Shadowboxin

Shadowboxin: Utils for real-time streamed LLM-generated UIs.
Bring your own component registry (BYO) and let an engine render UI described by model output.

Beta

This is an active WIP, and probably unrealiable right now

Modules (public subpaths)

• shadowboxin/client – prompt + adapter/provider wiring
• shadowboxin/engine – runtime that renders model-emitted UI
• shadowboxin/template-models – schema helpers/validators
• shadowboxin/template-utils – registry helpers and component bindings

Install

npm install shadowboxin zod react react-dom

Assumes a React app (Vite recommended). Shadowboxin ships ESM with subpath exports only (no root entry).

Quick Start

1) Create a tiny component registry (Container, Paragraph, Button)

Create src/sbx/registry.tsx:

import { z } from "zod"
import type { ComponentType, ReactNode } from "react"

import { createTemplateRegistry } from "shadowboxin/template-utils"
import { templateValidatorFor, parentValidatorFor } from "shadowboxin/template-models"

/** ───────── Paragraph ───────── */
type ParagraphProps = { content: string }
const Paragraph: ComponentType<ParagraphProps> = ({ content }) => <p>{content}</p>

export const ParagraphSet = {
  type: "paragraph",
  component: Paragraph,
  templateValidator: templateValidatorFor("paragraph", { content: z.string().min(1) }),
  instructions: {
    generalUsage: "Use for body copy.",
    fields: { content: "The paragraph text." }
  }
}

/** ───────── Container ───────── */
type ContainerProps = { children?: ReactNode }
const Container: ComponentType<ContainerProps> = ({ children }) => (
  <div data-sbx="container">{children}</div>
)

export const ContainerSet = {
  type: "container",
  component: Container,
  templateValidator: parentValidatorFor("container"),
  instructions: { generalUsage: "Groups other elements." }
}

/** ───────── Button ───────── */
type ButtonProps = { label: string; onSubmit: (payloads: unknown[]) => void }
const Button: ComponentType<ButtonProps> = ({ label, onSubmit }) => (
  <button type="button" onClick={() => onSubmit([])}>{label}</button>
)

export const ButtonSet = {
  type: "button",
  component: Button,
  templateValidator: templateValidatorFor("button", { label: z.string().min(1) }),
  instructions: {
    generalUsage: "Simple submit button; no fields.",
    fields: { label: "Button text." }
  }
}

/** ───────── Registry ───────── */
export const registry = createTemplateRegistry(ContainerSet, ParagraphSet, ButtonSet)

2) Wire the engine and a tiny local provider (no network)

Create src/sbx/setup.ts:

import {
  createAgentsProvider,
  createEngineAdapter,
  generateSystemPrompt,
  type ChatMessage,
  type EnginePort
} from "shadowboxin/client"

import { createEngine } from "shadowboxin/engine"
import { registry } from "./registry"

// Minimal example JSON the assistant could emit (Container → Paragraph + Button)
export const minimalExample = {
  id: "container",
  children: [
    { id: "paragraph", content: "Hello from Shadowboxin." },
    { id: "button", label: "Continue" }
  ]
} as const

// Build a system prompt using your registry instructions and the example JSON
const SYSTEM_PROMPT = generateSystemPrompt(
  "You respond by emitting UI templates only.",
  registry.instructions,
  { exampleJSON: minimalExample }
)

/** Local demo provider that yields one JSON response (no network). */
const provider = createAgentsProvider(async function* (_messages: ChatMessage[]) {
  // In production, plug in your streamed LLM responses here and yield deltas.
  yield JSON.stringify(minimalExample)
})

export async function startShadowboxin(root: HTMLElement) {
  let adapter!: ReturnType<typeof createEngineAdapter>

  const engine = await createEngine({
    registry,
    rootNode: root,
    onSubmit: (payloads) => adapter.submit(payloads),
    debug: false
  })

  const port: EnginePort = {
    next: (d) => engine.push(d),
    reset: () => engine.reset()
  }

  adapter = createEngineAdapter(port, provider, {
    systemPrompt: SYSTEM_PROMPT,
    initialUserMessage: "Render the minimal example."
  })

  await adapter.run()
}

3) Mount it in your app

Create src/main.tsx:

import { StrictMode, useEffect, useRef } from "react"
import { createRoot } from "react-dom/client"
import { startShadowboxin } from "./sbx/setup"

function App() {
  const ref = useRef<HTMLDivElement>(null)
  useEffect(() => {
    if (ref.current) startShadowboxin(ref.current)
  }, [])
  return <div ref={ref} id="gen-ui-root" />
}

createRoot(document.getElementById("root")!).render(
  <StrictMode>
    <App />
  </StrictMode>
)

Example UI JSON (what the assistant/provider can emit)

Create src/sbx/example.json (optional, for reference):

{
  "id": "container",
  "children": [
    { "id": "paragraph", "content": "Hello from Shadowboxin." },
    { "id": "button", "label": "Continue" }
  ]
}

How it works

• Registry exposes types/instructions/validators for your components.
• System Prompt is generated from registry instructions + an example JSON.
• Provider yields model output (as JSON strings); swap in your streamed LLM.
• Adapter + Engine connect the stream to the renderer and handle user onSubmit.