›_ Codex Issues
Codex Issues / #31353

Windows Desktop crash/hang with codex.exe MoBEX and node_repl/extension-host BEX64 after recent update

Open 💬 2 comments Opened Jul 7, 2026 by bpc-oss
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

Summary

Codex Desktop on Windows started crashing/hanging repeatedly during normal use. The workflow had been stable for several days while using the bundled chrome, browser, computer-use, and node_repl capabilities. The instability began suddenly on July 6-7, 2026.

This does not look like a single conversation corrupting state. Windows Error Reporting shows repeated codex.exe MoBEX crashes first, followed by a cascade involving node_repl.exe, extension-host.exe, and the main Codex app.

Environment

  • OS: Windows 11, x64, build family 10.0.26100
  • Current installed package: OpenAI.Codex_26.623.19656.0_x64__2p2nqsd0c76g0
  • Previously crashing package in WER: OpenAI.Codex_26.623.13972.0_x64__2p2nqsd0c76g0
  • Local CLI reported earlier: codex-cli 0.130.0-alpha.5

Crash / Hang Timeline

Recent Windows Error Reporting entries:

  • 2026-07-06 19:50-23:47: repeated MoBEX for OpenAI.Codex_26.623.13972.0_x64__2p2nqsd0c76g0, P5: codex.exe, P9: c0000409, P10: 0000000000000007
  • 2026-07-07 00:09-00:44: repeated same codex.exe MoBEX / c0000409
  • 2026-07-07 08:35 and 09:42: same codex.exe MoBEX / c0000409
  • 2026-07-07 09:58:
  • node_repl.exe crashed with BEX64, P10: 0000000000000007
  • extension-host.exe crashed with BEX64, P10: 0000000000000007
  • Codex main app reported MoAppCrash, module KERNELBASE.dll
  • Codex app then reported MoAppHang
  • 2026-07-07 11:19: Microsoft Store/AppX updated Codex from 26.623.13972.0 to 26.623.19656.0
  • 2026-07-07 11:23: bundled plugin cache folders for browser, chrome, and computer-use were updated locally

What I Tried / Ruled Out

  • Rotated a bloated/empty ~/.codex/logs_2.sqlite state file. This reduced disk pressure but did not explain the crash sequence.
  • Observed high-frequency inserts into ~/.codex/logs_2.sqlite, then temporarily installed a SQLite trigger blocking inserts into logs and checkpointed/truncated WAL. After sampling, MAX(id), row count, and WAL size stopped growing, but Codex Desktop still crashed/hung.
  • A very large conversation file was compacted from about 3.2 GB to a recent-turn subset and validated as well-formed JSONL. This reduced load pressure but did not eliminate the app instability.
  • The same user workflow was stable before July 6-7 with chrome, browser, computer-use, and node_repl, so the likely boundary is the Windows desktop app/runtime/plugin-host layer rather than the mere use of those capabilities.

Available Local Artifacts

Crash dumps exist locally under:

  • %LOCALAPPDATA%\CrashDumps\Codex.exe.3268324.dmp
  • %LOCALAPPDATA%\CrashDumps\codex.exe.3094072.dmp
  • %LOCALAPPDATA%\CrashDumps\codex.exe.3074908.dmp
  • several more codex.exe.*.dmp files from 2026-07-07

I can provide redacted WER details or crash dumps if maintainers want a specific artifact.

Expected Behavior

Codex Desktop should not crash or hang when opening/continuing a thread or when the bundled node_repl / browser-related plugin hosts are active.

Actual Behavior

Codex Desktop repeatedly crashes or becomes unresponsive. Windows records MoBEX / c0000409 for codex.exe, followed by BEX64 crashes in node_repl.exe and extension-host.exe, then MoAppCrash / MoAppHang for the main app.

Related Issues

This seems related to the Windows crash/performance reports in:

  • #28092
  • #26812
  • #16374
  • #29177
  • #27741
  • #28224

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #31115
  • #30178
  • #30558
  • #30925

Powered by Codex Action

Bricklink · 1 month ago

Adding a data point with the same signature, plus a set of local causes I was able to eliminate through testing.

Environment

  • Windows 11 Home 10.0.26100
  • Codex desktop 26.715.2305.0 (MSIX). Same behavior on 26.707.9981.0 before updating.
  • NVIDIA RTX 3090, driver 32.0.15.9597 (2026-03), no TDR or display events logged
  • codex CLI 0.144.1 on the same machine is completely unaffected

Timeline
Started the evening of 2026-07-15, immediately after Windows updates KB5101001, KB5101650, and KB5120102 were installed. Windows Error Reporting shows repeated Event 1000 crashes for ChatGPT.exe that evening with exception code 0xc06d007f and faulting module "unknown" (sample Report Id 4bc48318-5853-4d34-8cd3-ed5ec0577953), then Event 1002 AppHang entries on 2026-07-16 where Windows force-closed the app twice within one minute.

Current behavior on 26.715
The main window process is alive but oscillates between Responding and Not Responding indefinitely. Sampling (Get-Process).Responding every 10 seconds for 3 minutes gives patterns like HHHRHHHHHHRHHHHHRH (H = not responding, R = responding). It never settles, and CPU stays elevated on the main process while hung. No child processes (node_repl, extension-host, gpu) crash during the loop.

Eliminated locally, each verified with the same 10s sampling method

  • --disable-gpu launch: identical hang pattern
  • Clearing Cache, Code Cache, GPUCache, DawnGraphiteCache, and DawnWebGPUCache: identical hang pattern
  • GPU contention: GPU was at 19% utilization and 2.3/24GB during the hangs
  • Persisted state restore: removing three thread-browser-tabs-v1:* keys and trimming a 946KB prompt-history blob inside .codex-global-state.json improved time-to-first-responsive from roughly 60s to 8s, but the oscillating hang loop continued afterward

The prompt-history observation may be worth a look independently: it appears the full typed-prompt history for every thread is serialized into electron-persisted-atom-state, and ours had grown to nearly 1MB, which seems likely to be synchronously rewritten on the main thread.

Happy to pull WER report contents or run an instrumented build if that would help.