Launcher writes session settings to a predictable, world-readable path with a TOCTOU race allowing ANTHROPIC_BASE_URL hijack #25

Summary

The Auto Router caches model routing decisions in a module-level dictionary (_cache: dict[str, str] = {} in router.py). This cache is shared across all requests and all instances of ShimServer within the same Python process. A previous routing decision for one user message can influence routing for a semantically identical message in a different session, and the cache is never persisted or invalidated on model configuration changes — including when the user switches models via the picker.

Evidence

codex_shim/router.py:

_CACHE_MAX = 256
_cache: dict[str, str] = {}

Cache key construction:

def _cache_key(signal: dict[str, Any]) -> str:
    return "%s|%s" % (signal["has_images"], hash(signal["task"]))

The key is (has_images, hash(user_message_text)). Python's built-in hash() is not cryptographic and is not stable across restarts (PYTHONHASHSEED). More critically, the cache is never cleared when:

• A model is added or removed from models.json.
• A model's credentials are rotated.
• The user switches the active model via /api/switch.
• A configured model becomes unavailable (returns 401 or 429).

The eviction policy is a full cache clear when len(_cache) >= _CACHE_MAX — a poor replacement for LRU and means the cache flip-flops between 256 entries and 0.

Why this matters

1. Stale routing after config change: If a user reconfigures their models (removes a model, rotates an API key to a new model), the cache continues routing new requests to the old model slug for up to 256 cache entries.
2. Cross-request contamination: If two users share the same shim process (team/shared-server deployment), one user's cached routing decision affects another user's task routing.
3. Non-deterministic key: Python's hash() varies per process invocation with PYTHONHASHSEED randomisation. The cache is only valid for a single process lifetime, so any code that assumes cache keys are stable across restarts will be surprised.
4. Thundering herd on full eviction: The bulk-clear eviction means 256 concurrent requests could all miss the cache simultaneously and all fire the classifier, causing a sudden spike in API calls.

Root cause

The module-level mutable global was chosen for simplicity. The design does not account for cache invalidation on configuration changes, which are a normal operation (model switching via picker).

Recommended fix

• Clear the router cache whenever _set_active_model or switch_model is called.
• Use per-instance cache storage on ShimServer rather than a module-level dict, so multiple server instances don't share state.
• Replace bulk-clear eviction with LRU eviction (e.g., functools.lru_cache or collections.OrderedDict).
• Document that the cache key uses Python's hash() and is not stable across restarts.

Acceptance criteria

• Calling switch_model clears the router cache.
• Router cache is stored per ShimServer instance, not as a module global.
• LRU or FIFO eviction replaces the current bulk-clear strategy.
• Tests verify cache invalidation on model switch.

Suggested labels

bug, architecture, reliability

Priority

P2

Severity

Medium — incorrect routing after model reconfiguration; no security impact in single-user deployments.

Confidence

Confirmed — module-level _cache dict and bulk-clear eviction are explicit in the source; switch_model does not call reset_cache.