Langgraph Automation Framework

腓尼基星座

✦ 代理时代的建筑师工作台

腓尼基星座 将人工智能的“手动布线”转化为天体编排。通过在统一的框架中定义代理、MCP工具和RAG管道，开发人员可以在几分钟内从原始代码转移到可交付的Python捆绑包，而不是几天。

______________________________________________________________________

视觉

腓尼基星座是为原子时代建造的。开发人员成为架构师——在统一的JSON配置中定义代理、MCP工具、RAG管道、路由、内存和护栏。没有图表接线，没有样板。只是编排。

______________________________________________________________________

运作原理

                 ┌─────────────────────────────────────────┐
  English text   │  Phoenice Agentic Workbench (UI)        │
  or JSON ──────▶│  Templates · Builder · Monitor · Export │
                 └────────────────┬────────────────────────┘
                                  │ REST API
                 ┌────────────────▼────────────────────────┐
                 │  FastAPI Backend                        │
                 │  ├─ LLM Translation (LM Studio / OAI)  │
                 │  ├─ Orchestrator  (LangGraph runner)    │
                 │  ├─ Graph Factory (dynamic StateGraph)  │
                 │  ├─ Memory Manager (STM / LTM)          │
                 │  ├─ Guardrail Engine (input + output)   │
                 │  └─ Code Generator  ──▶  agent.py 🍎    │
                 └─────────────────────────────────────────┘

这 生成的捆绑包 （“苹果”）是一个独立的 agent.py +开发人员可以在任何地方部署的支持文件——不需要工作台依赖。

______________________________________________________________________

特性

______________________________________________________________________

建筑

api_backend.py         FastAPI — all REST endpoints
  ├─ config_manager.py       Load & validate config.json
  ├─ template_manager.py     Load, version, and save prompt templates
  ├─ llm_translator.py       English → JSON via LLM
  ├─ orchestrator.py         End-to-end workflow runner
  │    ├─ graph_factory.py         Build LangGraph StateGraph dynamically
  │    ├─ mcp_autobinder.py        Auto-discover & bind MCP tools
  │    ├─ memory_manager.py        STM (in-memory) + LTM (SQLite)
  │    └─ observability_manager.py Logging, tracing, metrics
  └─ code_generator.py       Generate deployable agent.py bundle

______________________________________________________________________

目录结构

backend/
  api_backend.py           FastAPI app + all REST endpoints
  orchestrator.py          Workflow runner
  graph_factory.py         Dynamic LangGraph StateGraph builder
  graph_factory_supervisor.py  Supervisor-style graph builder
  code_generator.py        Deployable agent.py bundle generator
  memory_manager.py        STM / LTM state management
  mcp_autobinder.py        MCP tool auto-discovery & binding
  tool_registry.py         Tool registry (metadata, health, versioning)
  observability_manager.py Logging, tracing, metrics, event hooks
  config_manager.py        Config loader / validator (Pydantic)
  template_manager.py      Prompt template management
  template_selector.py     Template search / filter
  llm_translator.py        LLM-powered config translation
  english_to_json.py       CLI: English → JSON (OpenAI / Gemini)
  config.json              Central backend config
  artifacts/               Per-run artifacts (auto-generated)
  prompt_templates/        Versioned enterprise prompt templates

frontend/                  React + Vite + TypeScript UI
  src/views/               Builder, Monitor, Export, Templates, Docs, Settings
  src/components/          Shared UI components

tests/                     Pytest test suite
COMMANDS.md                Developer command cheatsheet
ToDo.md                    Roadmap and task tracker

______________________________________________________________________

先决条件

______________________________________________________________________

快速入门

1·后端

# Create and activate virtual environment
python -m venv .venv
.\.venv\Scripts\Activate.ps1

# Install dependencies
pip install -r requirements.txt

# Start the backend
python -m uvicorn backend.api_backend:app --reload
# API available at http://localhost:8000

2·前端

cd frontend
npm install
npm run dev
# UI available at http://localhost:3000

3·LM工作室

1. 下载并打开 LM 工作室.
2. 加载模型（例如。 llama-3.2-3b-instruct).
3. 启动本地服务器（默认端口 1234).
4. 在菲尼斯→ 设置，单击 测试连接 以验证。

4·构建并导出您的第一个工作流程

1. 首选 模板，选择模板，单击 使用模板.
2. 调整中的节点 建造者 然后单击 跑.
3. 运行完成后，打开 出口 下载可部署的ZIP包。

______________________________________________________________________

工作流配置——快速参考

{
  "graph_name": "MyWorkflow",
  "version": "1.0",
  "state_schema": {
    "task":   { "type": "string" },
    "result": { "type": "string" }
  },
  "nodes": [
    {
      "id": "Planner",
      "type": "agent",
      "system_prompt": "Break the task into steps.",
      "next": "Executor"
    },
    {
      "id": "Executor",
      "type": "agent",
      "system_prompt": "Execute the plan.",
      "guardrails": { "pii": { "action": "redact" }, "harmful_content": { "action": "block" } },
      "output_schema": { "format": "json", "state_key": "result" },
      "next": "END"
    }
  ],
  "edges": [{ "from": "Planner", "to": "Executor" }],
  "memory": {
    "short_term": { "type": "graph_state", "max_entries": 100 },
    "long_term":  { "type": "sqlite", "ttl_days": 30 }
  },
  "retry_policy": { "max_retries": 3, "backoff_strategy": "exponential" },
  "observability_hooks": { "trace_nodes": true, "capture_agent_outputs": true }
}

看 COMMANDS.md 用于所有CLI命令和API烟雾测试。\ 查看应用内 帮助和文档 选项卡以获取完整的JSON模式参考。

______________________________________________________________________

并行执行

启用真正的并行扇出（使用LangGraph Send())通过添加a parallel_execution 阻止您的配置：

{
  "nodes": [
    { "id": "dispatcher", "type": "agent", "next": "collector" },
    { "id": "research_A", "type": "agent", "next": "collector" },
    { "id": "research_B", "type": "agent", "next": "collector" },
    { "id": "collector",  "type": "agent", "next": "END" }
  ],
  "parallel_execution": [
    {
      "group": "research_fanout",
      "nodes": ["research_A", "research_B"],
      "fan_in": "collector"
    }
  ]
}

会自动创建一个调度程序节点，以同时向所有组节点扇出。如果提供的有效节点少于2个，则优雅地回退到顺序执行。

______________________________________________________________________

本地检查点（崩溃恢复+时间旅行）

启用LangGraph的 SqliteSaver 检查点用于在每一步后保持图形状态：

{
  "checkpointing": {
    "enabled": true,
    "db_path": "artifacts/my_checkpoints.sqlite"
  }
}

每个工作流运行都会存储以下键控的状态快照 session_id.通过相同的方式恢复任何中断的跑步 session_id --该图将从中断的地方继续。对长时间运行的管道和调试很有用。

______________________________________________________________________

运行生成的捆绑包

每个出口 agent.py 是一个完全可运行的CLI脚本：

# Basic run with JSON input
python agent.py --input '{"document_text": "Analyze this document."}'

# Load input from file, output as machine-readable JSON
python agent.py --input-file data.json --output-format json

# Resume a previous session
python agent.py --session-id abc123 --resume

# Help
python agent.py --help

导出ZIP还包括 Dockerfile, docker-compose.yml，以及 .github/workflows/ci.yml 用于立即进行集装箱化部署和CI/CD。

______________________________________________________________________

生成的工件包

每次成功运行都会生成一个自包含的可部署包：

agent.py              Complete runnable agent (LangGraph + STM/LTM + guardrails)
requirements.txt      Pinned Python dependencies
docker-compose.yml    Ready-to-run Docker Compose config
.env.template         Environment variable template
run_config.json       Exact workflow config used for this run
run_status.json       Final run status and metadata
validation_report.json Code generation trust score and checks

使用以下方式部署：

pip install -r requirements.txt
python agent.py

______________________________________________________________________

测试

pytest tests/

______________________________________________________________________

许可证

MIT许可证——见 许可证.

______________________________________________________________________

有关开发人员命令备忘单，请参阅 COMMANDS.md.\ 有关路线图和未完成的任务，请参阅 ToDo.md.