yaml

@json-render/yaml

YAML wire format for @json-render/core. Progressive rendering and surgical edits via streaming YAML.

Key Concepts

• YAML wire format: Alternative to JSONL that uses code fences (yaml-spec, yaml-edit, yaml-patch, diff)
• Streaming parser: Incrementally parses YAML, emits JSON Patch operations via diffing
• Edit modes: Patch (RFC 6902), merge (RFC 7396), and unified diff
• AI SDK transform: TransformStream that converts YAML fences into json-render patches

Generating YAML Prompts

import { yamlPrompt } from "@json-render/yaml";
import { catalog } from "./catalog";

// Standalone mode (LLM outputs only YAML)
const systemPrompt = yamlPrompt(catalog, {
  mode: "standalone",
  editModes: ["merge"],
  customRules: ["Always use dark theme"],
});

// Inline mode (LLM responds conversationally, wraps YAML in fences)
const chatPrompt = yamlPrompt(catalog, { mode: "inline" });

Options:

• system (string) — Custom system message intro
• mode ("standalone" | "inline") — Output mode, default "standalone"
• customRules (string[]) — Additional rules appended to prompt
• editModes (EditMode[]) — Edit modes to document, default ["merge"]

AI SDK Transform

Use pipeYamlRender as a drop-in replacement for pipeJsonRender:

import { pipeYamlRender } from "@json-render/yaml";
import { createUIMessageStream, createUIMessageStreamResponse } from "ai";

const stream = createUIMessageStream({
  execute: async ({ writer }) => {
    writer.merge(pipeYamlRender(result.toUIMessageStream()));
  },
});
return createUIMessageStreamResponse({ stream });

For multi-turn edits, pass the previous spec:

pipeYamlRender(result.toUIMessageStream(), {
  previousSpec: currentSpec,
});

The transform recognizes four fence types:

• yaml-spec — Full spec, parsed progressively line-by-line
• yaml-edit — Partial YAML deep-merged with current spec (RFC 7396)
• yaml-patch — RFC 6902 JSON Patch lines; applied to the working spec, so later yaml-edit fences see them
• diff — Unified diff applied to serialized spec