39fe248f9f
- 新增 .claude/skills/new-project/SKILL.md(标准 skill 格式) - install-skill.sh 安装时组装 SKILL.md + verify.sh,单一源无重复 - verify.sh 改用索引数组替代 declare -A,兼容 macOS bash 3.2 - log_pass/log_skip 显式 return 0,避免 set -e 下非 verbose 误退出 - README 改为 skill 优先;补全 copier.yml/scripts/template 入库
186 lines
7.1 KiB
Django/Jinja
186 lines
7.1 KiB
Django/Jinja
---
|
||
paths:
|
||
- "agent/**"
|
||
---
|
||
|
||
# Agent 服务总览
|
||
|
||
Agent 服务的职责、技术栈、目录结构、配置{% if llm_provider == 'gemini-cli' %}、Skill 安装{% endif %}与测试规范。
|
||
|
||
## 概览(职责)
|
||
|
||
FastAPI 服务(port {{ agent_port }})。核心职责:{% if llm_provider == 'gemini-cli' %}PTY 桥接 LLM CLI 交互式会话{% else %}通过 SDK 调用 LLM API{% endif %} + DAG 编排(对话 → 风格预览 → 生成 → 截图 → 合成)+ 事件流式推送给 Backend。
|
||
|
||
## 技术栈
|
||
|
||
| 库 | 用途 |
|
||
|----|------|
|
||
| FastAPI + uvicorn | WebSocket 服务端 |
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
| pty(标准库) | 伪终端,让 LLM CLI 认为自己在真实终端 |
|
||
| asyncio | 异步 I/O,协调 PTY 读写与 WebSocket |
|
||
{% elif llm_provider == 'openai-api' -%}
|
||
| openai | OpenAI SDK,流式生成(`AsyncOpenAI`) |
|
||
| asyncio | 异步 I/O,协调 SDK 调用与 WebSocket |
|
||
{% elif llm_provider == 'anthropic-api' -%}
|
||
| anthropic | Anthropic SDK,流式生成(`AsyncAnthropic`) |
|
||
| asyncio | 异步 I/O,协调 SDK 调用与 WebSocket |
|
||
{% else -%}
|
||
| asyncio | 异步 I/O,协调 LLM 调用与 WebSocket |
|
||
{% endif -%}
|
||
| playwright(async) | 截取每页幻灯片截图 |
|
||
| python-pptx | 合成 PPTX |
|
||
| loguru | 结构化日志 |
|
||
| pytest + pytest-asyncio | 测试 |
|
||
|
||
## 包管理(uv)
|
||
|
||
```bash
|
||
# 添加依赖
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
uv add fastapi uvicorn playwright python-pptx loguru pydantic-settings pytest pytest-asyncio
|
||
{% elif llm_provider == 'openai-api' -%}
|
||
uv add fastapi uvicorn openai playwright python-pptx loguru pydantic-settings pytest pytest-asyncio
|
||
{% elif llm_provider == 'anthropic-api' -%}
|
||
uv add fastapi uvicorn anthropic playwright python-pptx loguru pydantic-settings pytest pytest-asyncio
|
||
{% else -%}
|
||
uv add fastapi uvicorn playwright python-pptx loguru pydantic-settings pytest pytest-asyncio
|
||
{% endif %}
|
||
# 安装 Playwright 浏览器(首次)
|
||
uv run playwright install chromium
|
||
|
||
# 运行服务
|
||
uv run python -m src.main
|
||
|
||
# 运行测试
|
||
uv run pytest
|
||
```
|
||
|
||
关键文件:`pyproject.toml`(依赖声明)、`uv.lock`(锁文件,必须提交 git)
|
||
|
||
## 数据库
|
||
|
||
**Agent 不操作数据库,不引入 SQLAlchemy / Alembic。**
|
||
|
||
Agent 是无状态处理服务。DB 由 Backend 单一持有。checkpoint 信息由 Backend 在 WebSocket 建连时作为参数传入,Agent 只读取,不写入任何持久化存储。
|
||
|
||
## 目录结构(clean-arch)
|
||
|
||
```
|
||
agent/
|
||
├── src/
|
||
│ ├── main.py # 接口适配器:FastAPI app,WebSocket 端点
|
||
│ │
|
||
│ ├── domain/ # 领域层:纯逻辑零 IO
|
||
│ │ ├── models.py
|
||
│ │ ├── state.py # 聚合根状态
|
||
│ │ ├── ports.py # 端口定义(LLMSession、Exporter 等)
|
||
│ │ └── exceptions.py
|
||
│ │
|
||
│ ├── dag/ # 应用层:用例编排
|
||
│ │ ├── nodes.py
|
||
│ │ └── runner.py
|
||
│ │
|
||
│ ├── llm/ # 扩展点①:LLMSession 适配器
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
│ │ ├── pty_bridge.py # 共享 PTY 机制
|
||
│ │ └── gemini_cli.py # GeminiCliSession 实现
|
||
{% elif llm_provider == 'openai-api' -%}
|
||
│ │ └── openai_api.py # OpenAISession 实现
|
||
{% elif llm_provider == 'anthropic-api' -%}
|
||
│ │ └── anthropic_api.py # AnthropicSession 实现
|
||
{% else -%}
|
||
│ │ └── custom_llm.py # 自定义 LLMSession 实现
|
||
{% endif -%}
|
||
│ │ └── registry.py
|
||
│ ├── providers/ # 扩展点②:ImageProvider 适配器
|
||
│ │ └── registry.py
|
||
│ ├── export/ # 扩展点③:Exporter 适配器
|
||
│ │ └── pptx_exporter.py
|
||
│ ├── storage/ # 扩展点④:StorageBackend 适配器
|
||
│ │ └── local_storage.py
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
│ ├── skills/
|
||
│ │ └── installer.py # CLI skill 自动安装
|
||
{% endif -%}
|
||
│ ├── utils/
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
│ │ ├── ansi.py # ANSI 转义码过滤
|
||
{% endif -%}
|
||
│ │ ├── concurrency.py
|
||
│ │ ├── files.py
|
||
│ │ └── ids.py
|
||
│ ├── middleware/
|
||
│ └── core/
|
||
│ ├── config.py
|
||
│ └── logging.py
|
||
└── tests/
|
||
├── conftest.py
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
├── test_pty_bridge.py
|
||
├── test_gemini_cli.py
|
||
├── test_installer.py
|
||
{% else -%}
|
||
├── test_llm_session.py
|
||
{% endif -%}
|
||
├── test_nodes.py
|
||
└── test_runner.py
|
||
```
|
||
|
||
## 配置(core/config.py)
|
||
|
||
```python
|
||
class Settings(BaseSettings):
|
||
env: Literal["dev", "prod"] = "dev"
|
||
work_dir: Path = Path.home() / "{{ project_slug }}"
|
||
port: int = {{ agent_port }}
|
||
log_level: str = "DEBUG"
|
||
{% if llm_provider == 'gemini-cli' %}
|
||
llm_cmd: str = "gemini"
|
||
skill_name: str = "frontend-slides"
|
||
skill_repo: str = "" # 从 .env 注入
|
||
pty_read_buffer_size: int = 4096
|
||
{% elif llm_provider == 'openai-api' %}
|
||
llm_api_key: str = "" # 从 .env 注入,不硬编码
|
||
llm_model: str = "gpt-4o"
|
||
llm_base_url: str = "" # 非空时覆盖 SDK 默认 endpoint
|
||
{% elif llm_provider == 'anthropic-api' %}
|
||
llm_api_key: str = "" # 从 .env 注入,不硬编码
|
||
llm_model: str = "claude-sonnet-4-6"
|
||
llm_base_url: str = ""
|
||
{% else %}
|
||
llm_api_key: str = ""
|
||
llm_model: str = ""
|
||
llm_base_url: str = ""
|
||
{% endif %}
|
||
generate_timeout_seconds: int = 300
|
||
preview_timeout_seconds: int = 180
|
||
dialog_timeout_seconds: int = 60
|
||
max_concurrent_sessions: int = 4
|
||
process_pool_workers: int = os.cpu_count() or 4
|
||
|
||
model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8")
|
||
```
|
||
{% if llm_provider == 'gemini-cli' %}
|
||
## Skill 安装(skills/installer.py)
|
||
|
||
Agent 启动时自动确保配置指定的 skill 已安装。`skill_name` 与 `skill_repo` 均来自 `Settings`,换 skill 只改 `.env`,不改代码。
|
||
|
||
> 注意:安装目的地路径(`~/.gemini/skills/`)是 Gemini CLI 约定;换 CLI 时 `installer.py` 需同步修改。
|
||
{% endif %}
|
||
## 测试规范
|
||
|
||
{% if llm_provider == 'gemini-cli' -%}
|
||
- PTY 测试:mock `pty.openpty`、`os.read/write`,验证 ANSI 过滤和 HTML 检测
|
||
- 超时测试:注入假 PTY 不输出,验证 `LLMTimeoutError` 被抛出
|
||
- skill 安装测试:mock `git clone`,验证成功/失败分支
|
||
{% else -%}
|
||
- LLM session 测试:mock `LLMSession.generate` 返回 `AsyncIterator[str]`,验证流式处理
|
||
- 超时测试:`generate` mock 为永不完成的迭代器,验证 `LLMTimeoutError` 被抛出
|
||
{% endif -%}
|
||
- DAG 节点测试:mock `send` 回调,验证状态转换和推送消息类型
|
||
- checkpoint 恢复测试:
|
||
- 注入 `checkpoint={status: "done"}` → 验证直接推送结果,不启动 LLM
|
||
- 注入 `checkpoint={status: "screenshotting"}` → 验证跳过 LLM 直接进截图节点
|
||
- 注入 `checkpoint={status: "dialog"}` → 验证推送 `checkpoint_lost`
|