Symptoms
- The same request sometimes succeeds later without a configuration change.
- Another model or route remains healthy.
Likely causes
- Provider outage, saturated capacity, disabled channels, or transient routing failure.
- A best-effort free route has no current upstream capacity.
How to diagnose
- Check the TokenReply status page and the exact model route.
- Preserve the HTTP status and a safe request identifier for support.
- Test one small request on an operational alternative.
Fixes
- Retry with capped exponential backoff when the operation is safe to repeat.
- Use a tested fallback model when application semantics permit.
- Do not rotate keys or rewrite payloads when evidence points to upstream health.
How to verify
The original or approved fallback route completes a small request without changing credentials.