Sharing an OpenClaw → codex-lb integration experience

[PYXXXX]

Hi, thanks again for codex-lb — it has been very useful.

I wanted to share an OpenClaw integration experience in case it is useful for docs or for other users trying a similar setup.

I initially followed the generic OpenAI-compatible route and connected OpenClaw to codex-lb through:

• http://127.0.0.1:2455/v1

That path does work, and it makes sense as the default recommendation because it is the simplest OpenAI-compatible setup.

However, after real OpenClaw-side testing, I ran into some compatibility issues on that route, especially around usage/accounting behavior.

Because of that, I looked at how OpenClaw natively talks to Codex in its OAuth-based flow, and traced it back to the openai-codex-responses path. From there, I tried aligning the chain more closely like this:

• OpenClaw
• openai-codex-responses
• codex-lb
• /backend-api/codex

That turned out to be a better fit for Codex-style behavior.

During that process, I also found an important client-side compatibility issue on the OpenClaw side:

• OpenClaw’s stock openai-codex-responses implementation assumes an official ChatGPT/Codex JWT
• it tries to forcibly extract chatgpt_account_id
• and it assumes chatgpt-account-id should always be sent

That works for the official ChatGPT/Codex backend, but it does not work well with third-party Codex-compatible backends such as codex-lb, where opaque bearer tokens may be valid and chatgpt-account-id may not be required.

What worked well for me was a narrow client-side adjustment:

• keep official chatgpt.com behavior strict
• but for non-official Codex-compatible base URLs, allow bearer-token passthrough even when chatgpt_account_id cannot be extracted
• only send chatgpt-account-id when an actual account ID was extracted

After that adjustment, the OpenClaw → openai-codex-responses → codex-lb → /backend-api/codex path behaved much better for my Codex-style workflow than the generic /v1 path.

Just sharing this in case it is useful for future OpenClaw integration notes or for other users trying the same route.