Ruflo
原名 Claude Flow,现名 Ruflo。面向 Claude Code 的多 Agent 编排框架,将单 Agent 编程助手扩展为带记忆、自学习、跨机器联邦协作的企业级 Agent 团队。
npm: ruflo / claude-flow | License: MIT | Author: RuvNet | Web UI: flo.ruv.io | Goal Planner: goal.ruv.io
核心定位
User → CLI/MCP → Router → Swarm → Agents → Memory → LLM Providers
↑ │
└──── Learning Loop ←──────┘
Ruflo 不是 LLM,不是 Agent 本身,而是编排层:
- 将 Claude Code 的单 Agent 能力扩展为多 Agent 协作
- 提供共享记忆(HNSW vector search)、自学习(SONA)、联邦安全通信
- 自身不运行 LLM,调用 Claude/GPT/Gemini/Ollama 等外部 Provider
安装路径(两条路)
| Claude Code 插件(轻量) | CLI 完整安装 | |
|---|---|---|
| 命令 | /plugin install ruflo-core@ruflo |
npx ruflo@latest init |
| 文件侵入 | 零 | .claude/ / .claude-flow/ / CLAUDE.md |
| MCP server | 不注册 | 注册,工具可用 |
| Hooks | 无 | 27 个 hooks |
| 适合 | 体验单个插件命令 | 生产使用 |
架构总览
┌──────────────────────────────────────────────────────────┐
│ Entry Layer │
│ CLI (49 commands) + MCP Server (300+ tools) │
└────────────────────┬─────────────────────────────────────┘
│
┌────────────────────▼─────────────────────────────────────┐
│ Routing Layer │
│ Q-Learning Router + MoE (8 Experts) + 130+ Skills │
│ 27 Hooks (pre-task / post-edit / post-task / ...) │
└────────────────────┬─────────────────────────────────────┘
│
┌────────────────────▼─────────────────────────────────────┐
│ Swarm Coordination Layer │
│ SwarmCoordinator (hierarchical/mesh/ring/star) │
│ Consensus: Raft / BFT / Gossip │
│ Queen-led 层级 + Claims (Human-Agent 协调) │
└────────────────────┬─────────────────────────────────────┘
│
┌────────────────────▼─────────────────────────────────────┐
│ 100+ Specialized Agents │
│ coder / tester / reviewer / architect / security-architect │
│ 15 个固定角色(ADR-002 DDD) │
└────────────────────┬─────────────────────────────────────┘
│
┌────────────────────▼─────────────────────────────────────┐
│ Memory & Intelligence │
│ AgentDB (HNSW, 150x-12500x faster) │
│ SONA 自学习 + ReasoningBank + Hyperbolic Poincaré │
│ RuVector: Int8 量化 + LoRA/MicroLoRA │
└──────────────────────────────────────────────────────────┘
V3 模块结构(v3/ 目录)
| Package | 路径 | 职责 |
|---|---|---|
@claude-flow/cli |
v3/@claude-flow/cli/ |
CLI 入口(26 commands,140+ subcommands) |
@claude-flow/guidance |
v3/@claude-flow/guidance/ |
Governance 控制面(compile/enforce/prove/evolve) |
@claude-flow/hooks |
v3/@claude-flow/hooks/ |
17 hooks + 12 背景 workers |
@claude-flow/memory |
v3/@claude-flow/memory/ |
AgentDB + HNSW 向量搜索 |
@claude-flow/shared |
v3/@claude-flow/shared/ |
共享类型(Agent/Task/Swarm/Event 类型定义) |
@claude-flow/security |
v3/@claude-flow/security/ |
输入验证 / 路径安全 / CVE 修复 |
@claude-flow/mcp |
v3/@claude-flow/mcp/ |
MCP Server 实现 |
@claude-flow/plugins |
v3/@claude-flow/plugins/ |
插件管理器 |
@claude-flow/swarm |
v3/@claude-flow/swarm/ |
Swarm 协调引擎 |
goal_ui |
v3/goal_ui/ |
Goal-Oriented Action Planning 前端(Vite+Supabase) |
Domain-Driven Design(ADR-002)
V3 采用 DDD 架构:
SwarmCoordinator (application) — 编排协调
├── Agent (domain/entity) — Agent 实体
├── Task (domain/entity) — Task 实体
├── SwarmState (domain/value-object)
└── Consensus Mechanism (domain/service)
HybridBackend / SQLiteBackend / AgentDBBackend (infrastructure)
└── Memory domain — MemoryEntity / Memory
关键类型(v3/@claude-flow/shared/src/types.ts)
15 个固定 Agent 角色:
queen-coordinator (#1) | security-architect (#2) | security-implementer (#3)
security-tester (#4) | core-architect (#5) | core-implementer (#6)
memory-specialist (#7) | swarm-specialist (#8) | mcp-specialist (#9)
integration-architect (#10) | cli-hooks-developer (#11) | neural-learning-dev (#12)
tdd-test-engineer (#13) | performance-engineer (#14) | release-engineer (#15)
Task 类型: security-audit/fix/test, architecture-design, implementation, memory-optimization, swarm-coordination, mcp-enhancement, integration, cli-development, neural-training, test-writing, benchmark, deployment, documentation
Swarm 拓扑: hierarchical-mesh / mesh / hierarchical / centralized
共识机制: Raft(leader 维护状态)/ BFT(Byzantine Fault Tolerance)/ Gossip
核心模块详解
1. SwarmCoordinator(v3/src/coordination/application/SwarmCoordinator.ts)
协调多 Agent Swarm 的核心引擎,基于 agentic-flow 的 AttentionCoordinator 模式:
export class SwarmCoordinator {
private topology: SwarmTopology;
private agents: Map<string, Agent>;
private connections: MeshConnection[];
private eventBus: EventEmitter;
async initialize(): Promise<void>
async shutdown(): Promise<void>
async spawnAgent(config: AgentConfig): Promise<Agent>
// ...
}
支持 hierarchical(Queen 控制)和 hierarchical-mesh(Queen + peer 通信)两种拓扑。
2. MCPServer(v3/src/infrastructure/mcp/MCPServer.ts)
Model Context Protocol 服务端,暴露 300+ tools 给 Claude Code:
export class MCPServer {
private toolRegistry: Map<string, MCPTool>;
private port: number;
private host: string;
async start(): Promise<void>
async stop(): Promise<void>
registerTool(tool: MCPTool): void
}
3. CLI(v3/@claude-flow/cli/src/index.ts)
现代化 CLI 架构,支持 lazy command loading:
export class CLI {
private parser: CommandParser; // 支持嵌套 subcommand
private output: OutputFormatter;
async run(args: string[]): Promise<void>
// 命令:init / agent / swarm / memory / mcp / hooks / config / ...
}
命令行注册表(commands/index.ts):26 个同步加载 + 24 个 lazy loading commands,减少启动时间 ~200ms。
4. Hooks 系统(@claude-flow/hooks)
27 个 hooks 覆盖完整开发周期:
| 阶段 | Hooks |
|---|---|
| pre | pre-task / pre-edit / pre-commit / pre-build |
| post | post-task / post-edit / post-commit / post-build |
| workers | ultralearn / audit / optimize / testgaps / document |
| routing | route / model-select / agent-boost |
5. 自学习系统(RuVector Intelligence)
| 组件 | 功能 | 性能 |
|---|---|---|
| SONA | 自优化神经注意力,< 0.05ms | |
| EWC++ | 弹性权重固化,防遗忘 | |
| Flash Attention | 2.49x - 7.47x 加速 | |
| HNSW | 向量索引搜索 | 150x - 12500x |
| Int8 Quantization | 内存优化 | 3.92x 内存节省 |
| LoRA/MicroLoRA | 低秩适配 |
6. 12 背景 Workers
自动触发:audit / optimize / testgaps / document / map / deepdive 等。
安全特性(ADR-096 加密存储 / ADR-097 联邦预算)
- 加密存储: AES-256-GCM vault,
RFE1magic byte,CLAUDE_FLOW_ENCRYPT_AT_REST=1启用 - 高风险 store 加密: sessions / terminals / .swarm/memory.db(sql.js SQLite + ONNX embeddings)
- 联邦预算熔断:
federation_send支持budget/maxHops/spent元数据,默认maxHops: 8 - 命令注入防护: github-safe.js / statusline.js / github-tools MCP / update/executor
- 环境变量门控:
LD_PRELOAD/NODE_OPTIONS/DYLD_*在terminal_create边界被validateEnv()拒绝 - 文件权限: session/terminal/memory stores 使用
fs-secure.writeFileRestricted(mode 0600) - MCP stdin DoS 防护:
bin/mcp-server.js/bin/cli.js限制 10MB 输入
Web UI
| 产品 | URL | 功能 |
|---|---|---|
| RuFlo Web UI | flo.ruv.io | 多模型 Chat,MCP 工具调用,WASM gallery,可自托管 |
| Goal Planner | goal.ruv.io | GOAP A* 规划,plain-English → 可执行计划,live agent dashboard |
32 个插件生态
| 分类 | 插件 |
|---|---|
| Core | ruflo-core, ruflo-swarm, ruflo-autopilot, ruflo-loop-workers, ruflo-workflows, ruflo-federation |
| Memory | ruflo-agentdb, ruflo-rag-memory, ruflo-ruvector, ruflo-knowledge-graph |
| Intelligence | ruflo-intelligence, ruflo-daa, ruflo-ruvllm, ruflo-goals |
| Code Quality | ruflo-testgen, ruflo-browser, ruflo-jujutsu, ruflo-docs |
| Security | ruflo-security-audit, ruflo-aidefence |
| Architecture | ruflo-adr, ruflo-ddd, ruflo-sparc |
| DevOps | ruflo-migrations, ruflo-observability, ruflo-cost-tracker |
验证系统(verification.md)
每条修复用 Ed25519 签名(keyed off git commit)。ruflo verify 命令:
- Fetch manifest
- 重新计算每个引用文件的 SHA-256
- 从 git commit 重新派生公钥
- 验证签名
相关 Wiki 条目
- entities/vue-termui — Vue 3 TUI 框架( Ruflo 的 Web UI vs 终端 UI 对比参考)
- pi-coding-agent — TypeScript 交互式 Coding Agent 参考
- deepseek-tui — Rust TUI Coding Agent 完整参考
- concepts/agent-loop — Agent 核心执行循环模式
- concepts/agent-swarm — 多 Agent 协作架构模式
- agent-cli-tui-learning-path — Agent-CLI/TUI 构建入手路径