Type Alias: OutputSchemaStrategy

OutputSchemaStrategy = "instruct" | "tool-forced" | "native"

Defined in: src/answer-validation/outputSchema.ts:104

How the schema reaches the model.

• 'instruct' (default) — the shape is described in the system prompt and the model is asked for JSON. Works on every provider, because it is only words. • 'tool-forced' — the shape is presented as a synthetic tool and the provider's tool choice is FORCED to it, so generation is constrained at the source instead of requested in prose. Requires a provider that declares carriesForcedToolChoice and permits the selected model through its optional supportsForcedToolChoice; a provider that does not is refused BY NAME at run start rather than quietly downgraded, because a strategy that silently falls back to the other one is config that lies. • 'native' — the provider's own structured output: the schema rides the request as Anthropic's output_config.format ({ type: 'json_schema', schema }) and the model's text answer is decoded under it. Unlike 'tool-forced' it keeps the agent's tools: the model may still call a tool first, and answers in JSON when it answers. Requires a provider that declares carriesNativeOutputSchema and permits the selected model through its optional supportsNativeOutputSchema; the same refusal BY NAME applies at run start. The schema sent is a wire copy made API-compatible by nativeOutputFormat.ts · nativeOutputFormat (every object closed with additionalProperties: false, the constraints the API does not take moved into the description); a schema the API cannot express is refused at build. The parser still judges every answer against the full contract, and a stop_reason of 'refusal', 'max_tokens' or 'model_context_window_exceeded' fails the attempt by name even when the text parses (a refusal is not re-asked).