hermes - 💡(How to fix) Fix concurrency: _openrouter_prewarm_done.is_set()+.set() non-atomic — can spawn duplicate prewarm threads

Bug

The prewarm guard in run_agent.py uses a non-atomic check-then-act on threading.Event:

if ... and not _openrouter_prewarm_done.is_set():
    _openrouter_prewarm_done.set()
    threading.Thread(target=fetch_model_metadata, ...).start()

is_set() + set() are two separate operations — not atomic. Two concurrent AIAgent.__init__ calls (every gateway request creates one) can both see is_set() == False, both call .set(), and both start a fetch_model_metadata thread. The comment explicitly says the guard exists to prevent per-message thread leaks.

While the Event limits the leak to a small burst at startup rather than per-message, it violates the stated invariant.

Fix

Replace the Event with a threading.Lock used as a one-shot gate via acquire(blocking=False): only the first caller gets the lock (never released), all subsequent callers see the lock held and skip thread spawn.