Clawdlinux Operator Cost And Monitoring

The operator exposes controller-runtime metrics and optional routing and cost metrics. Enabled components determine which series exist. Inspect exported /metrics before writing alerts.

Metrics

Documented routing series:

agentic_model_routing_total{provider, model, task_category}
agentic_tokens_used_total{provider, model, direction}
agentic_estimated_cost_usd{provider}

Token direction is input or output. Label order can differ. The default NoOpCostReporter records nothing. For local evaluation, AGENTIC_COST_TRACKING=memory enables these additional series:

agentic_workload_cost_usd{workload, namespace}
agentic_workload_tokens_total{workload, namespace, type}
clawdlinux_agent_cost_dollars{workload, namespace, model}

Memory reporting resets on restart. It is unsuitable for billing or durable chargeback. Cost estimates and reporter hooks do not establish complete token-budget enforcement.

Traces And Dashboards

The optional observability chart includes OTel Collector, Tempo, Prometheus, Grafana, ClickHouse, and Qdrant. Trace and log storage depends on enabled components. Routing instrumentation describes model.routing, task.classification, provider.resolution, and model.call spans. Check SDK initialization, exporter reachability, and sampling when traces are missing.

Inspect provider and model selection, token attributes, and error events when diagnosing routing failures. Verify alert expressions against exported metrics before using the cost alert sample.

Logs And Health

Find actual release-specific names before selecting logs or Grafana services:

kubectl -n agentic-system get deployments --show-labels
kubectl -n agentic-system get services
kubectl -n agentic-system logs deployment/<operator-deployment>

Documented health endpoints are /readyz and /healthz on port 8082; metrics use /metrics on port 8080.

Evidence Boundary

Trace storage and audit table creation do not prove same-run audit capture. Automatic complete per-run signed evidence is not connected. See evidence boundaries and troubleshooting.