Module proxy_usage 

On-disk record of the proxy’s measured provider spend (proxy_usage.json).

The proxy’s usage meter (proxy::usage_meter) aggregates the real provider usage of every finalized turn into these per-model sums and persists them; the dashboard, CLI, stats and the savings ledger (which run in other processes) read them back through load_persisted to learn the user’s real provider bill.

Structs

CohortUsage

Cumulative holdout cohort totals. Keyed by arm name: "control" | "treatment" for output savings (#895 Track B), "compression:control" | "compression:treatment" for input compression (#1905). The average tokens per turn are *_tokens / requests. Only populated while a holdout is active.

LastModel

The model the proxy saw last, with the time it saw it.

MeasuredSlice

Token/cost sums of the turns that reported a measured provider charge.

ModelUsage

Cumulative real token counts for one model. Cost is derived at read time so a pricing-table change re-values historical usage consistently.

PersistedUsage

On-disk shape of the measured spend totals.

ResponseObservation

Payload-free local measurement. Unknown cost stays unknown; neither HTTP status nor receipt of usage supplies the private runtime’s quality label.

Enums

UsageBoundary

Where provider usage was observed, not an assertion of answer quality or successful generation. Terminal usage also accompanies failed responses.

Functions

load_persisted

Cross-process read of the persisted measured spend (dashboard / CLI / ledger).

persisted_dominant_model

The model carrying the most measured tokens (excludes the “unknown” bucket). Used to value savings against the real dominant model when no explicit model is configured.

usage_path

Where the measured spend totals live in the data directory.