Bodhi Realtime Agent：基于 Gemini Live API 的实时语音智能体框架

Bodhi Realtime Agent Framework

TypeScript framework for building real-time voice agent applications using the Google Gemini Live API.

Features

• Real-time voice: Bidirectional audio streaming with Gemini Live API and server-side turn detection
• Multi-agent: Define multiple agents with distinct personas and tool sets; transfer between them mid-conversation
• Function tools: Inline (blocking) and background (non-blocking) tool execution with Zod validation
• Background subagents: Long-running tool calls hand off to Vercel AI SDK subagents while Gemini keeps talking
• Memory: Automatic extraction and persistence of durable user facts across sessions
• Session resumption: Transparent reconnection via Gemini resumption handles and audio buffering
• Observability: Type-safe EventBus and lifecycle hooks for logging, metrics, and debugging

Requirements

• Node.js >= 22
• A Google API key with Gemini Live API access
• pnpm (recommended)

Installation

pnpm add bodhi-realtime-agent

To integrate the framework into your own backend: use MultiClientTransport to accept WebSockets, build a SessionClientSender per connection, and create a VoiceSession with that sender plus your agents and tools. Runnable demos live in examples/ — start with pnpm tsx examples/hello_world/agent.ts and pnpm web-client.

Documentation

Core Concepts

VoiceSession

The top-level integration hub. It wires together all framework components and manages the full session lifecycle:

Client App  <--WebSocket-->  ClientTransport  <--audio-->  GeminiLiveTransport  <--WebSocket-->  Gemini Live API
                                    |                              |
                                    +--------- VoiceSession -------+
                                    |    (audio fast-path relay)    |
                                    |                              |
                              AgentRouter    ToolExecutor    ConversationContext

Audio flows on a fast-path directly between the client and Gemini transports, bypassing the EventBus for minimal latency.

Agents

Agents are the top-level personas that Gemini assumes. Each agent has its own system instructions and tool set.

const mainAgent: MainAgent = {
  name: 'main',
  instructions: 'You are a helpful assistant.',
  tools: [myTool],
  onEnter: async (ctx) => { /* agent activated */ },
  onExit: async (ctx) => { /* agent deactivated */ },
  onTurnCompleted: async (ctx, transcript) => { /* turn finished */ },
};

Agent transfers are triggered by a special transfer_to_agent tool. The framework intercepts this tool call automatically, disconnects from Gemini, reconnects with the new agent's config, and replays conversation context:

const transferToExpert: ToolDefinition = {
  name: 'transfer_to_agent',
  description: 'Transfer to the expert agent.',
  parameters: z.object({
    agent_name: z.literal('expert'),
  }),
  execution: 'inline',
  execute: async () => ({ status: 'transferred' }),
};

Tools

Tools are declared with a Zod schema (for both Gemini declaration and runtime validation) and an execution mode:

const myTool: ToolDefinition = {
  name: 'lookup',
  description: 'Look up information.',
  parameters: z.object({ query: z.string() }),
  execution: 'inline', // or 'background'
  timeout: 10_000,      // optional, default 30s
  execute: async (args, ctx) => {
    // ctx.abortSignal is triggered on cancellation/timeout
    return { answer: '42' };
  },
};

Session State Machine

Sessions follow a strict state machine:

CREATED --> CONNECTING --> ACTIVE --> RECONNECTING --> ACTIVE
                            |                           |
                        TRANSFERRING --> ACTIVE       CLOSED
                            |
                          CLOSED

• RECONNECTING: Triggered by GoAway signals or unexpected disconnects. Audio is buffered and replayed.
• TRANSFERRING: Active during agent transfers. Client audio is buffered until the new agent is connected.

EventBus

A type-safe, synchronous event bus for loose coupling between components:

session.eventBus.subscribe('agent.transfer', (payload) => {
  console.log(`Transfer: ${payload.fromAgent} -> ${payload.toAgent}`);
});

session.eventBus.subscribe('tool.result', (payload) => {
  console.log(`Tool ${payload.toolName}: ${payload.result}`);
});

Available events: session.start, session.close, session.stateChange, session.goaway, turn.start, turn.end, turn.interrupted, generation.start, generation.end, agent.enter, agent.exit, agent.transfer, agent.handoff, tool.call, tool.result, tool.cancel, gui.update, gui.notification.

GUI Events

The client WebSocket carries both audio and GUI events on the same connection using the native binary/text frame distinction:

• Binary frames: Raw PCM audio (16-bit, 16 kHz, mono)
• Text frames: JSON messages for GUI events

Server → Client (text frames):

{ "type": "gui.update",       "payload": { "sessionId": "...", "data": { ... } } }
{ "type": "gui.notification",  "payload": { "sessionId": "...", "message": "..." } }
{ "type": "ui.payload",       "payload": { "type": "choice", "requestId": "...", "data": { ... } } }

Client → Server (text frames):

{ "type": "ui.response", "payload": { "requestId": "...", "selectedOptionId": "..." } }

GUI events published on the EventBus (gui.update, gui.notification, subagent.ui.send) are automatically forwarded to the connected client. Client ui.response messages are published back to the EventBus as subagent.ui.response events, closing the loop for interactive subagent UIs.

Apps can send their own frames with session.sendJsonToClient(...) (or ctx.sendJsonToClient?.(...) inside a tool). Any JSON object is accepted as an application frame as long as its type is not a core frame type; core frames must keep their exact shape. See Custom frames from your app.

Hooks

Lifecycle hooks for observability (logging, metrics, alerting):

const session = new VoiceSession({
  // ...
  hooks: {
    onSessionStart: (e) => console.log(`Session started: ${e.sessionId}`),
    onSessionEnd: (e) => console.log(`Session ended after ${e.durationMs}ms`),
    onToolCall: (e) => console.log(`Tool: ${e.toolName} (${e.execution})`),
    onToolResult: (e) => console.log(`Result: ${e.status} in ${e.durationMs}ms`),
    onAgentTransfer: (e) => console.log(`${e.fromAgent} -> ${e.toAgent}`),
    onError: (e) => console.error(`[${e.component}] ${e.error.message}`),
  },
});

Memory

The memory system automatically extracts durable facts about the user from conversation using a merge-on-write strategy — each extraction produces the complete updated fact list (deduped, contradictions resolved):

import { JsonMemoryStore } from 'bodhi-realtime-agent';

const session = new VoiceSession({
  // ...required config
  memory: {
    store: new JsonMemoryStore('./memory'),
  },
});

Facts are persisted as JSON files (memory/{userId}.json) with structured directives and categorized facts:

{
  "directives": { "pacing": "slow" },
  "facts": [
    { "content": "Prefers dark mode", "category": "preference" },
    { "content": "Works at Acme Corp", "category": "entity" }
  ]
}

Project Structure

src/
  core/              # Central orchestration
    voice-session.ts     # Top-level integration hub
    session-manager.ts   # Session state machine
    event-bus.ts         # Type-safe event system
    conversation-context.ts  # Conversation timeline + context
    conversation-history-writer.ts  # EventBus-driven persistence
    hooks.ts             # Lifecycle hook manager
    session-store.ts     # Session checkpoint persistence
    errors.ts            # Error class hierarchy
  agent/             # Agent management
    agent-router.ts      # Agent transfers and subagent handoffs
    agent-context.ts     # Runtime context for agent hooks
    subagent-runner.ts   # Background subagent execution (AI SDK)
  tools/             # Tool execution
    tool-executor.ts     # Zod validation, timeout, cancellation
  transport/         # Network layer
    gemini-live-transport.ts  # Gemini Live API WebSocket
    client-transport.ts       # Client-facing WebSocket server
    audio-buffer.ts           # Bounded ring buffer for audio
    zod-to-schema.ts          # Zod → Gemini JSON Schema converter
  memory/            # User memory
    json-memory-store.ts      # JSON file-based memory persistence
    memory-distiller.ts       # LLM-powered fact extraction (merge-on-write)
    prompts.ts                # Extraction prompt template
  types/             # TypeScript interfaces and type definitions
test/                # Unit and integration tests (mirrors src/ structure)
examples/            # Runnable demos (see examples/README.md)

Development

pnpm install        # Install dependencies (single package — no nested workspaces)
pnpm build          # Build with tsup (ESM + CJS + declarations)
pnpm test           # Run tests with vitest
pnpm test:watch     # Run tests in watch mode
pnpm lint           # Check with Biome
pnpm lint:fix       # Auto-fix lint issues
pnpm typecheck      # TypeScript type checking

Optional example-only setup (Python venv for SpatialReal, and similar) is not part of pnpm install; run pnpm examples:setup and see examples/README.md.

Publishable browser embed packages (@bodhi_lab/bodhi-sdk-core, @bodhi_lab/bodhi-web-sdk, @bodhi_lab/bodhi-widget) live in a separate git repository, bodhi-lab-client-packages, typically cloned as a sibling of this repo.

Integration Tests

E2E tests require a Google API key and are skipped by default:

GOOGLE_API_KEY=your_key pnpm test

License

UNLICENSED