Structured output, and the chronicler gap

Structured output is the interesting part. You do not want prose back, you want a typed object that matches a schema. Define the schema in zod, then bind it to the model.

import * as z from "zod/v4";

const Decision = z.object({
  messageId: z.string(),
  category: z.enum(["A", "B", "C"]),
  confidence: z.enum(["low", "medium", "high"]),
  reason: z.string().optional(),
});

const structured = model.withStructuredOutput(Decision);
const result = await structured.invoke(messages); // typed as Decision

Under the hood, withStructuredOutput uses the provider’s tool-calling or structured-output mechanism to force the model to return JSON that fits the schema, then parses it for you. That parse-and-validate step is exactly what chronicler hand-rolls today.

withStructuredOutput should still exist in the 1.5.x line; verify the exact surface at build time, since the modern alternative inside an agent loop is toolStrategy, which lesson 2 of module 2 covers.

Here is what the framework gives you for free, set against what chronicler’s requestStructuredJson does by hand:

Map it across: requestStructuredJson’s safeParse maps to withStructuredOutput’s built-in parse. The four-attempt backoff maps, roughly, to LangChain’s model-level maxRetries / retry config. The fence-stripping is mostly unnecessary once you use tool-calling structured output, because the model returns structured content, not a markdown code block. The preservation-bias default is business logic that stays yours either way, no framework replaces a product decision.