Thanks, this was a real remaining path. I cherry-picked your follow-up as 9a9d46f, preserving authorship. Re-ran targeted canonical validation: scripts/run_tests.sh tests/run_agent/test_run_agent_codex_responses.py -- -q -> 71 passed. Live check after the follow-up: hermes chat -Q --provider openai-codex -m gpt-5.5 -q \"Reply exactly: LATEST_GPT55_FINAL_OK\" -> LATEST_GPT55_FINAL_OK.

Linux 修复命令，直接终端运行即可，然后重启gateway

`
cd /home/admin/.hermes/hermes-agent

0. 备份

cp venv/lib/python3.11/site-packages/openai/lib/_parsing/_responses.py
/tmp/_responses.py.bak.$(date +%Y%m%d-%H%M%S)

cp agent/transports/codex.py
/tmp/codex.py.bak.$(date +%Y%m%d-%H%M%S)

1. Patch OpenAI SDK：给 response.output 加空值保护

python3 - <<'PY'
from pathlib import Path

p = Path("venv/lib/python3.11/site-packages/openai/lib/_parsing/_responses.py")
src = p.read_text()

old = "for output in response.output:"
new = "for output in (response.output or []):"

if new in src:
print("✓ SDK patch already applied")
elif old in src:
p.write_text(src.replace(old, new, 1))
print("✓ SDK patch applied")
else:
raise SystemExit("✗ SDK patch failed: target line not found")
PY

2. Patch Hermes transport：避免 tools=None 传进 SDK

python3 - <<'PY'
from pathlib import Path

p = Path("agent/transports/codex.py")
src = p.read_text()

old = (
' "tools": response_tools,\n'
' "store": False,\n'
' }\n'
' if response_tools:\n'
' kwargs["tool_choice"] = "auto"\n'
' kwargs["parallel_tool_calls"] = True'
)

new = (
' "store": False,\n'
' }\n'
' if response_tools:\n'
' kwargs["tools"] = response_tools\n'
' kwargs["tool_choice"] = "auto"\n'
' kwargs["parallel_tool_calls"] = True'
)

if new in src:
print("✓ Transport patch already applied")
elif old in src:
p.write_text(src.replace(old, new, 1))
print("✓ Transport patch applied")
else:
print("⚠ Transport patch skipped: source differs or already patched")
PY

3. 验证关键行

grep -n "for output in (response.output or [])"
venv/lib/python3.11/site-packages/openai/lib/_parsing/_responses.py

grep -n '"tools": response_tools' agent/transports/codex.py || true
`

Thank you @serejaris ! I applied this patch with 2 Hermes installs and I'm back up and running with Hermes on codex again.

For the SDK-level root cause isolated here, I opened an upstream OpenAI Python SDK follow-up:

openai/openai-python#3315

The Hermes-side canonical fix in this PR recovers usable streamed output when the final Codex Responses parse hits response.output=None. The upstream SDK invariant is narrower: openai-python should not leak raw TypeError: 'NoneType' object is not iterable when parser/accessor code sees a Responses object with missing/null output.

Terminology note: this is the OpenAI Python SDK (openai-python) layer rather than the OpenAI Agents SDK layer.

Non-blocking suggestion: it may be worth adding one regression for the post-SDK-fix path where stream.get_final_response() succeeds but returns output=[] or output=None after text deltas were already streamed.

The implementation already handles _out is None or empty, so this would mostly document compatibility with a future openai-python fix that stops raising TypeError and instead exposes an empty final output.

This should keep #32884 explicitly covered for the SDK-layer follow-up in openai/openai-python#3315.

Added the regression in 21503a9: test_run_codex_stream_recovers_when_final_response_output_is_empty_after_deltas covers both output=None and output=[] after streamed text deltas when get_final_response() succeeds.

Validation: scripts/run_tests.sh tests/run_agent/test_run_agent_codex_responses.py -- -q -> 73 passed.

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.