Token usage monitoring
Flow
After a model turn, the runtime emits agent.turn.usage into the caller's durable event pipeline. AgentHost does not itself own the database. The payload includes input tokens, output tokens, total tokens, nano-AIU, and the response model when available.
The runtime also creates an Agentweaver activity for the model turn and records the agentweaver.token.usage metric. It adds run, project, agent, model, token, nano-AIU, duration, and first-token data when the provider supplies it.
The metric named agentweaver.token.usage measures nano-AIU cost, not token count. Its counter is incremented only for positive cost; token counts are separate event fields, span attributes and metric dimensions.
In production, run events use the EF/Postgres durable event stream. Local development can use the SQLite event stream. Subscribers replay events by cursor, so a web replica can show work performed by another replica.
Queries
Application Insights provides project metrics and run traces when telemetry is available. Stored data supplies fallback model and agent usage where supported; it does not reconstruct all missing telemetry charts or traces.
| Endpoint | Purpose |
|---|---|
GET /api/projects/{id}/metrics | Project throughput, model, agent, duration, first-token, and AI-credit metrics. |
GET /api/runs/{id}/token-breakdown | Per-agent token and AI-credit data for a run. |
GET /api/metrics/runs/{runId}/traces | Agent and LLM spans for a run. |
The retired /usage routes, token-usage projection store, and token_usage_records table are not part of the current architecture.
AI-credit unit
1 AIC (AI Credit) = 1,000,000,000 nano-AIU
display value = total_nano_aiu / 1_000_000_000Source
packages/Agentweaver.AgentRuntime/CopilotAIAgent.csapps/Agentweaver.Api/Endpoints/MetricsEndpoints.csapps/Agentweaver.Api/Metrics/MetricsDtos.csapps/Agentweaver.Api/Metrics/AppInsightsMetricsService.cs