Anomaly Detection

Statistical baselines on per-session metrics to catch abuse without static rules.

Metrics tracked

Metric	Description
tool_calls_per_5min	Tool invocations in rolling window
unique_tools_used	Distinct tool names in session
avg_response_length	Mean assistant message length
shell_exec_count	Shell tool usage
http_request_count	HTTP tool usage
injection_attempt_count	Prompt guard hits
egress_blocked_count	Egress filter blocks

Agents send JSON metric blobs on event_type: session_metrics via POST /api/events.

Algorithm

1. Load baseline mean and std dev per (instance_id, metric) from Postgres (minimum 10 samples).
2. Compute Z-score: (current - mean) / stdDev.
3. Flag anomaly when Z >= 3.0; critical alert when Z >= 5.0.

Baselines cache for one hour in memory.

Dashboard

Anomaly table and timeline chart show recent detections with metric name, Z-score, and session id.

Scheduler

Scout scheduler prunes old events and refreshes aggregates. Ensure Redis and Postgres are healthy for timely detection.

Tuning

Investigate false positives by raising thresholds in code (Z_ANOMALY_THRESHOLD) or narrowing which tools run unattended. Pair with Tool Policy for hard caps.