Task mode

task is run with a contract: the final answer must match your JSON schema. Tools run as usual along the way. Providers constrain the answer natively where they can, and the result is validated client-side everywhere, so the guarantee holds on all three providers.

schema = {
  type: "object",
  properties: {
    "calories" => { type: "integer" },
    "macros" => {
      type: "object",
      properties: {
        "protein_g" => { type: "number" },
        "carbs_g" => { type: "number" },
        "fat_g" => { type: "number" },
      },
      required: %w[protein_g carbs_g fat_g],
    },
  },
  required: %w[calories macros],
}

result = agent.task("Analyze the nutrition of this recipe.", schema: schema)
result.output  # => { "calories" => 540, "macros" => {...} }, parsed and validated

The fix loop

A validation failure goes back to the model with the exact violations, once by default (fixes: 1), then raises. Nothing silently passes through:

agent.task(input, schema: schema, fixes: 2)

A structured task is a real turn each time, so the fix exchange is visible in the session like everything else.