The OpenAI-compatible gateway

Serving openai_compatible/<gateway model id> from the deployment's environment: the two variables, the verbatim wire id, the guarded dial and its escape hatch, the claw-only routing and its C186 warning, the catalog (ADR-122) and what an unknown answer costs, the env-funded accounting.

Read backends.md — the section is the reference; this page is the pointer.

Provider/model configuration lives in the DB

The database is the seat of provider/model configuration — context window, prices, capabilities, the per-credential default model (spec and tracking: #2906). Every provider/model record resolves through one cascade, global → org → team → bot → bot-in-team → bot binding, which extends the credential audience chain (OrderedAPIKeyCandidates, pkg/secrets/byok.go) and the model-override fold (model.ModelOverrides). The env tables (OPENAI_COMPATIBLE_MODELS, OPENAI_COMPATIBLE_CATALOG_PROVIDER) are runner-side fallbacks, not the seat of truth.