fix(agent): close codex app-server sessions

[changeroa]

Summary

• close attached Codex app-server sessions when cached AIAgent instances release clients or close
• treat closed app-server stdin during turn/interrupt as non-fatal so timeout retirement can complete
• add regressions for release/close cleanup and broken-pipe interrupt retirement

Tests

• /Users/victor/.hermes/hermes-agent/venv/bin/python -m pytest -o addopts='' tests/run_agent/test_codex_app_server_integration.py::TestApiModeAccepted::test_agent_close_closes_attached_codex_app_server_session tests/run_agent/test_codex_app_server_integration.py::TestApiModeAccepted::test_agent_release_clients_closes_attached_codex_app_server_session tests/agent/transports/test_codex_app_server_session.py::TestRunTurn::test_deadline_interrupt_broken_pipe_still_retires_turn -q
• /Users/victor/.hermes/hermes-agent/venv/bin/python -m pytest -o addopts='' tests/run_agent/test_codex_app_server_integration.py tests/agent/transports/test_codex_app_server_session.py -q
• /Users/victor/.hermes/hermes-agent/venv/bin/python -m py_compile run_agent.py agent/transports/codex_app_server_session.py tests/run_agent/test_codex_app_server_integration.py tests/agent/transports/test_codex_app_server_session.py

[tonydwb]

Code Review Summary\n\nVerdict: LGTM\n\nFixes Codex app-server session lifecycle: closes attached sessions on agent release/close, treats broken-pipe during turn/interrupt as non-fatal. 4-file change with dedicated regression tests.\n\n### Looks Good\n- Well-scoped fix with clear root cause\n- Regression tests cover both cleanup paths and broken-pipe retirement\n- Minimal source changes (+18 lines across 2 files)\n---\nReviewed by Hermes Agent

[tonydwb]

Code Review: LGTM

Ensures codex app-server sessions are properly closed when the agent closes or releases clients. Extracts self._client and self._thread_id into local variables to avoid stale references during the turn loop. Adds RuntimeError handling for turn/interrupt when the transport is closed. Good test coverage (3 new tests: deadline interrupt broken pipe, close during turn, agent close/release integration).

Verdict: LGTM -- thorough resource lifecycle fix with tests.

[teknium1]

Thanks for the focused lifecycle fix. The premise still holds on current main: run_agent.py:3466-3473 and run_agent.py:3522-3529 close only the OpenAI/httpx client and leave _codex_session attached, while agent/transports/codex_app_server_session.py:300-310 can clear the session fields concurrently with run_turn() dereferences at :400-415 and :455-465.

The broken-pipe handling is also warranted: _issue_interrupt() currently catches only CodexAppServerError and TimeoutError (agent/transports/codex_app_server_session.py:788-801), whereas the transport raises RuntimeError for closed stdin/client states (agent/transports/codex_app_server.py:296-307). The added stable references and cleanup coverage match those current paths.

Automated hermes-sweeper review.

[jeff-mettel]

Link for discoverability: this PR is the closest match for #72548, which currently shows no linked PR in its timeline.

That issue asks specifically for _codex_session to be closed and cleared from both release_clients() and close(), with the reference cleared even when transport cleanup raises. This is the AIAgent-layer PR of the four open ones (#62105 / #62112 are the cron layer, #73429 the gateway/webhook layer), so it is the one that lines up.

---

Filed by an AI agent (Claude Opus 5) operating autonomously on @jeff-mettel's behalf.