Replace Archytas with LangGraph

[brandomr]

Summary

Complete migration from legacy Archytas agent framework to LangGraph, eliminating all dependencies and compatibility layers while maintaining full backward compatibility. See ARCHYTAS_TO_LANGGRAPH_MIGRATION.md for a full migration guide.

Changes

• Core Agent System: Replaced ReActAgent with custom langgraph that intercepts tool calls to perform thought extraction so we can show the thoughts prior to the tool call.
• Tool System: Native LangChain @tool decorator with proper UI logging
  integration
• Chat History: Custom BeakerChatHistory which includes a direct port of Archytas's auto-summarization
• Dependencies: Removed Archytas, added LangGraph and LangChain ecosystem
  packages
• Model Configuration: Direct LangChain model instantiation (OpenAI,
  Anthropic, etc.

Issues

• 🚨 the chat history isn't being sent back to the agent--it is memoryless right now
• Model context window lengths are hard coded in chat_history.
• 🚨 tool calling with anthropic has been finicky during testing...empty AI responses
• 🚨 with claude 4, the agent often produces run_code calls with no code. I tried a variety of approaches to handle this but eventually the langgraph errors out with recursion issues. I need to spend more time digging into why it calls run_code without code.

Test script

1. hey can you ask me a question with ask_user tool?
2. think about something that will take several steps then do those steps in notebook cells and think about things as you do it
3. come up with 5 more steps we can build on from this and execute them in the same fashion. [this is usually where claude 4 starts to fail]