File size: 8,498 Bytes
7bedc79
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
---
feature: agentlens
status: delivered
updated: 2026-11-19
source: agentlens
---

# AgentLens — AI 编程 Agent 配置迁移 + 产出体检

## Report

**What was built** — AgentLens v0.1.0: a CLI + MCP server for (1) cross-agent config migration (Claude/Cursor/Codex/Windsurf/Copilot), (2) config drift detection, and (3) adversarial code QA. All rules parsed into a unified IR and re-generated per target agent. Pure local, no server dependency.

**Verification** — `npm test`: 46/46 passing across 6 test files. `npx tsc --noEmit`: clean. `npm run build`: dist/ generated. E2E verified: scan (5 agents detected), move (claude→windsurf/cursor), drift (11 issues across 37 rules), check (3 critical vulnerabilities detected), install (auto-detected 3 tools on this machine), mcp (stdio server runs). VS Code extension bundled (32KB) and packaged as agentlens-0.1.0.vsix (30KB).

## [S1] Problem

开发者同时使用多个 AI 编程 Agent(Claude Code、Cursor、Codex、Windsurf 等),每个 Agent 有自己的配置格式(CLAUDE.md、.cursor/rules/、AGENTS.md 等)。两个真实痛点:

1. **配置迁移难**:用户在一个 Agent 里积累的规则、偏好、工作流习惯,换到另一个 Agent 时需要手动重写。规则散落在 N 个位置 × M 个仓库,"知识 90% 相同但维护 N 份"。现有同步工具只管复制文件,不管规则是否已经失效("一模一样的烂规则副本还是烂规则")。
2. **产出不信任**:46% 开发者不信任 AI 输出,66% 最大挫败是"几乎正确但不完全对"。开发者自己写 10+ 个 hook 做对抗式审查,但没有现成商业产品。官方文档承认插件评测是空白。

## [S2] Design

### 产品定位

一个 CLI + MCP Server,服务两类场景:
- **迁移(Move)**:解析、归一化、转换 AI Agent 的规则配置,一键在 Agent 之间搬运
- **体检(Check)**:对 Agent 产出的代码/改动做对抗式审查,输出风险报告

### [S2.1] 配置迁移引擎(核心差异化)

**输入**:源 Agent 类型 + 路径(自动检测当前项目里有哪些 Agent 配置)
**输出**:目标 Agent 格式的配置文件

支持的 Agent 配置格式(MVP 阶段):
| Agent | 配置文件 | 解析 |
|---|---|---|
| Claude Code | CLAUDE.md, .claude/rules/*.md | ✅ |
| Cursor | .cursor/rules/*.mdc | ✅ |
| Codex | AGENTS.md | ✅ |
| Windsurf | .windsurfrules | ✅ |
| 通用 | .editorconfig, .github/copilot-instructions.md | ✅ |

**归一化模型**:所有规则解析为统一中间表示(IR):
```
Rule {
  id: string
  category: "style" | "workflow" | "constraint" | "context" | "tooling"
  content: string          # 规则正文
  scope: "global" | "project" | "path-glob"
  source: { agent, file, line }
  status: "active" | "suspect" | "stale"
}
```

**漂移检测**(在迁移之外的附加价值):
- 同一规则在多处出现但内容不一致 → 标记 drift
- 规则引用的工具/路径/文件不存在 → 标记 stale
- 输出漂移报告,建议保留哪个版本

### [S2.2] Agent 产出体检引擎

**输入**:代码片段 / git diff / 文件路径
**输出**:结构化体检报告

检查维度(MVP 优先做前三个):
1. **正确性**:边界条件、空值处理、类型错误、逻辑漏洞
2. **安全**:注入、越权、敏感信息泄露、不安全的依赖
3. **规范符合**:是否违反用户配置中的规则(与 S2.1 联动)
4. 性能隐患(后续)
5. 测试覆盖建议(后续)

**运行方式**:
- CLI: `agentlens check <diff|file>`
- MCP Server: 注册为 MCP 工具,Claude Code / Cursor 等可直接调用
- Hook: 接入 Agent 的 post-write hook(如 Claude Code hooks)

### [S2.3] 技术架构

**跨 Agent 通用策略**(用户核心需求):不为每个 Agent 单独写插件,而是通过 **MCP Server + CLI** 双形态覆盖所有主流工具:

| 目标工具 | 接入方式 | 安装后效果 |
|---|---|---|
| Claude Code | MCP Server | 工具列表出现 agentlens 命令 |
| Cursor | MCP Server | .cursor/mcp.json 注册后可用 |
| Codex | MCP Server | MCP 配置注册后可用 |
| MiMo Desktop | MCP Server | MCP 设置里添加后可用 |
| DSH / WorkBuddy / ZCode | CLI + MCP | 支持 MCP 的走 MCP,否则走 CLI |
| 任何终端 | CLI | `npx agentlens` 直接用 |

**一键安装**:`npx agentlens install --target <agent>` 自动检测本机已装的 Agent,把 MCP 配置写入对应配置文件(.mcp.json / CLAUDE.md / .cursor/settings 等),装完后各 Agent 的工具面板/设置里即出现 AgentLens 命令(scan / move / check / drift)。

```
agentlens/
├── src/
│   ├── core/           # IR 模型、解析器注册表
│   ├── parsers/        # 各 Agent 配置解析器(claude.ts, cursor.ts, codex.ts, windsurf.ts...)
│   ├── migrator/       # IR → 目标格式生成器
│   ├── drift/          # 漂移检测引擎
│   ├── checker/        # 产出体检引擎
│   ├── mcp/            # MCP Server 入口
│   ├── install/        # 一键安装到各 Agent 的配置写入器
│   └── cli/            # CLI 入口
├── tests/
└── docs/
```

- **语言**:TypeScript(Node.js),与 MCP SDK(@modelcontextprotocol/typescript-sdk)天然亲和
- **分发**:npm 包(`npx agentlens`),MCP Registry 注册
- **无服务器依赖**:纯本地运行,代码不出机器(隐私卖点)

### [S2.4] 交互流程

**迁移流程**:
```
$ agentlens scan
  发现 3 个 Agent 配置:
  ✅ CLAUDE.md (47 条规则)
  ✅ .cursor/rules/ (12 条规则)  
  ✅ AGENTS.md (8 条规则)
  ⚠️ 检测到 3 处漂移、2 条疑似失效规则

$ agentlens move --from claude --to cursor
  转换 47 条规则 → .cursor/rules/agentlens.mdc
  ✅ 完成。预览?(y/n)

$ agentlens drift
  漂移报告:... 
```

**体检流程**:
```
$ agentlens check HEAD~1..HEAD
  检查 12 个文件变更...
  🔴 2 个高危 | 🟡 5 个警告 | ⚪ 8 个通过
  [详细报告]
```

## [S3] Out of Scope

- 不做云端同步/账号系统(纯本地工具)
- 不做 Agent 本身的替代或封装(不做 "又一个 AI 编程助手")
- 不做实时监听/自动修复(MVP 只做检测+报告)
- 不做非编程 Agent(如通用聊天机器人)的配置
- 不做付费/许可证系统(MVP 开源,后续再加商业版)

## Tasks

- [x] T1: 项目脚手架 — TypeScript + vitest + MCP SDK 初始化,CLI 骨架可运行 — acceptance: `npx agentlens --help` 输出帮助信息
- [x] T2: 配置解析器 — 解析 CLAUDE.md / .cursor/rules / AGENTS.md 为统一 IR — acceptance: 对 3 种格式的测试样本,解析出完整规则列表且字段正确 (covers: S2.1)
- [x] T3: 迁移生成器 — IR → 目标 Agent 格式输出 — acceptance: claude→cursor / cursor→claude / codex→claude 三组转换测试通过 (covers: S2.1)
- [x] T4: 漂移检测 — 检测同规则不一致、失效引用 — acceptance: 构造漂移样本能正确报告 drift/stale 项 (covers: S2.1)
- [x] T5: 体检引擎 MVP — 对 diff 做正确性+安全+规范检查 — acceptance: 对含已知漏洞的测试代码,报告能标出注入/越权/空指针 (covers: S2.2)
- [x] T6: MCP Server — 注册 scan/move/check/install 为 MCP 工具 — acceptance: Claude Desktop / Cursor 可通过 MCP 调用 agentlens 工具 (covers: S2.2, S2.3)
- [x] T7: 一键安装器 — `agentlens install` 自动检测本机 Agent 并写入 MCP 配置 — acceptance: 在装有 Claude Code 的机器上运行 install 后,Claude Code 工具列表出现 agentlens 命令 (covers: S2.3)
- [x] T8: CLI 集成 — scan/move/check/drift 子命令完整可用 — acceptance: 端到端跑通扫描→迁移→体检全流程 (covers: S2.4)
- [x] T9: 测试与文档 — 单测覆盖 + README + npm 发布准备 — acceptance: `npm test` 全绿,README 含安装和使用示例
- [x] T10: App 集成(用户新增需求)— VS Code/Cursor 扩展(.vsix,设置面板可见)+ 自动检测安装器 — acceptance: .vsix 可安装且命令出现在命令面板/设置;`agentlens install` 自动检测并写入各工具 MCP 配置 (covers: S2.3)

## Journey log

- 2026-11-19: 调研阶段发现"Agent 产出 QA"和"配置漂移检测"是 Agent 工具生态中竞争最少的两个缺口(详见 research/market-gap-opportunities/)。用户决定将两者合并为一个产品。