Problem

Hermes can fail on gpt-5.4 / Codex Responses streams with:

Invalid API response: response.output is empty

This can happen even on very small prompts, and leads Hermes to retry as if the provider returned a malformed or rate-limited response.

Root cause

In _run_codex_stream(), Hermes already tries several recovery paths:

• stream.get_final_response()
• response.output_item.done
• synthesized output from streamed text deltas

However, it does not recover the final response object attached to terminal stream events like response.completed.

If the provider/backend returns the usable final payload on event.response, while get_final_response() returns an object with output=[], Hermes currently treats that as an invalid response and retries.

Fix

• capture event.response for terminal stream events
• prefer that terminal response when get_final_response() has empty output
• add regression coverage for this response shape

User impact

This avoids false malformed/rate-limit retries on gpt-5.4 / Codex Responses streams and restores expected behavior for affected sessions.

Test coverage

Adds a regression test for the case where:

• the stream emits response.completed with a valid response
• get_final_response() returns output=[]
• Hermes should use the terminal event response instead of rejecting it