Summary

Fixes a Codex streaming edge case where the OpenAI SDK final response parser can raise:

TypeError: 'NoneType' object is not iterable

This was observed with the openai-codex provider against https://chatgpt.com/backend-api/codex. The stream had already produced usable assistant output, but stream.get_final_response() failed while parsing the final response. Hermes then treated the turn as a non-retryable provider/client failure even though output was available.

What changed

• Adds a recovery path around Codex stream.get_final_response() for the specific NoneType final parse error.
• If response.output_item.done events were collected, Hermes synthesizes a completed response from those output items.
• If no output items exist but text deltas were streamed and no tool calls are active, Hermes synthesizes a completed assistant message from the accumulated text deltas.
• Keeps unrelated TypeErrors unchanged by re-raising them.

Why

The Codex backend can stream valid output items/text before the SDK final parser fails. In that case, dropping the turn is worse than using the already streamed output, especially because the user may see the text but Hermes still marks the request as failed.

Observed log shape:

API call failed (attempt 1/3): TypeError
Provider: openai-codex  Model: gpt-5.5
Endpoint: https://chatgpt.com/backend-api/codex
Error: 'NoneType' object is not iterable
Non-retryable client error: 'NoneType' object is not iterable

After the fix, the same failure mode is logged as a warning and Hermes continues with the collected streamed output.

Tests

• source venv/bin/activate && python -m pytest tests/run_agent/test_run_agent_codex_responses.py -q
  • 67 passed
• source venv/bin/activate && python -m pytest tests/run_agent/test_streaming.py tests/run_agent/test_codex_xai_oauth_recovery.py -q
  • 77 passed