Change8

b7644

📦 llama-cppView on GitHub →
2 features🐛 2 fixes🔧 3 symbols

Summary

This release introduces support for reasoning/thinking content blocks in the Anthropic Messages API and includes significant refactoring of the Anthropic streaming state management for improved safety.

Migration Steps

  1. If using Anthropic API streaming with reasoning models, be aware of the new signature field and corrected event sequencing in the response stream.

✨ New Features

  • Added support for returning reasoning/thinking content blocks in Anthropic Messages API responses when using models with --reasoning-format deepseek and the thinking parameter enabled.
  • Thinking content blocks are now handled in non-streaming responses (appearing before text in the content array) and streaming responses (emitting thinking_delta events).

🐛 Bug Fixes

  • Fixed an issue in Anthropic Messages API streaming responses for reasoning models by adding a signature field and correcting duplicate content_block_start events.
  • Refactored Anthropic streaming state management to replace raw pointer usage with direct field copies, improving data flow clarity and safety.

🔧 Affected Symbols

Anthropic Messages APIupdate()to_json_anthropic()