[Windows app] codex.exe repeatedly crashes with 0xC0000005 at fixed RVA 0x819aa97
What version of the Codex App are you using (From “About Codex” dialog)?
Windows Store package 26.803.10989.0; app log client 26.803.81509; bundled Codex 0.147.0
What subscription do you have?
ChatGPT subscription (exact tier not exposed in local app diagnostics)
What platform is your computer?
Microsoft Windows 11 Pro; Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
On Windows 11, the native Codex backend repeatedly terminates while the ChatGPT/Codex desktop shell remains alive and displays:
ChatGPT stopped unexpectedly
Windows Application Error / WER evidence is consistent across every occurrence:
- Faulting process:
codex.exe - Exception code:
0xc0000005(read access violation) - Fault offset:
0x000000000819aa97 - Occurrences observed: 10 over two days
- Same fault offset in all 10 events
- Nine captured minidumps identify the crashing thread as
tokio-rt-worker
Clicking Restart recovers the app temporarily, but the same backend failure recurs. The Electron/ChatGPT shell itself remains running.
What steps can reproduce the bug?
- Install the current Microsoft Store package on Windows 11.
- Launch the ChatGPT/Codex desktop app and use Codex normally, or restart it after the crash-recovery screen appears.
- After a variable interval—from startup to later in a normal session—the
codex.exebackend terminates. - The shell displays "ChatGPT stopped unexpectedly."
- Click Restart; the failure eventually repeats with the same exception and RVA.
The failure is not tied to one prompt or conversation. It has occurred during startup as well as after completed turns.
A Claude Code profile and plugin catalog are present. Immediately before several crashes, structured app logs show activity including externalAgentConfig/detect, externalAgentConfig/import/readHistories, and plugin catalog/list operations. This is a possible trigger area, not a proven reproduction condition.
What is the expected behavior?
The native Codex backend should remain running during normal use. If plugin or external-agent metadata is malformed, it should be rejected with a handled error rather than terminating the process and forcing an app restart.
Additional information
Diagnostics:
- Microsoft Store package:
26.803.10989.0(package status OK) - App log client version:
26.803.81509 - Bundled Codex backend:
0.147.0 - OS: Windows 11 Pro, build 26200, x64
- Both the app and backend have valid OpenAI Authenticode signatures
- No Windows Defender or Code Integrity block was recorded
config.tomlparses successfully- The nine dumps all show a read violation at the same code address, but the invalid target address varies
- Faulting instruction:
movdqu xmm8, xmmword ptr [r8 + r15] - The repeated stack shape reaches Rust string/pattern/indexmap scanning code
- PDB identity:
codex.pdb, GUID1abe15db-ea7b-8db3-4c4c-44205044422e, age 1 - The exact higher-level function requires OpenAI's private symbols
Related issue #33981 is not an exact match: it involves an older build, an immediate ChatGPT.exe startup exit, and no Event 1000/WER report. This report concerns the child codex.exe backend on the current build, with a repeatable RVA and minidumps.
Raw minidumps and the structured log database are available for private transfer to OpenAI, but are intentionally not attached to this public issue because they may contain sensitive local/session data.
1 Comment
Fixed fault offset but the invalid target address varies every time...... that usually reads as memory corruption, not a plain logic bug. Worth ruling out RAM stability (XMP off, memtest) and doing a clean reinstall of the Store package. If it still crashes the same way on a different machine, the minidumps you've got are solid evidence for OpenAI.