Braintrust Tracing

This hooks into Claude Code's event system to trace every session, turn, and tool call to Braintrust for debugging and analysis. It captures the full hierarchy of LLM calls, tool executions, and token usage, with special handling for sub-agents (though those currently create orphaned traces you'll need to correlate manually via session IDs). The hook architecture is comprehensive, there are Python scripts to query traces, and the documentation is refreshingly honest about what works and what doesn't. If you're trying to understand why Claude Code made certain decisions or optimize token usage across complex sessions with nested agents, this gives you the visibility you need. The state management is file-based, so you can inspect everything without touching the API.