trajectory-logging

Agent action logging captures every tool name, parameter, result, and rationale as structured JSON using the ATIF schema. This enables post-hoc debugging and eval dataset construction without missing intermediate execution steps.

Prevent disk exhaustion during trajectory logging by using a ring buffer capped at 50K entries per file. This bounds disk usage while retaining recent agent actions for regression detection and debugging.

Yes, you can export agent trajectories to build an eval dataset using the provided eval-framework export tool. It converts structured JSONL trajectory logs into a compatible format for evaluation framework integration.

Detect regressions across agent runs for the same task by comparing structured trajectory logs. The regression detection utility analyzes JSONL entries across sessions to identify behavioral changes or errors.

Search across agent trajectories to find errors or patterns using the trajectory search utility. It indexes structured JSON logs containing tool names, parameters, and results to locate specific execution behaviors.

No, trajectory logging requires no external dependencies to capture agent sessions. Enable the PostToolUse hook to generate JSONL files recording session metadata, lifecycle events, and summary statistics.