Operability is the Admin surface for understanding how the governed runtime is behaving now. It complements Audit, which focuses on individual security/execution events, and OpenTelemetry/Prometheus, which export telemetry to external systems.
Execution metrics
The metrics endpoint can return:
- query and DML counts;
- success and failure counts plus success rate;
- p50 and p95 latency;
- latency sample size and whether latency is sampled;
- slow-query count;
- IP rate-limit count;
- MCP-key rate-limit count.
Filters can narrow the view by time range, database, access key, and tool name.
Database health
Database health items expose operational state for each managed connection:
| Field | Meaning |
|---|---|
Status | current health state |
LastCheckedAt | latest probe time |
LastSuccessAt | latest successful probe |
LatencyMs | probe latency when available |
ConsecutiveFailures | current failure streak |
OutageStartedAt | start of the current outage window when known |
LastError | most recent provider/probe error |
Health probing is controlled by the HEALTH_PROBE_* settings. A disabled background probe means this surface should not be interpreted as a continuously refreshed external monitoring system.
MCP-key usage
Key-usage results report the key identity/name, last use time, request count, successes/failures, rate-limit count, and rate-limit rejection rate.
Use this data to identify stale credentials, unexpectedly noisy agents, and keys that routinely hit their configured limit. Then inspect MCP Keys before changing scope or limits.
Outbound deliveries
The runtime records outbound delivery attempts such as configured alert/SIEM webhooks. Delivery items expose:
- category and status;
- attempt count;
- creation/delivery/last-attempt time;
- latest error.
The list endpoint accepts a limit parameter with a default of 100.
Retry a delivery
| Operation | Permission |
|---|---|
| metrics / DB health / key usage / delivery list | /runtime/operability → view |
| retry delivery | /runtime/operability → edit |
Retry is an explicit operator mutation. It is useful after correcting a transient destination/network problem, but repeated retries are not a substitute for fixing an invalid endpoint, secret, or downstream service.
Where to investigate next
- Audit — exact actors/actions/results and SQL execution context.
- Security Policies — runtime limits that may explain rejections.
- Observability — Prometheus, OTLP, traces, and external telemetry integration.
- Distributed Deployment — shared coordination when several nodes must agree on limits/state.