Duplicate of #32884 (same root cause: Codex Responses stream null output, #11179). Also competing with #32890 and #32891. All four PRs fix codex_runtime.py for the same SDK TypeError: 'NoneType' object is not iterable; #32888 and #32890 also fix auxiliary_client.py.

Tested. Confirms the fix.

Cherry-picked your commit (32a19b11c) onto Hermes v0.14.0 (main at 30928f945 from May 25) running against the official chatgpt.com/backend-api/codex endpoint with gpt-5.5. Same Telegram session that had been failing repeatedly with 'NoneType' object is not iterable on long-thread + post-compression turns is now landing the model's reply cleanly.

The dual-path coverage matters: I had been seeing agent.context_compressor: Failed to generate context summary fire on the same TypeError ~5s before the main-path failure (so the bug breaks the compressor's own LLM call too, not just conversation_loop). PRs that only patch codex_runtime.py would still let the secondary failure through. Including auxiliary_client.py in this fix means both paths recover.

Runtime evidence:

• ✅ from agent.codex_runtime import run_codex_stream clean
• ✅ from agent.auxiliary_client import ... clean
• ✅ The 3 added regression tests pass under pytest -k 'null or recover'
• ✅ End-to-end: a follow-up turn on the previously-broken Telegram session completed with the right answer (QB Time tool call + summary), no more Non-retryable client error in agent.log

This is one of 8 racing PRs for the same bug. Of the dual-path fixes (yours, #32890, #11182), yours has the smallest diff (+158/-15 vs +250/-62 and the 4-month-stale #11182). Cleanest path to merge IMO.

Will run for a day; will report back if anything regresses. Thanks for shipping this fast.

Closing as duplicate — the Codex null-output fix has been merged via #32963 (cherry-picked from @carltonawong's PR #32890, the one Gille reviewed). Thanks for jumping on the outage so quickly; appreciate the help. Closes #11179.