Features:
-
anthropic: support adaptive thinking for hosted Claude by dl-alexandre
-
add thinking and reasoning options to deepseek (#754) by Bill Huang
-
google-vertex: grammar-constrained structured output for Claude via output_config.format by Arthur Valingot
-
anthropic: honor explicit cache_control from ContentPart metadata by Arthur Valingot
Bug Fixes:
-
google: pass thinking config to Gemini image generation (#757) by Shawn Vo
-
ci: clear Elixir 1.20 compile warnings (#758) by mikehostetler
-
openai_codex: keep HTTP responses stateless (#756) by BlueHotDog
-
bedrock: preserve Cohere inference profile IDs by dl-alexandre
-
remove deepseek's max reasoning level, and use xhigh instead by Bill Huang
-
anthropic: map refusal/pause_turn/model_context_window_exceeded stop_reasons instead of collapsing to :error (#749) by Arthur Valingot
-
Fix reasoning_effort not being passed through to OpenAI Chat Completions API (#753) by green-david
-
google_vertex: Allow for specifying google_thinking_level for Google Vertex (#752) by green-david
-
mistral: parse nested thinking chunks (#744) by mikehostetler
-
openai: preserve Responses file_id inputs (#740) by mikehostetler
-
anthropic: return adaptive thinking text by green-david
-
openai: send previous_response_id with store: false on Responses API by bowd