Bug type

Regression (worked before, now fails)

Summary

Summary

On macOS, OpenClaw can complete openai-codex OAuth successfully and enumerate Codex models, but actual requests to Codex models never complete. The agent always falls back to Gemini.

Environment

• OpenClaw: 2026.3.7 (GitHub source build)
• macOS (Apple Silicon)
• Proxy: Clash 127.0.0.1:7897
• OpenAI Codex OAuth login succeeds
• Official codex CLI works with the same account and proxy

Steps to reproduce

1. Install OpenClaw from GitHub source so active version is 2026.3.7
2. Set proxy env:
   • HTTP_PROXY=http://127.0.0.1:7897
   • HTTPS_PROXY=http://127.0.0.1:7897
   • ALL_PROXY=http://127.0.0.1:7897
   • NO_PROXY=127.0.0.1,localhost
3. Run:
   • openclaw models auth login --provider openai-codex
4. Complete browser OAuth successfully
5. Set main model to one of:
   • openai-codex/gpt-5.1-codex-mini
   • openai-codex/gpt-5.3-codex
   • openai-codex/gpt-5.4
6. Restart gateway
7. Run:
   • openclaw agent --agent main --message 'ping' --json

Expected behavior

OpenClaw should answer using the configured openai-codex/* model.

Actual behavior

OpenClaw starts with the Codex model configured, but the request does not complete through Codex. The final reply always comes from:

• google-gemini-cli/gemini-3-flash-preview

Evidence

• OAuth succeeds:
  • Auth profile: openai-codex:default (openai-codex/oauth)
• Codex models enumerate correctly:
  • openai-codex/gpt-5.1-codex-mini
  • openai-codex/gpt-5.3-codex
  • openai-codex/gpt-5.4
• Gateway confirms configured model:
  • agent model: openai-codex/gpt-5.4
• But actual agent response metadata shows:
  • provider: google-gemini-cli
  • model: gemini-3-flash-preview

OpenClaw version

2026.3.7

Operating system

macos 14.5

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response