I see in the latest generated Anthropic PR (#4897) we parse "refusal" from the Anthropic API stop reason.

Currently this would be parsed into AiResponse.FinishReason as "unknown". Just wanted to mention in case it slips through the cracks.

As a side note, if refusals are ultimately surfaced as successful responses, the retry/model fallback behavior via AiPlan's while function seems unable provide input for what to do next when there is a refusal.