doctor --fix migrates working openai-codex + PI runtime setup to codex runtime

Summary

openclaw doctor --fix currently migrates a working openai-codex/* configuration to openai/* with the Codex runtime. That may be intended for the newer native Codex path, but it breaks or invalidates a setup where ChatGPT+/Codex OAuth is intentionally used together with the OpenClaw PI runtime.

I would like OpenClaw to support or at least not overwrite this configuration:

• model/provider route: openai-codex/gpt-*
• auth: openai-codex:<account>
• runtime: agentRuntime.id = "pi" on the model entries

This setup allows Codex OAuth / ChatGPT+ usage while still using the PI runtime.

Environment

• OpenClaw version: 2026.5.18 Stable
• macOS
• Codex OAuth / ChatGPT+ account configured

Current behavior

After running:

openclaw doctor --non-interactive --fix

OpenClaw migrates the config away from openai-codex/* toward openai/* with the Codex runtime.

In my case this is not the desired state. The working local setup is:

primary:   openai-codex/gpt-5.5
fallback:  openai-codex/gpt-5.4
manual:    openai-codex/gpt-5.4-mini

agentRuntime:
  openai-codex/gpt-5.5      -> pi
  openai-codex/gpt-5.4      -> pi
  openai-codex/gpt-5.4-mini -> pi

auth.order.openai:
  - openai-codex:<account>

With that setup, status/usage shows the expected Codex OAuth usage while the runtime remains OpenClaw Pi Default.

Expected behavior

There should be a supported way to use Codex OAuth / ChatGPT+ auth with the PI runtime without doctor --fix rewriting it as legacy or invalid config.

Possible fixes could be one of:

1. Treat openai-codex/* + agentRuntime.id="pi" as a valid configuration.
2. Add an explicit config flag or migration mode for "use Codex OAuth with PI runtime".
3. Make doctor --fix warn but not rewrite this setup automatically.
4. Document the intended supported replacement if this path should no longer be used.

Why this matters

The current migration makes it hard to keep a stable setup where:

• billing/auth goes through Codex OAuth / ChatGPT+ rather than normal OpenAI API budget
• the agent runtime remains PI
• openclaw doctor --fix can still be used safely after updates

Right now I need a local guard script to undo the doctor migration after updates, which suggests the migration behavior is too aggressive for this valid use case.

Question

Is using Codex OAuth / ChatGPT+ with the PI runtime supposed to be supported? If yes, can doctor preserve this setup? If no, what is the recommended supported configuration for this use case?