Token usage and cost visibility
Agentweaver exposes recorded token consumption and AI Credit (AIC) cost through run inspection, graph cost chips, dashboards, and Observability. Read the scope, time range, and data source before comparing totals. A selected-range telemetry chart is not the same measurement as a durable run aggregate. For API contracts see the reference; for event/projection internals see the deep dive.
Run and graph cost surfaces
Run usage is projected from agent.turn.usage events. CostChip delegates its label to costChipLabel: positive totalNanoAiu displays AIC, positive tokens without cost display a compact token count, and absent/zero values produce no chip. The richer AiCredits control shares this formatting for run/graph contexts. The absence of a chip is not evidence that execution was free.
Coordinator inspection exposes AI-credit context and per-agent usage where available. Graph usage is associated with the relevant coordinator or child context, not inferred from a neighboring node. Open an orchestration and inspect its usage controls alongside the selected task's messages and artifacts.
The current board RunCard focuses on status, stage, agent, approvals, retry, and archive; it does not render CostChip or fetch supplemental run usage. Use run inspection and telemetry surfaces rather than expecting a cost beside every board status badge.
Sources: apps/web/src/components/CostChip.tsx:13, apps/web/src/components/costChipFormat.ts:16, apps/web/src/components/AiCredits.tsx:109, apps/web/src/components/board/RunCard.tsx:146.
Scope and surface mapping
| Surface | Scope and source | How to read it |
|---|---|---|
| Run usage API / run inspection | Persisted usage aggregates for that run; project usage APIs have their own endpoint-defined scope. | Treat these as run aggregates, not automatically as telemetry for the dashboard's selected range. |
| Project Dashboard | Dashboard state plus GET /api/projects/{id}/metrics?from=...&to=.... The Activity selector offers 7/30/90-day ranges. | The current Model metrics and agent leaderboard Cost use the selected-range metrics response, not a client-side sum of /runs/{id}/usage calls. |
| Fleet Overview | AI usage & performance aggregates project observability metrics for the recent projects shown on that page. | Its 7/30/90-day selector scopes that rollup; do not call it an exhaustive all-project billing total. |
| Project Observability → Overview | Selected-range project metrics. | Inspect model mix, credit trends, response time, and first-token timing with the range visible. |
| Project Observability → Agents | Metrics grouped by agent over the selected range. | Agent token breakdown is a dimensioned telemetry view, not proof that every run reported usage. |
| Project Observability → Traces | Recorded spans for a selected run, with persisted run events for tool detail. | Cost sums the model spans loaded into the trace tree; incomplete telemetry or pagination can differ from a durable full-run aggregate. |
Sources: apps/web/src/pages/DashboardPage.tsx:675, :881, :989, :1018; apps/web/src/pages/OverviewPage.tsx:299, :382; apps/web/src/pages/observability/ObservabilityOverviewPage.tsx:95; apps/web/src/pages/observability/ObservabilityAgentsPage.tsx:95, :190; apps/web/src/components/runs/traceTree.ts:308.
Empty and partial telemetry
Charts use explicit empty states such as No data for AI credits in this range, No response-time data for this range, and No first-token data for this range. Missing or delayed telemetry is not a healthy zero-cost result. Check the selected project/range and query availability before interpreting an empty trend or comparing it with per-model rows.
The previous AI-credit popover, dashboard, and Overview usage images were 1×1 placeholders; the Agents image was also a placeholder. They are omitted. The earlier Observability Overview capture had zero total AIC alongside nonzero model rows and empty trends; it is not used as an illustration of coherent totals. No replacement captures or visual verification are claimed.
Source: apps/web/src/components/dashboard/ModelPerformancePanels.tsx:332, :354, :386, :390.
Understanding the numbers
| Term | What it counts |
|---|---|
| Input tokens | Recorded prompt tokens sent to the model. |
| Output tokens | Recorded completion tokens returned by the model. |
| Total tokens | Reported aggregate for the specified run or telemetry scope; compare like scopes and ranges. |
| AIC | Nano-AIU divided by 1,000,000,000. formatAic renders values below 1 AIC with four decimal places. This is a credit unit, not a currency amount. |
| Per-model / per-agent breakdown | Usage grouped by the dimension supplied by the endpoint. Missing attribution or incomplete telemetry must not be filled with invented rows. |
| Trace invocation cost | Sum of descendant model-call costs; agent and tool nodes do not add a second copy of their own model totals. |
Sources: apps/web/src/components/costChipFormat.ts:1; apps/web/src/components/runs/traceTree.ts:308–327; apps/Agentweaver.Api/Metrics/MetricsDtos.cs:159, :203.
Graph layout
Cost chips, pod indicators, and status metadata contribute to card height. Shared DAG layout uses rendered-height hints so graphs can accommodate that metadata; this is presentation behavior, not another source of usage totals. See apps/web/src/utils/dagLayout.ts and apps/web/src/pages/CoordinatorRunPage.tsx.