Harmonia’s monitoring surface is assembled from durable Firestore records and bounded activity telemetry. It does not treat process memory or model narration as operational truth.
How Harmonia uses it
/api/metrics aggregates model calls, token usage, estimated cost, reservations, job status, and receipt counts./dashboard/monitoring presents operational health and recent execution data.- Structured agent activity records expose role, phase, status, operation, duration, and safe summaries without prompt or response bodies.
- Notifications and failure records surface unresolved operator attention.
- Trace IDs allow deeper diagnosis in the configured tracing backend.
Metrics use integer microdollars for aggregation and preserve pricing-catalog versioning. Provider estimates remain distinct from observed billing data.
Security and authority
Monitoring endpoints are tenant-scoped. Metrics cannot approve, retry, replay, publish, or change a budget. Operator actions leave the monitoring view and enter dedicated deterministic handlers.
Failure behavior
Missing usage, stale state, unresolved effects, and failed verifications remain visible. The UI does not infer health from absence of logs.
Evidence status
Aggregation, activity contracts, redaction, and dashboard rendering have automated coverage. Production monitoring requires authenticated records from the deployed services.
Relevant files
src/app/api/metrics/route.tssrc/app/dashboard/monitoring/page.tsxsrc/lib/observability/repository.tssrc/lib/observability/schema.tsagent/harmonia_agent/activity_models.pydocs/observability.mdx
Last modified on August 28, 2026
