[Linux/X11][Pantheon] Codex coincides with repeated Gala compositor crashes and GPU fatal

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

26.901.20858

What subscription do you have?

Unknown / not included in report

What platform is your computer?

Linux 7.0.0-31-generic x86_64 x86_64 — elementary OS 8.1, Pantheon X11

What issue are you seeing?

Using the Codex/ChatGPT Linux desktop app repeatedly coincides with the Pantheon Gala compositor crashing. Within 2–3 seconds, Codex and Google Chrome also terminate, disrupting the desktop session.

The ordering in systemd-coredump is Gala first, then Chromium-based applications. This shows a strong timing correlation but does not establish whether Codex triggers Gala or is a secondary victim.

Immediately after Gala fails, Codex logs repeated renderer/zygote and GPU failures ending with:

Network service crashed or was terminated, restarting service.
Failed to send GetTerminationStatus message to zygote
Electron renderer process gone reason=clean-exit exit_code=0
GPU process launch failed: error_code=1002
FATAL: GPU process isn't usable. Goodbye.

This occurred at least four times on 2026-09-07. No out-of-memory event was found.

What steps can reproduce the bug?

The failure is intermittent; I have not isolated a single keystroke or UI action.

1. Run Codex desktop 26.901.20858 on elementary OS 8.1 / Pantheon X11 with AMD Radeon 680M (amdgpu, Mesa 25.2.8).
2. Use the app normally with hardware acceleration enabled.
3. Gala crashes with SIGSEGV; 2–3 seconds later ChatGPT/Codex and Chrome terminate with SIGTRAP.

Observed clusters (local time UTC+08):

• 08:20:55 Gala SIGSEGV; 08:20:58 Chrome and ChatGPT SIGTRAP
• 08:25:31 Gala SIGSEGV; 08:25:33 Chrome SIGTRAP; 08:25:34 ChatGPT SIGTRAP
• 08:42:00 Gala SIGSEGV; 08:42:03 Chrome and ChatGPT SIGTRAP
• 18:14:26 Gala SIGSEGV; 18:14:28 Chrome SIGTRAP; 18:14:29 ChatGPT SIGTRAP; 18:14:38 Gala SIGSEGV

What is the expected behavior?

Codex should remain stable and must not destabilize the compositor or desktop session. If Gala independently fails, Codex should recover its GPU/renderer processes gracefully instead of terminating.

Additional information

Environment:

• App package: chatgpt 26.901.20858 amd64
• Apt candidate available but not yet tested: 26.901.51231
• Codex CLI: 0.153.2
• GPU: AMD Radeon 680M (Rembrandt), amdgpu
• Mesa/OpenGL: Mesa 25.2.8, OpenGL 4.6, direct rendering enabled
• RAM: 19 GiB; about 15 GiB available when inspected; no swap pressure

The latest ChatGPT main process produced a 41.7 MiB systemd core dump after SIGTRAP. It was not attached because core memory may contain private data; the stripped local stack only contains ChatGPT offsets (top frame ChatGPT+0xcbfbe35).

Before the latest failure, the renderer repeatedly logged ResizeObserver loop completed with undelivered notifications.

Please investigate Electron/Chromium GPU-process compatibility with Pantheon/Gala on AMD/Mesa under X11. Is there a supported way to disable hardware acceleration for diagnosis, and does 26.901.51231 contain a relevant fix?

Privacy: account details, conversation/thread/request IDs, hostname, username, machine ID, and core memory are omitted.

Start by reproducing the failure in the reported elementary OS 8.1/Pantheon X11 environment with AMD Radeon 680M, Mesa 25.2.8, and hardware acceleration enabled. Review the Gala, Chromium, renderer, zygote, and GPU failure logs, then compare behavior with ChatGPT 26.901.51231. Done means the crash relationship is characterized, a supported acceleration diagnostic is identified, or the relevant Electron/Chromium compatibility fix is documented.