behavior-analyzer

To analyze agent session behavior and compute a health score, run the analyzer against workspace logs like .self-eval-reflections.jsonl and .cache-monitor.json. It processes data from self-eval, cache-monitor, evolve, and yolo_classifier sources to generate per-session scores and actionable recommendations.

OpenClaw session anomaly detection requires access to workspace data files including .self-eval-reflections.jsonl, .evolve-rule-applications.jsonl, .cache-monitor.json, and .permission-decisions.jsonl. These sources provide the self-eval, cache, and classifier data needed to detect anomalies.

You can integrate session behavior analysis into an automated workflow using the exposed Python API and CLI. This allows you to programmatically compute session health scores and apply trend analysis within your existing automation pipelines.

Trend analysis improves governance and risk management by tracking session health scores over time. It identifies persistent anomalies and provides automated recommendations for policy updates, enabling safer automation and historical persistence for auditing.

The best way to monitor OpenClaw agent sessions for real-time anomalies is to analyze multiple data sources simultaneously. Integrating self-eval, cache-monitor, and evolve data computes a comprehensive health score and detects behavioral deviations.

A limitation when analyzing yolo_classifier data is that the session health score depends entirely on the availability and quality of workspace logs. Without properly populated .permission-decisions.jsonl files, anomaly detection and trend analysis will be incomplete.