freeze before normal account initialization completes.
5. When the failure occurs, system networking becomes unavailable and requires a Mac restart.

Reproduction history: three occurrences with the same visible symptoms; one occurrence has a matching configd watchdog stackshot.

What is the expected behavior?

Account switching should complete or fail cleanly. Codex service processes should recover during an account switch or relaunch and should not remain involved in a system networking lock condition that requires rebooting the Mac.

Additional information

No VPN or active Docker workload was in use. A dormant com.docker.vmnetd helper was loaded.

A sanitized diagnostic excerpt is attached. The full macOS stackshot and raw Codex logs contain local and account metadata, so I have not posted them publicly; I can provide them privately to OpenAI maintainers if requested.

Related but not exact duplicates:

• #23967 reports temporary local-network latency and packet loss during prompt submission in long conversations. This report instead involves ChatGPT account switching, a stalled relaunch/login, and a watchdog-terminated configd service requiring a reboot.
• #27601 reports Codex Desktop live-state failures after account switching, but does not report a system networking outage or the configd / Chrome_ChildIOThread stack signature.

codex-configd-sanitized.txt

Start with the attached codex-configd-sanitized.txt and the Codex app-server teardown and launch logs around the EPIPE, relaunch, and account lookup events. Compare a long-session account switch with a healthy launch, including the Codex Service and configd interaction. Done means account switching completes or fails cleanly without leaving networking unusable or requiring a reboot.