Usage & Analytics
Every request is written to a usage ledger, and the dashboard reads that ledger several ways. This page covers what is recorded and how to read it.
What is recorded per request
| Field | Meaning |
|---|---|
| Provider | Which provider served it — a registry id, or a custom node's id |
| Model | The model that answered |
| Prompt / completion tokens | Both directions |
| Cost | From the pricing table, in your display currency |
| Compression mode | rtk, caveman, adhd, ponytail, stacked, or none |
| Tokens saved | What RTK actually returned, not an estimate |
| Latency | Time to first token and total |
| Outcome | Success, or the failure reason and HTTP status |
| Fallback chain | Which candidates were tried before one succeeded |
Live stream
GET /api/usage/stream is a Server-Sent Events feed the dashboard's topology
animation and recent-request list read. GET /api/usage/stats and
GET /api/usage/chart serve the aggregates.
Dashboard views
The Usage page switches between several views of the same data
(web/src/components/analytics/UsageDashboard.svelte):
| View | Component | Shows |
|---|---|---|
| Summary KPIs | SummaryKpiCards.svelte |
Totals for the selected period |
| Chart | UsageChart.svelte |
Volume over time |
| Breakdown | UsageBreakdownTable.svelte |
Grouped by provider, model, or key |
| Provider topology | ProviderTopologyCard.svelte |
Which providers talk to which |
| Activity heatmap | ActivityHeatmap.svelte |
When traffic happens |
| Combo flow | ComboFlowCard.svelte |
How a combo's chain resolved |
| Node health | NodeHealthTable.svelte |
Per-node errors and cooldowns |
| Request details | RequestDetailsTab.svelte |
One request's full payload |
Custom ranges
The period selector is not limited to fixed presets — 14d, 12h, 7d all
work, because the range parser accepts a number plus a unit
(internal/analyticsrange/range.go).
Compression analytics
GET /api/analytics/compression reports what the token savers actually did:
- Total requests and total tokens saved
- Average savings ratio
- Average duration
- Mode breakdown, including
stacked(RTK combined with a style saver)
The mode is resolved on the live pipeline, so a request that ran RTK and Caveman
is labelled stacked rather than being attributed to one of them.
Cost display
Costs are computed from the pricing table (internal/pricing/) and shown in your
chosen currency — USD, IDR, EUR and ten more. The exchange-rate table is editable
and can be refreshed from a public provider with
POST /api/settings/currency/refresh.
Quota tracking
The Quota Tracker page reads per-provider quota and the reset schedules:
- Codex reset credits —
GET /api/usage/{connectionId}/reset-credits, andPOST …/reset-credits/consumeto spend one. - GLM reset cards (Z.ai international) —
GET|POST /api/usage/{connectionId}/glm-reset-cards. - Auto-ping schedule — set a background interval that keeps quota current without traffic.
Cache analytics
GET /api/cache and GET /api/cache/entries read the semantic response cache:
hit rate, entry count, and the entries themselves. DELETE /api/cache clears it.
The dashboard reports the cache the engine actually serves from. That matters: a second cache instance would report zero entries while the engine served hits, and clearing from the dashboard would leave the engine still serving them.
Retention
Usage history is pruned on a schedule. Pruning keeps a carryover per dimension, so spend already counted is not billed twice after the rows are gone — the invariant is covered by tests.
Tracing
GET /debug/traces gives p50/p95 latency per provider+model, which is the
fastest way to find the provider that is actually slow.