Catalog ingestor coverage 7/7 → 8/8 — Groq closes the last drift gap

2026-05-08

What We Built

Added the Groq catalog ingestor — closing coverage from 7/7 to 8/8. The 7/7 metric was set against the seven providers BR explicitly listed in its routing intelligence, but Groq had been quietly serving traffic without an ingestor. Surfaced during R23 closure verification when /v1/intelligence/status returned providers including groq despite the prior round having no Groq-shaped ingestor.

Why It Matters

Closes the catalog drift surface across every routed provider. No hand-maintained liveness for any model BR routes to.

How It Works

Groq's /openai/v1/models is OpenAI-compatible (Bearer auth, data: [{id, owned_by, active, context_window, max_completion_tokens, ...}]). Three nuances worth coding around explicitly:

1. Not a meta-router. Unlike Perplexity, Groq's owned_by reports

the architectural origin (OpenAI / Meta / Canopy / Groq itself), not a passthrough indicator. ALL listed models go under provider=groq because that's where BR routes them.

1. active flag matters. Groq surfaces deprecated models with

active: false. The ingestor filters them out so liveness signals reflect what's actually routable.

1. Bonus capability fields. context_window → maxInputTokens,

max_completion_tokens → maxOutputTokens. Same partial-payload protection pattern as the Moonshot ingestor.

The Numbers

• Provider ingestor coverage: 7/7 → 8/8
• New paired regression tests: 10 (all-active emission, capability

propagation, inactive filtering, error paths, default endpoint)

• Source files: 1 ingestor + 1 test + 1 wiring edit

Lockstep Checklist

• [x] API Routes: unchanged — diagnostics auto-surface via /v1/intelligence/status
• [x] TS SDK: unchanged
• [x] Python SDK: unchanged
• [x] MCP Schemas: unchanged
• [x] Master Record: unchanged