codex - 💡(How to fix) Fix CLI 0.135.0 crashes with STATUS_ILLEGAL_INSTRUCTION on Haswell Windows CPU

Summary

After Codex CLI auto-updated to @openai/[email protected] on Windows, the CLI native binary started crashing immediately on a Haswell-era CPU with Windows exit code -1073741795 (0xC000001D, STATUS_ILLEGAL_INSTRUCTION).

Downgrading to @openai/[email protected] restores startup on the same machine. This looks like a Windows x64 native binary CPU-instruction baseline regression in 0.135.0.

Environment

• Product: Codex CLI installed via npm (@openai/codex)
• Affected version: 0.135.0
• Last known working version on the same host: 0.134.0
• OS: Microsoft Windows NT 10.0.22631.0 x64
• CPU: Intel(R) Xeon(R) CPU E3-1230 v3 @ 3.30GHz
  • Haswell generation, 2013
  • Supports AVX2 but does not support AVX-512
  • 4 cores / 8 logical processors
• Node.js: v24.15.0
• npm: 11.12.1
• Current workaround state: pinned back to codex-cli 0.134.0

Observed behavior

At approximately 2026-05-31 12:38 CST, Codex auto-updated to 0.135.0. After that, the normal launcher could no longer start Codex. It reported:

Codex CLI exited with code -1073741795

-1073741795 is 0xC000001D, i.e. STATUS_ILLEGAL_INSTRUCTION on Windows. The launcher fell back to its recovery shell / cmd prompt.

In local diagnosis, light wrapper paths such as codex --help did not reproduce the immediate crash, but paths that start the real interactive / execution native runtime (codex exec or the normal TUI launcher) crashed immediately.

Expected behavior

@openai/[email protected] should start on Windows x64 machines that satisfy the intended support baseline, or should fail gracefully with a clear CPU capability error before launching the native runtime.

If Codex CLI intends to support older x86-64 CPUs such as Haswell, the Windows x64 binary should not be built with instructions unavailable on that baseline. If Haswell is no longer supported, the installer / launcher should detect and report that explicitly instead of crashing with STATUS_ILLEGAL_INSTRUCTION.

Reproduction outline

I am not reinstalling 0.135.0 on this production host again because it breaks the local agent launcher, but the observed sequence was:

1. Windows x64 host with Intel Xeon E3-1230 v3 / Haswell CPU.
2. Codex CLI auto-updates or is installed as @openai/[email protected].
3. Start Codex through the normal CLI/TUI launcher or execution path.
4. Process exits immediately with Windows code -1073741795 / 0xC000001D.
5. Downgrade to @openai/[email protected] on the same machine.
6. Codex starts again; launcher no longer hits the immediate illegal-instruction crash.

Workaround

Pin Codex CLI to 0.134.0:

npm install -g @openai/codex@0.134.0 --registry https://registry.npmmirror.com
codex --version
# codex-cli 0.134.0

I also added a local launcher pre-flight that checks the installed Codex version and re-pins to 0.134.0 if a later auto-update reappears, because the trigger for the auto-update has not yet been identified.

Related observations

There are existing Windows 0xC000001D / binary startup issues in this repo, but I did not find an exact duplicate for Codex CLI 0.135.0 + Haswell CPU + npm package @openai/codex-win32-x64.

This may be related to a build-toolchain or CPU target change in the 0.135.0 Windows x64 native binary. Please check whether the Windows x64 artifact is being built with a CPU baseline above generic x86-64 / Haswell-compatible instructions, or whether a dependency/runtime path introduced an AVX-512 requirement.