给 Vercel AI SDK 应用加上多智能体编排

Browser (useChat)
    |
    v
POST /api/chat
    |
    +-- Phase 1: OMA runTeam()
    |     coordinator decomposes goal
    |     -> researcher agent gathers info
    |     -> writer agent drafts article
    |     (shared memory passes context between agents)
    |
    +-- Phase 2: AI SDK streamText()
    |     streams the team's output to the browser
    |
    v
useChat renders streamed response

{
  "private": true,
  "scripts": {
    "dev": "next dev",
    "build": "next build"
  },
  "dependencies": {
    "@ai-sdk/openai-compatible": "^3.0.0",
    "@ai-sdk/react": "^4.0.0",
    "@open-multi-agent/core": "^1.18.0",
    "ai": "^7.0.0",
    "next": "^16.0.0",
    "react": "^19.0.0",
    "react-dom": "^19.0.0"
  }
}

import {
  streamText,
  convertToModelMessages,
  createUIMessageStreamResponse,
  toUIMessageStream,
  type UIMessage,
} from 'ai'
import { createOpenAICompatible } from '@ai-sdk/openai-compatible'
import { OpenMultiAgent } from '@open-multi-agent/core'
import type { AgentConfig } from '@open-multi-agent/core'

export const maxDuration = 360

// --- Provider setup (swap this for your preferred LLM) ---
const BASE_URL = 'https://api.deepseek.com'
const MODEL = 'deepseek-v4-flash'

const provider = createOpenAICompatible({
  name: 'deepseek',
  baseURL: `${BASE_URL}/v1`,
  apiKey: process.env.DEEPSEEK_API_KEY,
})

// --- Agent definitions ---
const researcher: AgentConfig = {
  name: 'researcher',
  model: MODEL,
  provider: 'openai',
  baseURL: BASE_URL,
  apiKey: process.env.DEEPSEEK_API_KEY,
  systemPrompt: `You are a research specialist. Given a topic, provide thorough,
factual research with key findings, relevant data points, and important context.
Be concise but comprehensive. Output structured notes, not prose.`,
  maxTurns: 3,
  temperature: 0.2,
}

const writer: AgentConfig = {
  name: 'writer',
  model: MODEL,
  provider: 'openai',
  baseURL: BASE_URL,
  apiKey: process.env.DEEPSEEK_API_KEY,
  systemPrompt: `You are an expert writer. Using research from team members
(available in shared memory), write a well-structured, engaging article
with clear headings and concise paragraphs.`,
  maxTurns: 3,
  temperature: 0.4,
}

function extractText(message: UIMessage): string {
  return message.parts
    .filter((p): p is { type: 'text'; text: string } => p.type === 'text')
    .map((p) => p.text)
    .join('')
}

export async function POST(req: Request) {
  const { messages }: { messages: UIMessage[] } = await req.json()
  const lastText = extractText(messages.at(-1)!)

  // --- Phase 1: OMA multi-agent orchestration ---
  const orchestrator = new OpenMultiAgent({
    defaultModel: MODEL,
    defaultProvider: 'openai',
    defaultBaseURL: BASE_URL,
    defaultApiKey: process.env.DEEPSEEK_API_KEY,
  })

  const team = orchestrator.createTeam('research-writing', {
    name: 'research-writing',
    agents: [researcher, writer],
    sharedMemory: true,
  })

  const teamResult = await orchestrator.runTeam(
    team,
    `Research and write an article about: ${lastText}`,
    // Execution routing sends a short goal down the single-agent path. An
    // explicit mode outranks the router, so this keeps the demo on the
    // researcher + writer team topology.
    { mode: 'team' },
  )

  // The team path publishes the synthesized answer under 'coordinator'; the
  // single-agent path publishes it under the winning agent's own name. Read
  // both so this route survives whichever topology runs.
  const teamOutput =
    teamResult.agentResults.get('coordinator')?.output
    ?? teamResult.agentResults.get('writer')?.output
    ?? ''

  // A failed run still leaves the coordinator's unparsed plan under
  // 'coordinator', so check `success` rather than testing for an empty string.
  if (!teamResult.success || teamOutput === '') {
    return new Response(
      `The agent team did not produce an article: ${
        teamResult.errorInfo?.message ?? teamResult.status?.code ?? 'unknown error'
      }`,
      { status: 500 },
    )
  }

  // --- Phase 2: Stream result via Vercel AI SDK ---
  const result = streamText({
    model: provider(MODEL),
    system: `You are presenting research from a multi-agent team.
The team has already done the work. Relay their output faithfully
in a well-formatted way.

## Team Output
${teamOutput}`,
    messages: await convertToModelMessages(messages),
  })

  return createUIMessageStreamResponse({
    stream: toUIMessageStream({ stream: result.fullStream }),
  })
}

'use client'

import { useState } from 'react'
import { useChat } from '@ai-sdk/react'

export default function Home() {
  const { messages, sendMessage, status, error } = useChat()
  const [input, setInput] = useState('')

  const isLoading = status === 'submitted' || status === 'streaming'

  const handleSubmit = async (e: React.FormEvent) => {
    e.preventDefault()
    if (!input.trim() || isLoading) return
    const text = input
    setInput('')
    await sendMessage({ text })
  }

  return (
    <main style={{ maxWidth: 720, margin: '0 auto', padding: '32px 16px' }}>
      <h1>Research Team</h1>

      {messages.map((m) => (
        <div key={m.id} style={{ marginBottom: 24 }}>
          <strong>{m.role === 'user' ? 'You' : 'Research Team'}</strong>
          <div style={{ whiteSpace: 'pre-wrap' }}>
            {m.parts
              .filter(
                (p): p is { type: 'text'; text: string } =>
                  p.type === 'text',
              )
              .map((p) => p.text)
              .join('')}
          </div>
        </div>
      ))}

      {isLoading && status === 'submitted' && (
        <p>Agents are collaborating -- this takes a few minutes...</p>
      )}

      {error && <p style={{ color: 'red' }}>Error: {error.message}</p>}

      <form onSubmit={handleSubmit} style={{ display: 'flex', gap: 8 }}>
        <input
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Enter a topic to research..."
          disabled={isLoading}
          style={{ flex: 1, padding: '10px 14px' }}
        />
        <button type="submit" disabled={isLoading || !input.trim()}>
          Send
        </button>
      </form>
    </main>
  )
}

import { openai } from '@ai-sdk/openai'
import { AISdkAdapter } from '@open-multi-agent/core/ai-sdk'

const researcher: AgentConfig = {
  name: 'researcher',
  model: 'gpt-4o',
  adapter: new AISdkAdapter(openai('gpt-4o')),
  systemPrompt: 'You are a research specialist.',
}