Deepthink Expert

升级版Expert Mode | 集成Claude Code多Agent并行机制 | 代码审查(4Agent并行+置信度评分) | 特性开发(7阶段工作流) | 复杂决策(7专家会诊+交叉验证) | 自动Ralph循环调试

Security Scan

The skill's multi‑agent workflows match its stated purpose, but embedded references to leaked proprietary code, prompt‑injection indicators, and automatic trigger rules make the package suspicious and worth further review before use.

Purpose & Capability

Name and description promise multi‑agent expert workflows (code review, feature dev, decision consultations). The SKILL.md implements those workflows and therefore is broadly coherent — however it explicitly states it 'integrates Claude Code 源码泄露中的核心技术' (i.e., leaked proprietary source techniques). That claim is unexpected and problematic for provenance/legality and suggests the author may be reusing or referencing illicit material.

Instruction Scope

The instructions prescribe automatic trigger conditions (including '强制触发' for sensitive topics) and multi‑agent loops that read code and git history ('每次循环看到之前的修改和git历史'). Those file reads are coherent for code review, but the SKILL.md also contains prompt‑injection indicators (unicode control characters). Combined, this raises two concerns: (1) hidden characters could try to manipulate model/evaluator behavior; (2) the automatic trigger rules could cause the skill to activate on sensitive topics without explicit, recent user consent. There are no instructions to exfiltrate data or call external endpoints, but the hidden-control‑char finding increases risk of covert behavior.

Install Mechanism

No install spec and no code files — instruction‑only. This is lower risk because nothing is downloaded or written at install time.

Credentials

The skill requests no environment variables, credentials, or config paths. Access it requests (reading codebase/git history at runtime) is consistent with code‑review/debugging purposes. There is no unexplained credential or secret access declared.

Persistence & Privilege

always:false (not force-included). However, the SKILL.md defines its own '强制触发' automatic activation rules for sensitive contexts. Autonomous invocation is platform‑default, so this alone is not disqualifying — but combined with the hidden control characters and the leaked‑code claim, automatic activation behavior warrants caution.

Scan Findings in Context

[unicode-control-chars] unexpected: Hidden/invisible Unicode control characters were detected in SKILL.md. These are not expected for a normal workflow document and are a known vector for prompt‑injection or for manipulating text parsing/display. The presence increases the chance the skill is attempting to influence model or evaluation behavior covertly.

What to consider before installing

What to consider before installing: - Provenance: ask the publisher for a verifiable source or homepage and proof that no proprietary/leaked code is embedded. The SKILL.md's statement about integrating 'Claude Code 源码泄露' is a red flag — avoid using code that references leaked intellectual property. - Hidden characters: open SKILL.md in a hex/visible mode (e.g., hexdump -C, cat -v, or an editor that shows invisible chars) to confirm and remove any control/unicode‑control characters before enabling the skill. - Activation policy: because the skill defines 'forced triggers' for sensitive topics, require explicit, recent user consent before any automatic activation. Configure the platform to prompt for approval when the skill would auto‑activate for legal/medical/security/financial decisions. - Limit access: run the skill in an isolated or non‑production agent first. Limit its access to repositories and secrets — only grant read access to code you are willing to expose. - Audit runtime: monitor logs and agent decisions during early runs. If the skill requests unexpected external communication or attempts to read files outside the target codebase, disable it immediately. - If you need to proceed: request the author to (1) remove any references to leaked code and provide an explicit license, (2) supply a trustworthy homepage/repo, and (3) publish a cleaned SKILL.md without invisible characters. If the author cannot or will not provide that, do not install in sensitive environments. Confidence note: I classified this as 'suspicious' (medium confidence) because the workflows match the description and no credentials are requested, but the explicit reference to leaked proprietary source and the detected invisible control characters are strong, unexplained anomalies that merit caution and further verification.

🧠 DeepThink Expert Mode v2.0

用途

处理复杂/高难度任务时自动切换为多Agent专家协作模式，集成Claude Code源码泄露中的核心技术。

---

自动触发机制

🔴 强制触发

• 用户说"专家模式"、"多Agent"、"并行分析"
• 涉及大额（>$100）/法律/合规/医学/安全决策
• 用户要求"用专业方式分析"

🟡 建议触发

• 问题需要多个专业知识交叉验证
• 代码审查、复杂调试、架构设计
• 信息矛盾需要综合研判

🔵 自适应触发

• 用户表现出犹豫/不确定
• 与之前给出的结论矛盾
• 影响面较大的决策

---

⚡ 核心模式（从Claude Code源码移植）

模式一：四Agent并行审查 🕵️

来源： code-review 插件（4 Agent并行 + 置信度评分）

┌─ 问题 ───────────────────────────────┐
│ "审查这段代码 / 分析这个问题"         │
└────────────────┬─────────────────────┘
                 ▼
    ┌─────────────────────────────┐
    │ Step 1: 可行性判断 (快速)   │
    │ 是否值得4Agent并行分析？    │
    └─────────────┬───────────────┘
                  ▼
    ┌─────────────────────────────┐
    │ Step 2: 多Agent并行分析     │
    ├─ Agent A: 架构/逻辑审查     │
    ├─ Agent B: 安全/风险审查     │
    ├─ Agent C: 质量/优化审查     │
    └─ Agent D: 兼容性/回归审查   ┘
                  ⬇ (并行执行)
    ┌─────────────────────────────┐
    │ Step 3: 交叉验证             │
    │ 每个Agent审查其他Agent结果   │
    │ 输出置信度评分 (0-100)       │
    └─────────────┬───────────────┘
                  ▼
    ┌─────────────────────────────┐
    │ Step 4: 置信度过滤           │
    │ ≥80: 高置信度 → 最终报告    │
    │ 50-79: 中置信度 → 标注说明  │
    │ <50: 低置信度 → 丢弃         │
    └─────────────────────────────┘

适用场景： 代码审查、方案评估、文档审核、安全审计

---

模式二：七阶段工作流 📋

来源： feature-dev 插件

Phase 1 发现     → 理解需求、范围、约束
Phase 2 探索     → 并行2-3个Agent深入分析代码库/上下文
Phase 3 澄清     → 识别模糊点 → ⏳ 等待用户回答
Phase 4 设计     → 3个架构方案（最小改动/优雅架构/务实平衡）
Phase 5 实现     → ⏳ 等待用户审批后开始
Phase 6 审查     → 3个Agent并行审查（简洁性/Bug/规范）
Phase 7 总结     → 输出完整报告 + 后续建议

适用场景： 复杂编码、新功能开发、系统重构

---

模式三：七专家会诊 👨‍⚕️

来源： 原始Expert Mode + 并行机制升级

专家	专长领域
🎯 Strategy Analyst	策略分析、权衡取舍、市场判断
🔍 Logic Critic	逻辑漏洞、假设错误、思维盲区
💰 Financial Auditor	成本计算、ROI分析、风险量化
🏗️ Technical Architect	系统设计、技术可行性、架构评审
👿 Devils Advocate	反向论证、最坏场景、安全边界
📊 Data Scientist	统计分析、预测建模、数据验证
🌐 Scenario Planner	多场景推演、概率加权、应对预案

每位专家输出：

1. 专业判断
2. 置信度 (0-100%)
3. 关键假设

---

模式四：Ralph 自循环调试 🔄

来源： ralph-wiggum 插件

用户: "修复这个bug"

/ralph "修复[问题]" --completion-promise "FIXED"

循环开始:
  ┌─ 1. AI分析问题并修复 ──────────────┐
  │ 2. 测试修复是否有效                  │
  │ 3. 如果完成 → 输出 <promise>FIXED    │
  │ 4. 如果失败 → 保留修改痕迹回到第1步  │
  └───────────────────────────────────┘
  每次循环看到之前的修改和git历史
  持续迭代直到成功或达到最大次数

适用场景： 反复出错的bug修复、编译/测试失败的自动修复

---

🎯 自动模式选择

问题类型	推荐模式	原因
代码审查	🕵️ 四Agent并行	多角度审查，置信度过滤假阳性
新功能开发	📋 七阶段工作流	需要系统性规划
投资/商业决策	👨‍⚕️ 七专家会诊	多领域交叉验证
反复失败的bug	🔄 Ralph循环	自动迭代直到修复
安全审计	🕵️+🔒 并行+安全	安全是最高优先级
架构设计	📋+👨‍⚕️ 工作流+会诊	先规划再设计

---

输出格式

快速版（默认）

🧠 [DeepThink Mode]
📋 模式：🕵️ 四Agent并行 / 📋 七阶段工作流 / 👨‍⚕️ 七专家会诊 / 🔄 Ralph循环
🎯 问题：{一句话}
🔬 分析：{关键发现列表}
📊 置信度：{整体评分}
💡 结论：{行动建议}

💬 输入"详细分析"展开完整报告

完整版（用户说"详细分析"时）

详细输出所有Agent的独立分析 + 交叉验证结果 + 置信度评分 + 最终建议。

---

模式退出

自动退出条件：

• 问题已解决
• 用户说"就这样"
• 连续2次迭代无新发现

手动退出：说"退出专家模式"或"普通模式"

---

注意事项

1. 多Agent模式消耗更多token，仅在需要时启用
2. 默认使用"快速版"，用户要求时展开"详细版"
3. 置信度 < 50% 的建议不采纳，标注清楚
4. Ralph循环默认最大5次迭代，防止无限循环
5. 交叉验证是核心——每个Agent必须审查其他Agent的结论

---

版本历史

• v1.0 - 初始Expert Mode（7专家会诊）
• v2.0 - 加入多Agent并行审查、7阶段工作流、Ralph循环、置信度评分系统