Variable: UNSUPPORTED\_CHAT\_FIELDS

const UNSUPPORTED_CHAT_FIELDS: readonly string[]

Top-level request fields an OpenAI client may send that Databricks Model Serving rejects outright, failing the whole turn rather than ignoring them.

Databricks validates the chat body strictly, so an unrecognized key comes back as parallel_tool_calls: Extra inputs are not permitted (the gateway’s pydantic validation) or json: unknown field "parallel_tool_calls" (its strict JSON decode), depending on the endpoint. Everything listed here is either OpenAI-platform bookkeeping with no bearing on the completion, or - in the case of parallel_tool_calls - a real setting Databricks has no way to accept, so dropping it is the only way the request can succeed at all.

Callers that translate a request field-by-field (an allowlist, as openaiResponses.responsesToChat does) never need this; it exists for the paths that forward a client body largely as-is.