What happened

A valid empty final response from openai-codex is being treated as incomplete by Hermes.

In a Telegram group, I told the bot not to reply. Instead of staying silent, Hermes retried 3 times and then sent:

⚠️ Codex response remained incomplete after 3 continuation attempts

Expected

If Codex completes with an empty final answer, Hermes should preserve that completion state instead of treating it as incomplete.

Actual

Hermes normalizes the response to finish_reason = "incomplete" and enters the continuation loop.

Repro

I reproduced this with the live Codex Responses API and the same prompt/history.

Provider result:

• response.status = completed
• one message item
• message.status = completed
• message.phase = final_answer
• empty final text

Hermes result:

• _run_codex_stream() receives completed
• _normalize_codex_response() returns finish_reason = "incomplete"

Suspected cause

In run_agent.py, _normalize_codex_response() has this branch:

elif reasoning_items_raw and not final_text:
    finish_reason = "incomplete"

That appears to misclassify a completed empty final-answer turn as incomplete.