Hermes "provider failed after retries; check logs"

Likely cause

Two causes, and neither is fixed by retrying — which is what makes the message misleading. First, a retired or renamed model: when your configured model id no longer exists on the provider, Hermes still reports it as a retry-exhausted failure and tells you to check logs, hiding the recovery step of picking a valid model (issue #55323). Second, capacity: when a provider returns HTTP 503 "model temporarily unavailable due to upstream capacity limits", Hermes retries the same overloaded provider three times and gives up, instead of treating a 5xx as a trigger for the fallback chain (issue #68771).

The fix

1. 1 Look up your configured model id in the provider's current model list and correct it if it has been retired or renamed. This is the cause retrying can never solve.
2. 2 If the model id is valid, read the logged status. A 503 means upstream capacity, not a fault in your setup.
3. 3 Configure fallback_providers so a 503 rolls over to a second provider rather than exhausting three retries against the overloaded one.
4. 4 For known-busy models, put a lower-demand model earlier in the fallback chain during peak hours.
5. 5 Update Hermes — the 5xx-triggers-fallback behaviour is a tracked change.