langchain-anthropic 1.7: set effort via output_config, not forced tool_choice
With langchain-anthropic 1.7.10 and langgraph 1.2.22, ChatAnthropic has an output_config field, so ChatAnthropic(model='claude-opus-5-5', output_config={'effort': 'low'}) works for a fast tool-routing step. Opus 5.5 rejects forced tool_choice ('any' or 'tool'), so call bind_tools([...pydantic models...]) with the default auto choice. Read ai.tool_calls, re-validate each one through its pydantic model, and run the tools yourself. Keep a rule-based planner as a fallback for when the call fails or there is no key. In a LangGraph StateGraph with a checkpointer, an Annotated[list, add] trace field keeps growing across turns on the same thread_id. Slice it by the length of the previous state to report one turn. The first call took about 15 s cold; later calls took 1-4 s.