CommitBrief renders a code review as cards, JSON schema v1, or a CI exit code — which means the LLM has to hand back structured findings, not prose. Every provider can do that. The catch is that no two of them do it the same way, and some don't really do it at all.
There's exactly one schema the whole system targets. Getting four native APIs to honor it takes four completely different mechanisms; getting three more is a matter of asking nicely and not trusting the answer. This is how that works, and what happens when a model ignores the contract anyway.
TL;DR
One schema, many dialects. Every provider targets the same Finding shape, expressed through whatever structured-output mechanism that vendor offers — tool_use, strict json_schema, responseSchema, or just format: "json".
Structured output is a spectrum, not a guarantee. It runs from "the API enforces the shape" down to "we asked in the prompt."
