0.28

• New model: Claude Fable 5.1 (claude-fable-5.1). Like Fable 5 it always thinks, so -o thinking 0 raises an error. #89
• Reasoning traces are now requested by default for models that think by default. Claude 4.7 and later models leave the thinking trace out of the response unless asked, so the plugin now sends thinking.display: summarized whenever thinking is on - the summarized reasoning is streamed to standard error and stored in the logs. Pass -R/--hide-reasoning to send display: omitted instead, which leaves the trace out of the response entirely. #88
• Requests declined by Anthropic's safety classifiers now raise a llm_anthropic.ClaudeRefusal exception (a subclass of llm.ModelError) instead of silently returning an empty response. The exception message includes the refusal category and explanation from the API, also available as .category and .explanation on the exception. #90
• Thinking token counts reported by the API are now recorded in token_details, along with the service_tier if it is anything other than standard.