hermes - 💡(How to fix) Fix concurrency: providers/_discover_providers() sets _discovered=True before registry is populated (TOCTOU) [1 pull requests]

Bug

providers/__init__.py::_discover_providers() sets _discovered = True at line 154 before any plugin imports run — before _REGISTRY and _ALIASES are populated.

def _discover_providers() -> None:
    global _discovered
    if _discovered:
        return
    _discovered = True          # ← flag set HERE, registry still empty
    # 1. Bundled plugins — directory scan + imports
    # 2. User plugins — directory scan + imports
    # 3. Legacy per-file modules

Race window: Thread A sets _discovered = True, Thread B enters get_provider_profile("nvidia"), sees the flag, skips discovery, reads an empty _REGISTRY, and returns None. Unlike agent/transports.get_transport(), get_provider_profile() has no retry on miss — it returns None directly.

def get_provider_profile(name: str) -> ProviderProfile | None:
    if not _discovered:
        _discover_providers()
    canonical = _ALIASES.get(name, name)
    return _REGISTRY.get(canonical)   # ← returns None for valid provider

A None profile causes the call site to fall back to generic provider configuration (no model-specific routing, temperature rules, or context limits).

Fix: Add _discover_lock = threading.Lock() and apply double-checked locking — fast lock-free path when already discovered, slow path acquires lock, re-checks, runs all imports, then sets _discovered = True last.

Affected files

• providers/__init__.py — _discover_providers(), get_provider_profile(), list_providers()