0.2.23

langgraph 0.2.23

Summary of Changes

• Added support for token-by-token message streaming in LangGraph applications using stream_mode="messages" #1716
• Added support for custom output streaming from nodes using stream_mode="custom" #1716
• Enhanced chat_agent_executor to handle tools with return_direct flag, allowing tools to return directly without going back to the agent #1780
• Improved concurrent streaming with queue-based approach for better performance #1716
• Added msgpack as a dependency for efficient serialization #1716
• Fixed docstring example for reducer function in StateGraph #1733

Detailed Changes

langgraph.pregel.types.StreamMode

• Added new stream modes:
  • "messages": Stream LLM messages token-by-token
  • "custom": Stream custom output from nodes via write parameter

langgraph.pregel.messages.StreamMessagesHandler

• Added new handler to capture and stream LLM messages token-by-token
• Supports both chat model outputs and chain outputs containing messages
• Ties into the overall streaming system to provide real-time updates

langgraph.utils.queue.AsyncQueue and SyncQueue

• Added new queue implementations to support concurrent streaming
• AsyncQueue provides async queue with wait functionality
• SyncQueue provides thread-safe queue with wait functionality

langgraph.prebuilt.chat_agent_executor

• Modified should_continue function to return "tools" instead of "continue" and "__end__" instead of "end"
• Added route_tool_responses function to handle tools configured with return_direct
• Changed the edge connection pattern to support direct returns from tools
• Improved support for tools that should return directly without going back to the agent

langgraph.graph.state.StateGraph

• Fixed docstring example for reducer function return type (changed from int to list)