fix(cron): close codex app-server session when a job run finishes

[Shaance]

What does this PR do?

Fixes a process leak when model.openai_runtime: codex_app_server is active: every cron run inside the long-lived gateway process could leave behind one codex app-server subprocess and its MCP children.

The fix now follows the cron agent's existing lifecycle boundary:

• agent/codex_runtime.py provides an idempotent, exception-safe close_codex_session(agent) helper.
• cron/scheduler.py invokes that helper from _teardown_cron_agent(), so cleanup happens on both inline and delivery-deferred teardown paths.
• agent/transports/codex_app_server_session.py retains stable client/thread references during run_turn(), allowing inactivity-timeout teardown to close the transport concurrently without a NoneType race.
• A closed transport racing turn/interrupt is treated as non-fatal because the turn is already stopping.

Related Issue

Fixes #62101

Why

Cron constructs a fresh AIAgent per run. The Codex runtime lazily attaches a CodexAppServerSession, but AIAgent.close() does not currently own that session. In a one-shot CLI process, parent exit masks the problem; in the long-lived gateway, the client reader threads keep the subprocess and its pipes alive indefinitely.

Cleanup originally landed immediately after ThreadPoolExecutor.shutdown(wait=False). Review correctly identified that the worker can still be inside run_turn() at that point. This revision moves cleanup to the centralized cron teardown path and makes the active turn safe when close happens concurrently.

Validation

• pytest -o addopts='' tests/agent/test_codex_runtime_close.py tests/agent/transports/test_codex_app_server_session.py tests/cron/test_run_one_job.py -q — 79 passed
• pytest -o addopts='' tests/cron/test_cron_inactivity_timeout.py tests/cron/test_codex_execution_paths.py -q — 13 passed
• python -m py_compile for all changed production and test modules — passed
• Mutation check: temporarily restoring the unsafe self._client.is_alive() dereference made the new concurrency regression fail with the original AttributeError; restoring the fix returned the suite to green.

The focused regressions cover:

• session close and reference clearing
• no-op cleanup without a session
• cleanup despite a session close error
• close racing an active run_turn()
• close racing turn/interrupt
• centralized scheduler teardown, including when agent.close() fails

Manual verification

With provider: openai-codex and openai_runtime: codex_app_server, a triggered cron run returns the gateway's Codex child-process count to baseline after delivery/teardown instead of leaking one app-server per run.

Checklist

• Bug fix is scoped to Codex app-server cron lifecycle cleanup
• Cleanup uses the existing centralized cron teardown path
• Concurrent-close behavior has regression coverage
• Focused agent/session/cron tests pass
• Broader cron inactivity and Codex execution-path tests pass
• Tested on macOS arm64 with Python 3.11

[alt-glitch]

This was generated by AI during triage.

Duplicate of #62105 — same fix for #62101 (cron per-run codex app-server subprocess leak): both add an idempotent close_codex_session(agent) helper in agent/codex_runtime.py and call it from cron run_job()'s finally block. #62105 was created first (14:01:10Z vs 14:11:28Z), so it is canonical. Related: #62101 (target issue), #55450 (broader run_agent.py session close, same family).

[teknium1]

Thanks for tracing the missing cron-owned Codex session cleanup; current main does leave _codex_session outside AIAgent.close() (run_agent.py:3463-3544) while cron creates a fresh agent per run (cron/scheduler.py:3044-3075).

Problems

• The proposed cron/scheduler.py:3144 close can run while the worker is still active: the pool is stopped with wait=False (cron/scheduler.py:3133-3134) and agent.interrupt() follows later (cron/scheduler.py:3157-3159). CodexAppServerSession.close() clears self._client (agent/transports/codex_app_server_session.py:300-310), but the active run_turn() loop reads self._client.is_alive() (agent/transports/codex_app_server_session.py:455-465). This is a close-vs-turn race.

Suggested changes

• Make run_turn() safe against concurrent close, then route cleanup through _teardown_cron_agent() (cron/scheduler.py:3324-3346), which already covers both inline and deferred cron teardown.
• Add a scheduler-path regression; the added tests currently validate only the standalone helper.

Automated hermes-sweeper review.