swarm-agent 0.1.4

Lightweight, concurrency-safe, multi-tenant Feishu AI Agent framework

Swarm

多租户飞书 AI Agent 框架

并发安全 · 轻量级 · 面向企业

---

中文 · English

---

Swarm 中文

Swarm 是一个轻量级、并发安全的多租户飞书 AI Agent 框架。提供简洁的 Agent 循环加上生产环境所需组件：WebSocket 长连接、记忆系统、工具系统、用户 OAuth 权限、定时任务和完善的日志系统。

借鉴与致谢

Swarm 的项目结构和核心设计参考了两个优秀的 AI Agent 项目：

• nanobot — Agent 循环、工具系统、消息总线的设计灵感
• hermes-agent — 扁平化项目结构、插件协议、技能系统

为什么选择 Swarm

• 多租户隔离 — 基于 chat_id 的完全隔离，存储、运行时、权限三层防护
• 并发安全 — RequestContext 显式传递 + 每会话 asyncio.Lock，杜绝全局状态竞态
• WebSocket 原生 — 无需公网 IP、无需内网穿透、无需开放端口
• 用户身份操作 — 飞书 OAuth 原生支持，工具以用户身份执行而非应用身份
• 小核心 — 可读的内部实现，易于理解和自托管
• 企业就绪 — 结构化日志、审计追踪、优雅关闭、Docker 部署

快速开始

git clone https://github.com/poplangfan/swarm.git ~/.swarm
cd ~/.swarm
pip install -e .
swarm init
# 编辑 ~/.swarm/config.yaml 填入凭证
swarm chat

安装后 ~/.swarm/ 包含完整源码 + 所有数据：

~/.swarm/
├── agent/            # 源代码
├── gateway/          # 源代码
├── tools/            # 源代码
├── cli/              # 源代码
├── ...               # 其他模块
├── config.yaml       # 配置文件
├── sessions.db       # 会话数据
├── chroma/           # 长期记忆
├── logs/             # 日志
└── skills/           # 自定义技能

也可以从 PyPI 安装：pip install swarm-agent && swarm init

# config.yaml
llm:
  provider: "anthropic"
  base_url: "https://api.deepseek.com/anthropic"
  api_key: "${LLM_API_KEY}"
  model: "deepseek-v4-pro"

feishu:
  app_id: "${FEISHU_APP_ID}"
  app_secret: "${FEISHU_APP_SECRET}"

核心特性

架构

• 5 状态 AgentLoop: RESTORE → BUILD → RUN → SAVE → RESPOND
• 消息总线: asyncio.Queue 解耦 I/O 和 Agent 逻辑
• RequestContext: 不可变请求上下文，杜绝全局可变状态
• ContextBuilder: 分层系统提示组装（identity + memory + skills）

多租户隔离

chat_id=A → asyncio.Lock(A) → ChromaDB:mem_A → SQLite:session_A
chat_id=B → asyncio.Lock(B) → ChromaDB:mem_B → SQLite:session_B
                    ↑ 并发                     ↑ 零交叉污染

记忆系统

• 短期记忆: SQLite 按 chat_id 分区存储
• 长期记忆: ChromaDB 向量存储，按 chat_id 隔离
• Dream: 两阶段 LLM 辅助记忆巩固
• Recall: 混合召回（向量相似度 + 时间衰减 + 重要性加权）
• 知识图谱: 跨会话实体和关系提取
• Compressor: 轨迹保留的上下文压缩

用户授权

首次消息 → OAuth 授权链接
用户点击 → 飞书授权
Swarm 用 code 换取 user_access_token
Token 加密存储 (AES-256-GCM) + 自动刷新
后续消息 → 用户身份注入工具调用

工具系统

• 内置工具：网页搜索、URL 抓取、飞书消息、文件管理、定时任务
• MCP 协议：客户端 + 服务端，与其他 AI Agent 互操作
• Markdown Skills：无代码 Agent 行为定义（10 个内置）
• 插件协议：安装/启用/禁用生命周期
• 权限系统：声明式工具访问控制

飞书集成

• WebSocket 长连接（无需 webhook）
• 全消息类型：文本、富文本、图片、音频、文件、贴纸、交互卡片、分享、合并转发
• CardKit 流式输出
• 交互卡片动作处理
• 群管理事件（加入/退出）
• Emoji 状态反馈

运维

• 结构化 JSON 日志（structlog）
• 日志轮转 + gzip 压缩
• 审计日志
• 优雅关闭 + 断点保存
• Docker 部署
• systemd/supervisor 服务模板

架构图

飞书 WS → 消息总线 → AgentLoop(5状态) → 消息总线 → Delivery → 飞书回复
                         │
               ┌─────────┼──────────┐
               ▼         ▼          ▼
           Session    Memory      Tools
           (SQLite)  (ChromaDB)  (Registry)

               Auth (OAuth + 加密Token存储)

项目结构

swarm/
├── agent/         # AgentLoop, Runner, Context, Subagent
├── session/       # Session CRUD, Goal State, Continuation
├── memory/        # ChromaDB, Short-term, Dream, Compressor, KG
├── providers/     # OpenAI, Anthropic, Fallback, Retry
├── gateway/       # Feishu WS, Messages, Reply, Streaming, Events
├── tools/         # Registry, Builtins (web, message, file, cron)
├── auth/          # OAuth, Token Store (AES-256-GCM 加密)
├── events/        # Pub/Sub 事件总线, Metrics, Logging
├── cron/          # APScheduler, SQLite 持久化, NL 解析
├── mcp/           # MCP 客户端 + 服务端
├── plugins/       # 清单协议, 加载器
├── skills/        # Markdown skills 加载器
├── skills_builtin/ # 10 个内置技能
├── delivery/      # 出站队列, 限流, 重试
├── state/         # 持久化状态, 快照, 迁移
├── cli/           # Typer + Rich + prompt_toolkit
├── bus/           # 异步消息总线
├── config/        # Pydantic v2 schema + YAML 加载器
├── docker/        # Dockerfile + Compose
├── tests/         # 315 单元/集成测试
└── logging_/      # structlog + 轮转 + 压缩 + 审计

文档

• 快速开始
• 架构指南
• 配置说明
• 飞书设置
• 认证授权
• 工具开发
• Skills 开发
• 插件开发
• 日志监控
• 部署指南

环境要求

• Python 3.10+
• 飞书账号（可创建应用）
• LLM API Key（DeepSeek、OpenAI 或 Anthropic）

协议

MIT License — 详见 LICENSE

---

Swarm English

Swarm is a lightweight, concurrency-safe, multi-tenant Feishu AI Agent framework. It provides a clean agent loop with production-ready components: WebSocket long connection, memory system, tool system, user OAuth permissions, cron scheduler, and comprehensive logging.

Acknowledgments

Swarm's project structure and core design draw inspiration from two excellent AI agent projects:

• nanobot — Agent loop, tool system, and message bus design
• hermes-agent — Flat project structure, plugin protocol, and skill system

Why Swarm

• Multi-tenant isolation — chat_id-based isolation at storage, runtime, and auth layers
• Concurrency-safe — RequestContext explicit passing, per-chat asyncio.Lock, no global state races
• WebSocket native — No public IP, no ngrok, no open ports needed
• User-permissioned — Native Feishu OAuth — tools operate as the requesting user, not as the bot
• Small core — Readable internals, easy to understand, extend, and self-host
• Enterprise-ready — Structured logging, audit trail, graceful shutdown, Docker deployment

Quick Start

git clone https://github.com/poplangfan/swarm.git ~/.swarm
cd ~/.swarm
pip install -e .
swarm init
# Edit ~/.swarm/config.yaml with your credentials
swarm chat

After setup, ~/.swarm/ contains both full source code and all data.

Also available via PyPI: pip install swarm-agent && swarm init

# config.yaml
llm:
  provider: "anthropic"
  base_url: "https://api.deepseek.com/anthropic"
  api_key: "${LLM_API_KEY}"
  model: "deepseek-v4-pro"

feishu:
  app_id: "${FEISHU_APP_ID}"
  app_secret: "${FEISHU_APP_SECRET}"

Core Features

Architecture

• 5-State AgentLoop: RESTORE → BUILD → RUN → SAVE → RESPOND
• Message Bus: asyncio.Queue decoupling for I/O and agent logic
• RequestContext: Immutable per-request isolation — no global mutable state
• ContextBuilder: Layered system prompt assembly (identity + memory + skills)

Multi-Tenant Isolation

chat_id=A → asyncio.Lock(A) → ChromaDB:mem_A → SQLite:session_A
chat_id=B → asyncio.Lock(B) → ChromaDB:mem_B → SQLite:session_B
                    ↑ concurrent             ↑ zero cross-contamination

Memory System

• Short-term: SQLite per-chat message store
• Long-term: ChromaDB vector memory with per-chat collections
• Dream: Two-phase LLM-assisted memory consolidation
• Recall: Hybrid search (vector similarity + time decay + importance)
• Knowledge Graph: Cross-session entity and relationship extraction
• Compressor: Trajectory-preserving context window management

User Authorization

First message → OAuth authorization link
User clicks link → authorizes in Feishu
Swarm exchanges code for user_access_token
Token encrypted (AES-256-GCM) + auto-refresh
Subsequent messages → user identity injected into tools

Tools & Skills

• Built-in tools: Web search, URL fetch, Feishu messages, file management, cron
• MCP: Client + server — interoperate with other AI agents
• Markdown Skills: Code-free agent behaviors (10 built-in)
• Plugins: Manifest protocol with install/enable/disable lifecycle
• Permissions: Declarative access control for tools

Feishu Integration

• WebSocket long connection (no webhook needed)
• All message types: text, post, image, audio, file, sticker, interactive, share, merge_forward
• CardKit streaming output engine
• Interactive card action handling
• Group chat member join/leave events
• Emoji reactions for processing state feedback

Operations

• Structured JSON logging (structlog)
• Log rotation with gzip compression
• Audit logging for compliance
• Graceful shutdown with checkpoint preservation
• Docker deployment
• systemd/supervisor service templates

Architecture

Feishu WS → MessageBus → AgentLoop(5 states) → MessageBus → Delivery → Feishu Reply
                              │
                    ┌─────────┼──────────┐
                    ▼         ▼          ▼
               Session    Memory      Tools
               (SQLite)  (ChromaDB)  (Registry)

                    Auth (OAuth + Encrypted Token Store)

Project Structure

swarm/
├── agent/         # AgentLoop, Runner, Context, Subagent
├── session/       # Session CRUD, Goal State, Continuation
├── memory/        # ChromaDB, Short-term, Dream, Compressor, KG
├── providers/     # OpenAI, Anthropic, Fallback, Retry
├── gateway/       # Feishu WS, Messages, Reply, Streaming, Events
├── tools/         # Registry, Builtins (web, message, file, cron)
├── auth/          # OAuth, Token Store (AES-256-GCM encryption)
├── events/        # Pub/Sub event bus, Metrics, Logging subscribers
├── cron/          # APScheduler, SQLite persistence, NL parsing
├── mcp/           # MCP client + server
├── plugins/       # Manifest protocol, Loader
├── skills/        # Markdown skills loader
├── skills_builtin/ # 10 built-in skills
├── delivery/      # Outbound queuing, rate limiting, retry
├── state/         # Persistent state, snapshots, migration
├── cli/           # Typer + Rich + prompt_toolkit
├── bus/           # Async message bus
├── config/        # Pydantic v2 schema + YAML loader
├── docker/        # Dockerfile + Compose
├── tests/         # 315 unit/integration tests
└── logging_/      # structlog + rotation + compression + audit

## Documentation

- [Quickstart Guide](docs/quickstart.md)
- [Architecture Guide](docs/architecture.md)
- [Configuration](docs/configuration.md)
- [Feishu Setup](docs/feishu-setup.md)
- [Authentication](docs/auth.md)
- [Tools Guide](docs/tools.md)
- [Skills Guide](docs/skills.md)
- [Plugin Development](docs/plugins.md)
- [Logging & Monitoring](docs/logging.md)
- [Deployment Guide](docs/deployment.md)

## Requirements

- Python 3.10+
- Feishu account with app creation permissions
- LLM API key (DeepSeek, OpenAI, or Anthropic)

## License

MIT License — see [LICENSE](LICENSE)