Monitoring

Health endpoint

curl -s http://localhost:8080/health | jq

Returns a JSON snapshot covering: model liveness, Redis, PostgreSQL, browser pool, behavioral queue, CPI. Use this for an external uptime check (UptimeRobot, Pingdom, custom cron).

/health page

The dashboard /health page is the live snapshot for the operator. Panels:

Panel	What it tells you
Model	Active provider, model name, last ping result
CPI	Cognitive Pressure Index 0–100 (composite of active goals, error rate, latency, memory growth, CPU). > 80 sets agent:cpi_high and pauses non-critical jobs
Memory	Container RSS, growth rate
Queues	events:incoming, events:outgoing, behavioral:pending (with cap warnings)
Goals	Active count, paused count, error count
Integrations	Circuit breaker state per connector (CLOSED / OPEN / HALF_OPEN)
Learning Queue	behavioral:pending depth with thresholds (yellow ≥ 20, red ≥ 40 of 50 cap)

Cognitive Pressure Index (CPI)

CPI is a 0–100 composite metric updated by the cpi_monitor job (5 min). Components and weights:

Component	Weight
Active goals	20%
Error rate (last 5 min audit log)	25%
Latency p95	20%
Memory growth	15%
CPU	20%

Thresholds:

• HIGH > 80 → sets agent:cpi_high flag (TTL 10 min). Background jobs (autonomous, dream, perception) check this flag and skip if set.
• CLEAR ≤ 60 → flag cleared.

Visible at /health and /cognitive.

Self-Integrity Monitor

scheduler/integrity.py runs every 6 h. It cross-checks:

• Declared self-model strengths vs actual skill success rates (skill_success_rates)
• Epistemic state drift (sudden confidence changes)
• AuditLog error spikes

Writes a structured agent:integrity_report JSON in Redis. Visible at /cognitive (Integrity tab). Drift larger than threshold triggers a Telegram alert.

/metrics page

The dashboard /metrics page aggregates the audit_log table:

• Latency histograms per skill
• Error rates per skill
• Token usage by provider
• Skill call frequencies

Use it to spot regressions ("yesterday browser was 8s p95, today it's 25s") and to catch token-spend anomalies.

/traces page

For per-response forensic analysis. See Logs.

Recommended weekly checks

1. Audit log review — /audit filtered to the last 7 days. Look for unexpected skill.shell or skill.self_improve actions.
2. Behavioral rules — /behavioral-rules. Disable rules that no longer reflect your preferences. Watch for behavioral.rule_conflict_detected warnings in logs.
3. Model spend — provider dashboards (Anthropic, OpenAI, etc.). Background jobs (dream, perception, autonomous goals, behavioral learner) consume tokens.
4. Disk usage — docker system df -v or check the Docker named volumes (docker volume ls). Screenshots and browser sessions can grow.
5. Self-improve proposals — /self-improve. Approve only what you've reviewed; reject the rest.
6. Goal failures — /goals filtered to status FAILED. Use Replan or close the goal.
7. Container health — docker compose ps; all services should be healthy.
8. VACUUM ANALYZE is run weekly automatically by db_maintenance. If you're seeing slow queries, manually run VACUUM ANALYZE and inspect pg_stat_user_tables.

Operational checklist

Activity	Frequency
Verify all containers healthy	Daily
Inspect /audit for unexpected actions	Weekly
Review /behavioral-rules; disable stale ones	Weekly
Check model spend on provider dashboards	Weekly
Review /self-improve proposals	Weekly
Run external uptime check against /health	Always
Backup databases and memory	Daily
Update OS packages	Monthly
Rotate secrets (DASHBOARD_PASSWORD, DASHBOARD_SECRET, MEDIA_SIGNING_SECRET)	Quarterly
Rotate model API keys	Per provider's policy
Renew TLS certs	Automatic via certbot, verify quarterly
Run regression suite	Automatic on every rebuild
Forensic audit (mandatory tests)	Weekly
Forensic audit (edge + adversarial)	Monthly
Forensic audit (state + cross-layer + stress)	Quarterly

Alerting

WASP does not ship with a built-in alerting system. Recommended:

• External uptime check on /health.
• Telegram-side alerts: the agent itself sends 🤖 notifications for autonomous actions, integrity drift, and threshold breaches.
• For external infrastructure alerts (disk full, host down), use a separate monitoring agent (Prometheus, Datadog, etc.) on the host.

See also

• Logs — log surfaces and event names
• Scaling — capacity, backups, recovery
• Audit Logs — what's logged
• Health Dashboard — page reference