Skip to main content
ELI5: declare the shape you want per stage; the harness picks each provider’s native structured-output mechanism (Anthropic output_format: json_schema, Gemini structured output, OpenAI structured responses) and validates the result. You never write provider-specific JSON-mode flags.

​
Where it lives: per-stage output_schema

There is deliberately no top-level structured_output field. Structure belongs to stages — a stage that produces data declares its shape: 
state_machine:
  start: extract
  stage_order: [extract, report]
  states:
    extract:
      output_schema:
        type: object
        properties:
          issues:
            type: array
            items:
              type: object
              properties:
                file: { type: string }
                severity: { type: string }
                summary: { type: string }
              required: [file, severity]
        required: [issues]
  on_event:
    extract_completed: report
  terminal_states: [done]
The extract stage must emit JSON matching this schema; the next stage receives it as typed input. Supported type values: object, array, string, number, boolean, with recursive properties/items and required.

​
Provider support

Implemented and probed on Anthropic (output_format {type: json_schema} and strict JSON mode), Gemini (structured output), and the OpenAI Responses path. Local models honor the shape on a best-effort basis — validate downstream if your local model matters.

​
When to use

TaskUse structured output?
Stage feeds another stage or a programYes — always declare output_schema
Extraction, classification, triageYes
Prose deliverables (reports, reviews)No — let the stage write markdown to the workpad

​
See also