v1.3.0

Costs come from exactly two places now: the vendor-reported amount, or the unified price list. New vendor: Groq. We recommend leaving unified pricing on — vendor APIs keep changing or dropping fields, and one day a field we rely on may be gone.

Added

• groq-official crawl vendor (opt-in under Settings → Vendors). Syncs the same /platform/v1/organizations/{org}/activity endpoint the Groq console Usage page uses (Stytch JWT + org header, orgs auto-discovered), one record per day × model × API key with vendor-reported USD cost. Keeps unifiedPassthrough: per-second audio billing (e.g. whisper num_seconds_billed) has no token-rate expression, so the vendor spend stays authoritative in both modes.
• CI now reads the crawl artifacts. A new test suite parses every committed price snapshot and model snapshot, recomputes each vendor's rates.preset.json/rates.source.json from its snapshots, and rejects any shrink that is not named in tools/snapshot-shrink-allowlist.json — these files previously passed CI without anything reading them.

Changed

• No more rate-table estimation. A record bills its own vendorCost when the vendor reports one, and bills 0 when it does not. The author-maintained fallback tables no longer price anything at render time; they survive only as an input to the unified preset and as peak/off-peak window definitions.
• Consequences: with unified pricing off, deepseek-official and mimo (tokens only, no reported amount) and opencode records the Console sends cost: null for bill 0; their savings column is 0. With unified pricing on (the default) token vendors are unchanged — deepseek/mimo are priced from their official group.
• Recommendation: keep unified pricing on. Vendor-reported fields shift without notice (opencode already removed usage/rows and now sends cost: null for some rows); the unified list prices from the makers' published rates with daily watches, so history stays comparable even when a vendor stops sending a field.
• Settings → Vendors cost-basis line for a token-only vendor reads cost 0 (no vendor cost) instead of cost estimated.
• A partial scrape is refused, not committed. Both watchers now compare the freshly crawled model count against the committed one and abort the write when it drops, naming the ids that would disappear (--allow-shrink, or the workflow's allow_shrink input, accepts a real retirement). Previously only wholly empty payloads were rejected, so a truncated response could silently gut a snapshot.
• The shrink policy is per source. A count drop on a stable catalog (commandcode) still blocks the write, but opencode's gateway catalog is regenerated per request and legitimately rotates, so a drop there is recorded and warned about instead of failing the daily run.
• A failing source no longer blocks the healthy one. model-watch and price-watch still fail red, but the sources that parsed are committed in the same run instead of being discarded with the failure.
• The openai parser fails closed on a duplicate model id instead of keeping whichever row it read last.
• rates.source.json provenance is now enforced. For parser-backed official vendors it is a generated mirror of the price snapshots; the daily job refuses to overwrite a file it would not itself have produced, so hand-editing one there no longer disappears silently.