Metrics & traces

Prometheus

GET /metrics — unauthenticated so Prometheus can scrape it; restrict it at the network layer.

Metric	Type	Labels	Meaning
aegis_runs_total	counter	route, status	Runs finished, by final status.
aegis_run_duration_seconds	histogram	route	Wall-clock time per run.
aegis_exporter_failures_total	counter	exporter	Evidence records an exporter failed to deliver.

scrape_configs:
  - job_name: aegis
    scrape_interval: 15s
    static_configs:
      - targets: ["aegis:8000"]

Useful queries:

sum by (route) (rate(aegis_runs_total{status="blocked"}[5m]))          # blocks per second
histogram_quantile(0.95, sum by (le, route) (rate(aegis_run_duration_seconds_bucket[5m])))
sum(aegis_runs_total{status="paused"}) - sum(aegis_runs_total{status="denied"})  # rough review backlog

Point your own Prometheus at /metrics with the scrape config above; Aegis doesn't bundle a monitoring stack.

OpenTelemetry

Each run executes inside one span:

Span	Attributes
aegis.run	run.id, run.route, run.principal_id, run.status

Spans go to the process's global tracer provider (tracer name aegis.server). Aegis doesn't install an exporter; either run under auto-instrumentation —

pip install opentelemetry-distro opentelemetry-exporter-otlp
OTEL_SERVICE_NAME=aegis OTEL_EXPORTER_OTLP_ENDPOINT=http://collector:4317 \
  opentelemetry-instrument aegis serve --config aegis.yaml

— or pass a tracer= to create_app() when embedding the server.

Beyond metrics

The per-run event log (aegis explain, GET /v1/runs/{id}) and the evidence ledger carry the verdict-level detail that metrics aggregate away.