thesoze/feature-crew
GitHub: thesoze/feature-crew
一个基于分层LLM的自主代码生成、测试和审查框架，旨在大幅降低自动化开发成本并提升效率。
Stars: 0 | Forks: 0
# Feature Crew — 自主开发流程
一个经过生产验证的、可复用的自主代码生成、测试和审查框架。最初为 Argus 项目构建，现作为独立模板提取，适用于任何 Python 项目。
**成本：** 每项任务约 $0.0005（本地 LLM + 低成本 API 审查），而传统方法成本为 $0.15 以上。
**速度：** 每项任务 30-60 秒（代码生成 + 自审计 + 审查）。
**可靠性：** 仅使用新鲜上下文审查（无会话上下文膨胀）。
## 快速开始
```
# stral_tier1.py
git clone https://github.com/thesoze/feature-crew.git myproject-crew
cd myproject-crew
# 21. guardrails/mistral-tier1.md
cp config.example.toml config.toml
# 22. feature_crew/audit/plugins/bandit.py
# 23. feature_crew/review/checks/security.py
docker-compose up -d # Redis + optional Spark service
# 24. feature_crew/review.py
python -m feature_crew tier1 orchestrator
# 25. Terminal 1: Redis
python -m feature_crew enqueue \
--spec "Add a function named greet() that returns 'Hello'" \
--spec-slug "test-greet-function"
```
## 架构
```
Task Queue (Redis)
↓
Spark Service (tier-specific task dispatch)
├─ Tier 1 (XS, low-risk): Local LLM (Qwen, Llama, etc.)
├─ Tier 2 (S/M, standard): Claude Sonnet
└─ Tier 3 (L/XL, high-risk): Claude Opus + security review
↓
Code Generation (via Aider or direct LLM)
↓
Self-Audit (pytest, ruff, mypy, compile)
↓
Fresh-Context Review (read diff + audit only, never conversation history)
↓
Merge & Deploy (or re-queue with feedback, up to 3 attempts)
```
## 核心组件
### 1. 任务路由器
- **文件：** `feature_crew/router.py`
- **分类：** 按代码行数、复杂性、风险分为 XS/S/M/L/XL
- **输出：** 包含 `tier`、`model`、`crew_plan` 的任务负载
### 2. Spark 服务
- **文件：** `spark/tier1_service.py`
- **职责：** 管理任务队列、创建工作树、跟踪状态
- **协议：** REST API (GET /next, POST /complete, POST /feedback)
- **存储：** Redis 队列 + 任务元数据
### 3. 编排器
- **文件：** `feature_crew/orchestrator.py`
- **层级：** 每个层级可插拔的编排器
- `Tier1Orchestrator` — 本地 LLM 分派（Qwen、Llama）
- `Tier2Orchestrator` — Claude Sonnet（通过 API）
- `Tier3Orchestrator` — Claude Opus + 安全审查
- **方法：** 轮询 Spark 服务、分派、等待审计、审查、合并
### 4. 自审计
- **文件：** `feature_crew/audit/`
- **工具：** pytest、ruff、mypy、py_compile（可配置）
- **判定：** `ready_for_review` | `needs_retry`
- **输出：** 包含所有检查结果的 JSON 报告
### 5. 新鲜上下文审查
- **文件：** `feature_crew/review.py`
- **仅读取：** 规格说明 + 差异 + 审计报告
- **不读取：** 会话历史、任务元数据、先前尝试
- **检查：** 自定义验证链（每层级 5+ 项检查）
- **输出：** `{action: approve|reject, reason, feedback}`
### 6. 命令行界面
- **命令：**
- `enqueue` — 将任务添加到队列
- `tier1 orchestrator` — 启动第一层级轮询循环
- `tier2 orchestrator` — 启动第二层级轮询循环
- `status` — 检查队列状态
- `config` — 显示/测试配置
## 配置
### 26. Terminal 2: Spark service
```
[project]
name = "myproject"
repo_path = "/path/to/myproject"
description = "My awesome project"
[infrastructure]
redis_url = "redis://localhost:6379"
spark_url = "http://localhost:30001"
[tier1]
enabled = true
model = "qwen3-235b" # or "llama3-70b", "mistral-7b", etc.
guardrail_file = "guardrails/tier1-f2.md" # optional: teach-to-ask guardrail
timeout_seconds = 300
[tier2]
enabled = true
model = "claude-sonnet-4-6" # Claude API
api_key = "sk-..." # or read from .env
cost_budget = 5.0 # USD per day
[tier3]
enabled = false # tier-3 for high-risk only
model = "claude-opus-4-7"
security_review = true
[self_audit]
tools = ["pytest", "ruff", "mypy", "compile"]
pytest_args = ["tests/", "-v", "--tb=short"]
ruff_args = ["check", "src/", "--select=E9,F63,F7,F82"]
[review]
# 27. Terminal 3: Orchestrator
checks = [
"audit_verdict",
"diff_size",
"file_count",
"spec_adherence",
"suspicious_patterns",
]
max_files_per_tier1 = 3
max_files_per_tier2 = 10
```
### 28. Using docker-compose
可选的防护规则，用于教导第一层级 LLM 在规格说明模糊时提问：
```
# 29. Using systemd (Linux)
When given a vague spec, respond with:
[ASK]:
Vague spec indicators:
- "improve", "clean up", "refactor" (without specifics)
- "make it better", "optimize", "simplify"
- "handle edge cases" (which ones?)
Example:
Spec: "Improve the login flow"
Response: [ASK]: Do you want to (a) add rate limiting, (b) fix a specific bug, (c) add 2FA?
```
## 使用模式
### 模式 1：快速第一层级（开发）
```
# 30. Using launchd (macOS)
python -m feature_crew enqueue \
--spec "Add validation to email field in UserForm" \
--spec-slug "validate-email-form"
# 31. Spark service health
# 32. Queue status
```
### 模式 2：第二层级（标准功能）
```
# 33. Logs
python -m feature_crew enqueue \
--spec "Add OAuth2 login endpoint with database migration" \
--spec-slug "oauth2-login" \
--tier 2
# 34. Task success rate
# 35. Cost tracking
```
### 模式 3：第三层级（高风险）
```
# 36. Latency per tier
python -m feature_crew enqueue \
--spec "Implement HMAC verification for webhook signatures" \
--spec-slug "webhook-hmac" \
--tier 3
# Yes, that's 36 lines. Now, I need to translate each one to Simplified Chinese, keeping professional terms in English.
# Let me translate each line:
```
## 可扩展性
### 添加新的 LLM 模型
1. **创建层级编排器：**
# 1. Clone this repo – Translation: 克隆此仓库
class MistralTier1Orchestrator(BaseTier1Orchestrator):
async def dispatch_to_llm(self, spec, worktree_path, feedback):
# 调用 Mistral API 或本地 vLLM 端点
pass
2. **在配置中注册：**
[tier1]
model = "mistral-7b"
api_endpoint = "http://localhost:8000"
3. **可选防护规则：**
# 2. Configure for your project – Translation: 为您的项目配置
# 针对 Mistral 特性的自定义防护规则
### 添加自审计工具
1. **创建审计插件：**
# 3. Edit config.toml: set LLM models, Redis URL, project paths – Translation: 编辑 config.toml：设置 LLM 模型、Redis URL、项目路径
def run_bandit(task_id, worktree_path, files_modified):
result = subprocess.run(["bandit", "-r", "src/"])
return {"exit_code": result.returncode, "passed": result.returncode == 0}
2. **在配置中注册：**
[self_audit]
tools = ["pytest", "ruff", "mypy", "bandit"]
### 添加审查检查
1. **创建自定义验证器：**
# - Here, config.toml, LLM, Redis are kept in English as per instruction.
def check_no_hardcoded_secrets(spec, diff, audit_report):
if "password" in diff or "secret" in diff:
return False, "Found hardcoded secrets in diff"
return True, None
2. **注册：**
# 4. Start the infrastructure – Translation: 启动基础设施
CHECKS = [
check_audit_verdict,
check_diff_size,
check_no_hardcoded_secrets, # 自定义
]
## 部署
### 本地开发
```
# 5. Run the pipeline – Translation: 运行流水线
docker run -d -p 6379:6379 redis:latest
# 6. Enqueue a task – Translation: 将任务加入队列
python spark/tier1_service.py
# 7. config.toml – This is a file name, so keep it as is: config.toml
python -m feature_crew tier1 orchestrator
```
### 生产环境
```
# 8. Validation checks (ordered, stop on first failure) – Translation: 验证检查（有序，第一个失败时停止）
docker-compose -f docker-compose.prod.yml up -d
# 9. guardrails/tier1-f2.md – File name, keep as is: guardrails/tier1-f2.md
sudo systemctl start feature-crew-orchestrator.service
# 10. Tier-1 F2 Guardrail – Translation: Tier-1 F2 防护栏 (but keep Tier-1 and F2 in English? F2 might be a specific term. From context, it seems like a tier or type. I'll keep "Tier-1" and "F2" in English, and translate "Guardrail" to "防护栏".)
launchctl load ~/Library/LaunchAgents/com.feature-crew.plist
```
## 监控
### 健康检查
```
# - Let's see the instruction: keep professional terms in English. "Tier-1" and "F2" might be part of a naming convention, so I'll keep them as is. Translation: Tier-1 F2 防护栏
curl http://localhost:30001/health
# 11. Single command for dev/test features – Translation: 用于开发/测试功能的单个命令
python -m feature_crew status
# 12. Orchestrator auto-routes to tier-1 (Qwen3), audits locally, reviews cheaply – Translation: Orchestrator 自动路由到 tier-1 (Qwen3)，本地审计，低成本审查
tail -f logs/orchestrator.log
tail -f logs/spark.log
```
### 指标
```
# - Orchestrator is a term, so keep it in English? The instruction says keep tool/framework names. Orchestrator might be a component name, so I'll keep it as "Orchestrator". Similarly, Qwen3 is a model name, keep in English.
redis-cli keys "task:*:status" | wc -l
# 13. Cost: ~$0.0005, Time: ~40s – Translation: 成本：约 $0.0005，时间：约 40 秒
python -m feature_crew metrics --duration 7d
# - Keep the currency symbol and units as is? Probably translate "Cost" and "Time" to Chinese, but keep the numbers and symbols. Since it's a heading, I'll translate the words.
python -m feature_crew metrics --tier 1 --metric latency
```
## 故障排除
| 问题 | 原因 | 修复方法 |
|-------|-------|----------|
| 任务在"分派"状态挂起 | LLM 端点无法访问 | 检查配置中的 vLLM/API 端点 URL |
| 自审计静默失败 | 工具未安装 | `pip install pytest ruff mypy` |
| 审查拒绝所有差异 | 检查规则过于严格 | 在 config + review.py 中调整阈值 |
| 合并冲突 | 主分支有新提交 | 实现智能合并（变基 + 重试） |
## 项目结构
```
feature-crew/
├── feature_crew/
│ ├── __init__.py
│ ├── cli.py # CLI commands
│ ├── router.py # Task triage (tier classification)
│ ├── orchestrators/ # Pluggable per-tier orchestrators
│ │ ├── base.py
│ │ ├── tier1_qwen.py
│ │ ├── tier1_mistral.py
│ │ ├── tier2_sonnet.py
│ │ └── tier3_opus.py
│ ├── audit/ # Self-audit framework
│ │ ├── runner.py
│ │ └── plugins/
│ │ ├── pytest.py
│ │ ├── ruff.py
│ │ ├── mypy.py
│ │ └── compile.py
│ ├── review/ # Fresh-context review
│ │ ├── checks.py
│ │ └── validators/
│ ├── models/ # Data classes (Task, AuditReport, etc.)
│ └── config.py # Configuration loader
├── spark/ # Spark service (tier-specific dispatch)
│ ├── tier1_service.py
│ ├── requirements.txt
│ └── README.md
├── guardrails/ # LLM guardrails (optional)
│ ├── tier1-f2.md
│ └── tier2-instructions.md
├── scripts/
│ ├── enqueue.py
│ ├── status.py
│ └── metrics.py
├── tests/
│ ├── test_router.py
│ ├── test_orchestrator.py
│ ├── test_audit.py
│ └── test_review.py
├── docker-compose.yml # Local dev
├── docker-compose.prod.yml # Production
├── config.example.toml # Template config
├── pyproject.toml # Python dependencies
└── README.md # This file
```
## 参与贡献
Feature Crew 设计为可扩展，以适应您的特定项目：
1. **分叉或克隆**此仓库
2. **自定义** config.toml、guardrails/ 和审查检查
3. 在采用第二/三层级之前，先用几个第一层级任务**测试**
4. **监控**成功率、成本、延迟
5. **迭代** — 收集失败案例，完善防护规则，调整检查项
## 许可证
MIT 许可证 — 可在您的项目中自由使用。
## 致谢
在 **Argus** 项目（一个拥有 35+ 个 OSINT 源的常驻个人 AI 代理）中构建并经过实战检验。现作为可复用模板提取，以便将相同的自主开发工作流带入任何 Python 项目。
## 支持
- **问题：** 在 GitHub 上提交 issue
- **文档：** 阅读 `docs/` 文件夹以深入了解架构
- **示例：** 查看 `examples/` 获取项目特定配置
标签：AI代码审查, AI辅助开发, Docker容器化, LLM模型, LLM集成, Python项目, Redis队列, REST API, 云API, 代码审查, 代码测试, 代码生成框架, 任务调度, 任务队列管理, 低成本开发, 分层处理, 安全审查, 安全规则引擎, 开发效率, 开发框架, 开源框架, 微服务架构, 快速开发, 成本优化, 技术栈优化, 持续集成, 搜索引擎查询, 智能开发, 本地LLM, 网络可观测性, 自动化代码生成, 请求拦截, 软件开发管道, 逆向工具