Provider Routing

Stable static priority

1. The requested model resolves to configured provider candidates.
2. Enabled candidates retain their configured static priority.
3. OPEN circuit candidates are excluded before selection.
4. The first available candidate is selected and recorded in the routing snapshot.

Non-streaming Chat fallback

For non-streaming Chat only, a retryable primary-provider failure can execute the next configured candidate within the fixed attempt budget. Provider attempts and executed fallback events are recorded. Streaming and image fallback are not enabled.

Inspect the Provider Attempts for a specific request through its Generation detail.

Health and circuit behavior

Stored circuit state filters OPEN providers. A HALF_OPEN candidate remains available and is marked as requiring a probe in the routing snapshot. Separate probe coordination exists, but this does not make every ordinary user request an automatic probe. If every candidate is unavailable, Qufas returns an explicit all_providers_unavailable error.

No smart ranking

Qufas does not currently rank providers by price, latency, AI scoring, or automatic “best provider” selection. Requests follow stable configured priority rather than random or intelligent switching.