openclaw - 💡(How to fix) Fix Windows: spawn codex ENOENT - codex/* models unusable due to raw spawn() not resolving PATHEXT [2 comments, 2 participants]

Summary

On Windows, the gateway fails to spawn the codex CLI when using a codex/* model (e.g. codex/gpt-5.5). Every request errors with spawn codex ENOENT and falls back to another model, even though codex is installed globally via npm i -g @openai/codex and is resolvable from the same shell that started PM2.

Root cause appears to be child_process.spawn('codex', …) without shell: true on Windows. Node on Windows only resolves .exe via raw spawn; npm global bins ship as codex (shell script), codex.cmd, and codex.ps1, so raw spawn('codex') can never find an executable and throws ENOENT.

Environment

• OS: Windows 11
• OpenClaw: 2026.4.22 (00bd2cf)
• Node: via nvm / system (npm global prefix: C:\Users\<user>\AppData\Roaming\npm)
• @openai/codex: 0.124.0 (installed globally)
• where.exe codex → codex, codex.cmd, codex.ps1 present in %APPDATA%\npm (no .exe)
• codex --version works from PowerShell: codex-cli 0.124.0
• PM2: launched via pm2 start ecosystem.gateway.json from PowerShell where codex is on PATH

Repro

1. Windows + PowerShell, npm i -g @openai/codex, confirm Get-Command codex resolves to codex.ps1 / codex.cmd.
2. openclaw config set agents.defaults.model.primary codex/gpt-5.5
3. pm2 kill; pm2 start ecosystem.gateway.json (so child inherits fresh PATH).
4. Gateway boots clean: [gateway] agent model: codex/gpt-5.5.
5. Send any request:

   openclaw agent --session-id t1 --message "reply: hi" --thinking off --json

6. Request falls back to github-copilot/gpt-5.4; gateway error log shows:

[diagnostic] lane task error: lane=main durationMs=14633 error="Error: spawn codex ENOENT"
[diagnostic] lane task error: lane=session:agent:main:explicit:<sid> durationMs=14638 error="Error: spawn codex ENOENT"
[model-fallback/decision] decision=candidate_failed requested=codex/gpt-5.5 candidate=codex/gpt-5.5 reason=unknown next=github-copilot/gpt-5.4 detail=spawn codex ENOENT

Related (embedded ACP probe path, same underlying issue):

[plugins] embedded acpx runtime backend probe failed: embedded ACP runtime probe failed
  (agent=codex; command='C:\Users\<user>\.openclaw\acpx\codex-acp-wrapper.mjs';
   cwd=...; spawn EFTYPE)

Expected

On Windows, spawning codex should resolve codex.cmd via PATHEXT (same way where codex does), so codex/* models work out of the box after npm i -g @openai/codex.

Suggested fix

In the Codex harness / acpx wrapper spawn sites, one of:

1. Pass shell: true to child_process.spawn on Windows (simplest, but has quoting caveats).
2. Use cross-spawn (respects PATHEXT, no shell, safest for arg escaping).
3. Manually resolve the binary using PATHEXT before spawning:

   const bin = process.platform === 'win32'
     ? (which.sync('codex.cmd', { nothrow: true }) || which.sync('codex', { nothrow: true }))
     : 'codex';
   spawn(bin, args);

Option 2 (cross-spawn) is the usual answer here and also fixes the spawn EFTYPE seen on the .mjs wrapper path (that one likely needs node <wrapper>.mjs instead of invoking the .mjs directly on Windows).

Workaround

None that doesn't involve shimming a codex.exe onto PATH or running the gateway under WSL.

Impact

All codex/* models (gpt-5.5, gpt-5.4-pro, gpt-5.4, gpt-5.3-codex-spark, gpt-5.2-codex) are unusable as primary or successful-fallback on Windows unless the user manually creates a .exe shim. Requests appear to "work" only because fallback silently routes to github-copilot models.