trajectory-logging

What problem does it solve?

Log every agent action as structured JSON using the ATIF (Agent Trajectory and Interaction Format) schema. Use a ring buffer to prevent disk exhaustion and enable post-hoc debugging, eval dataset construction, and regression detection.

Core Features & Use Cases

• Structured trajectory entries: capture tool name, parameters, results, rationale, timestamp, agent ID, and session ID.
• Ring-buffer storage: capped at 50K entries per file to bound disk usage.
• Session metadata and lifecycle: record start/end times plus summary statistics for each session.
• Eval-framework export: provide an export tool to convert trajectories for eval-framework compatibility.
• Regression detection and search: compare runs for the same task and search across trajectories for patterns or errors.

Quick Start

Enable trajectory logging by enabling the PostToolUse hook, then run a sample session to generate a trajectory file at .harness/trajectories/{agent-id}-{session-id}.jsonl.