Response
Wire-level constraint on what shape the model is allowed to emit. Mirrors v6's responseFormat (per call-settings.ts:103-128 → historical parity gap #20). Distinct from Output — Output is the post-decode wrapper that turns text into typed values; ResponseFormat is what the provider sees at call time.
Providers that support constrained decoding (OpenAI response_format, Anthropic output_format, on-device grammar-constrained sampling) use this to force JSON shape compliance during decoding rather than letting the model wander into prose. Providers without that capability ignore it — the schema is still useful as a hint in the prompt-injection middleware.
Two variants:
Text — default, no constraint, prose allowed.
Json — model must produce JSON; optional schema constrains shape.
Sealed interface form is the port-idiomatic shape.
Inheritors
Types
JSON-shaped response. schemaJson is the JSON-Schema constraint (optional — null means "any valid JSON object"); schemaName and schemaDescription are passed verbatim to providers that expose named-schema modes (OpenAI structured outputs, Gemini's response_schema).