fix(gateway): cron cleanup on failure + session finalize guard

[jjjojoj]

Summary

Two small bug fixes for the gateway runner:

#12175 — Cron ticker leaks on failure shutdown

start_gateway() returned early on should_exit_with_failure before calling cron_stop.set() / cron_thread.join(), leaking the daemon cron thread. Moved cleanup into a finally block so it always runs.

#12176 — /new emits on_session_finalize(session_id=None) with no prior session

_handle_reset_command() always fired on_session_finalize, even when old_entry was None (first session). Added a guard so the finalize hook only fires when there is an actual old session to finalize.

Test plan

• Verify cron thread is joined on failure exit path
• Verify /new on first session does not emit on_session_finalize
• Verify /new on existing session still emits on_session_finalize with correct session_id

[DavidMetcalfe]

Thanks for the contribution @jjjojoj. This PR is effectively superseded and can be closed. Here's why:

All three unrelated fixes have already landed on main:

• Fix 1 (Skills cache invalidation, Bug: disabled skills still appear in system prompt due to stale LRU cache #12294): Merged via PR fix(agent): invalidate skills prompt cache when disabled skills change #12645 (commit fd119a1c). The current code already includes disabled = get_disabled_skill_names(...) before the cache-key construction, using tuple(sorted(disabled)) in the key.

• Fix 2 (Session-finalize guard): On main (gateway/slash_commands.py:194) the code now computes _old_sid = old_entry.session_id if old_entry else None and passes it to the hook. The hook receives session_id=None when there is no old entry and handles it correctly.

• Fix 3 (Voice file routing, Background task media delivery passes extract_media tuples into send_document() #12064): Merged via commit 498fc678. The current code (gateway/run.py:12118-12135) already routes media through should_send_media_as_audio → send_voice, with send_video and send_image_file handling the other cases.

The cron cleanup fix (Fix 4) is stale: The diff does not incorporate the additional cleanup steps now present on main (cron_provider.stop(), housekeeping_thread.join(), _planned_stop_watcher). The try: pass finally: approach also adds complexity beyond what's needed.

A clean replacement PR exists: PR #55031 applies the same core idea (running cleanup before returning on failure) as a minimal reorder of existing code — no new logic, no new patterns.

The diagnosis work here was useful for identifying the issue.

[teknium1]

Thanks for identifying the failure-path lifecycle gap. The cron diagnosis is still valid on current main: gateway/run.py:20749-20752 returns before the cleanup beginning at gateway/run.py:20763.

Problems

• The PR's synchronous cron_thread.join() is no longer compatible with the current lifecycle contract. Current main documents that joining there blocks the event loop and can drop an in-flight cron delivery; it instead uses cooperative _await_thread_exit() (gateway/run.py:20756-20775).
• The bundled prompt-cache and background-media fixes are already present on main (agent/prompt_builder.py:1479-1489; gateway/run.py:13367-13401). The proposed finalize-hook guard conflicts with the documented nullable callback contract (gateway/slash_commands.py:222-234; website/docs/user-guide/features/hooks.md:837-845).
• No regression test covers the failure return path.

Suggested changes

• Salvage only the cron fix by moving the failure return after the existing cooperative cleanup sequence, then add a failure-path regression test.

Automated hermes-sweeper review.