What's New
Token deduplication. Usage records are now deduplicated by API request ID, eliminating double-counted entries that inflated total token numbers. Observed ~17% reduction in total token count on real-world datasets.
This directly addresses the remaining total-token overcount reported in #18. Combined with v4.2.1's model attribution fix, the Model Mix chart and token totals should now closely match other measurement tools.
Run python3 measure.py collect --rebuild after updating to recalculate historical data with dedup.
Full changelog: v4.2.1...v4.2.2