Skill Collision Engine

Skill碰撞引擎 —— 让Skill之间自动四向碰撞生成新Skill的元引擎，Skill工厂的自进化核心

Security Scan

The skill's description and high-level design are plausible for a 'meta' skill, but the runtime instructions claim Python APIs and ongoing behaviors without shipping any code or declaring needed credentials — several mismatches warrant caution.

Purpose & Capability

The stated purpose (a meta-engine that collides Skills and self-evolves) is coherent with the described operations (reading other Skills, external knowledge sources, producing new Skills). However the SKILL.md includes a Python API (SkillCollisionEngine class and methods) and install commands while the package contains no code files — this is an inconsistency between claimed capability and delivered artifacts.

Instruction Scope

Instructions direct the agent to read verification logs across 'all published Skill' (VERIFICATION_LOG.md), pull from multiple external sources (ArXiv, GitHub, YouTube, web/PDF, human dialog), and run periodic 'heartbeat' tasks. Those actions imply broad filesystem and network access and potential aggregation of sensitive or private Skill data; yet the skill does not declare limits or required permissions. The SKILL.md also gives executable examples (Python import) that do not exist in the package.

Install Mechanism

This is an instruction-only skill with no install spec and no bundled binaries — lowest install risk. The README suggests 'clawhub install' or copying files, but nothing will write or execute shipped code because no code files are present.

Credentials

No environment variables or credentials are declared even though the SKILL.md describes accessing external services (GitHub, YouTube transcripts) and reading many local verification logs. Either the skill expects to run with existing agent permissions (which gives it broad access) or it omits declaring required API keys — both are notable omissions that reduce transparency.

Persistence & Privilege

The skill is not set to always:true and uses normal autonomous invocation defaults. However the documentation prescribes recurring 'heartbeat' and daily auto-collision tasks and self_evolve routines that would run periodically and read/aggregate data; without shipped code or explicit scheduler hooks it's unclear how/when these run. If the agent executes these autonomously with broad file/network access, the blast radius increases.

Scan Findings in Context

[no_regex_findings] expected: The package is instruction-only and contains no code files, so the regex scanner had nothing to analyze. That absence is expected for a purely-doc skill, but it also means there is no executable implementation to inspect against the SKILL.md claims (e.g., Python class).

What to consider before installing

Do not install blindly. Key points to consider before proceeding: - The skill ships only documentation (no code). The SKILL.md advertises a Python API and install commands but no Python module is present — ask the author for the actual implementation or a link to the exact code that will run. - The runtime instructions ask the agent to read verification logs across all Skills and to fetch from many external sources; verify what files and external endpoints it will access and whether those contain sensitive data. - The skill does not declare any API keys or credentials even though some sources (YouTube, GitHub with high rate) may require them; demand explicit configuration guidance and least-privilege requirements. - If you consider using it, run it in a restricted sandbox first (no access to private Skills, secrets, or system directories), require explicit consent for scanning other Skills, and audit any produced outputs and network calls. - If you cannot obtain the referenced implementation (the SkillCollisionEngine module) or a clear security/configuration model from the author, treat this skill as untrusted and avoid granting it broad file or network access.

元数据

字段	值
名称	skill-collision-engine
版本	1.0.0
作者	KingOfZhao
发布日期	2026-03-31
置信度	96%

核心哲学

认知世界的本质是无穷层级的框架节点。 Skill碰撞引擎让节点之间自动碰撞，产生新的认知节点。 这不是简单的Skill组合，是认知层面的结构化涌现。

Skill A ⊗ Skill B → 四向认知碰撞 → 新 Skill C

self-evolution ⊗ diepre-vision → vision-action-evolution-loop ✓ (已验证)
self-evolution ⊗ arxiv-collision → [待碰撞]
diepre-vision ⊗ diepre-embodied → [待碰撞]
任意 Skill ⊗ 任意外部知识 → [待碰撞]

三大核心能力

1. Skill-Skill 碰撞（认知涌现，非机械组合）

不是简单拼接两个SKILL.md，而是对两个Skill的认知框架做四向碰撞：

输入:
  Skill A 的已知集合 FA + 未知集合 VA + 推理框架 RA
  Skill B 的已知集合 FB + 未知集合 VB + 推理框架 RB

碰撞过程:
  正面碰撞: FA ∩ FB 的重叠区域 → 直接可迁移的认知
  反面碰撞: FA 中的已知 是否能解决 VB 中的未知？
  侧面碰撞: RA 和 RB 的非主要特性碰撞 → 隐藏新认知
  整体碰撞: A+B 的研究方向趋势 → 新的框架方向

输出:
  新 Skill C 的已知集合 FC = (FA ∩ FB) ∪ 碰撞新增认知
  新 Skill C 的未知集合 VC = VA ∪ VB - 碰撞已解决
  新 Skill C 的推理框架 RC = RA ⊗ RB 的融合框架
  置信度: 每个认知点独立标注

为什么不是机械组合？

• 机械组合 = A的SKILL.md + B的SKILL.md 拼接 → 没有新认知
• 认知碰撞 = A和B的已知/未知交集 → 产生A和B都不单独拥有的涌现认知
• 关键：涌现认知出现在 FA∩VB 和 FB∩VA 的交叉区域

2. SOUL 领域适配（通用框架 → 专用Skill）

SOUL哲学是通用认知框架，任何人/任何职业都可以直接使用：

通用 SOUL 框架:
  1. 已知 vs 未知    → 每个职业有不同的"已知"和"未知"
  2. 四向碰撞推理    → 推理方法不变，碰撞对象变
  3. 人机闭环        → 每个职业有不同的实践验证方式
  4. 文件即记忆      → 每个职业有不同的记忆类型
  5. 置信度+红线      → 每个职业有不同的红线

适配流程:
  输入: 职业名称 + 该职业的核心知识领域
  输出: 领域专用 Skill（SOUL框架 + 领域已知/未知 + 领域四向碰撞实例）

示例适配:
  程序员 → programmer-cognition（代码审查四向碰撞、bug模式已知/未知）
  医生   → physician-cognition（症状推理四向碰撞、诊断已知/未知）
  教师   → educator-cognition（学情分析四向碰撞、教学效果已知/未知）
  企业家 → entrepreneur-cognition（市场判断四向碰撞、商业模式已知/未知）
  设计师 → designer-cognition（用户研究四向碰撞、设计决策已知/未知）
  科研人员 → researcher-cognition（文献碰撞四向碰撞、假设已知/未知）

3. 工厂元进化（自改进碰撞质量）

引擎通过验证反馈持续优化自身：

进化循环:
  1. 读取所有已发布 Skill 的 VERIFICATION_LOG.md
  2. 分析: 哪些碰撞产生的Skill验证置信度高？哪些低？
  3. 提取: 高置信度碰撞的共同特征（输入类型、碰撞方向权重、领域特征）
  4. 更新碰撞参数:
     - direction_weights = {正面: 0.3, 反面: 0.25, 侧面: 0.2, 整体: 0.25}
     - domain_affinity_map = {包装: [diepre, vision], AI: [evolution, arxiv], ...}
     - threshold_adjustments = 基于历史置信度的动态阈值
  5. 下次碰撞使用优化后的参数

支持的外部知识源

1. ArXiv 论文        → arxiv-collision-cognition（已有）
2. GitHub 仓库/README → 代码能力碰撞
3. Wikipedia 条目     → 基础概念碰撞
4. 网页/PDF           → 通用知识碰撞
5. YouTube/播客文字稿  → 口述知识碰撞
6. 人类对话记录        → 实践知识碰撞（最高价值）

安装命令

clawhub install skill-collision-engine
# 或手动安装
cp -r skills/skill-collision-engine ~/.openclaw/skills/

调用方式

from skills.skill_collision_engine import SkillCollisionEngine

engine = SkillCollisionEngine(workspace=".", skills_dir="./skills")

# 1. Skill-Skill 碰撞
result = engine.collide_skills(
    skill_a="self-evolution-cognition",
    skill_b="diepre-vision-cognition",
    context="包装行业具身智能"
)
print(result.new_skill_name)      # "vision-action-evolution-loop"
print(result.emergent_insights)   # 涌现认知列表
print(result.confidence)          # 0.96

# 2. SOUL 领域适配
profession_skill = engine.adapt_soul(
    profession="程序员",
    domain_knowledge=["代码审查", "bug模式识别", "架构设计"],
    domain_unknown=["生产环境行为", "并发安全"],
    practice_methods=["Code Review", "单元测试", "生产监控"]
)
# 输出: programmer-cognition Skill

# 3. 外部知识碰撞
result = engine.collide_with_external(
    skill="diepre-vision-cognition",
    source_type="arxiv",
    source_id="2410.05340",
    context="VLM生成CAD的技术可行性"
)

# 4. 工厂元进化
engine.self_evolve()  # 读取验证日志，优化碰撞参数

当前 Skill 碰撞矩阵（已发布6个 Skill）

                  self-evo  diepre-vis  human-ai  arxiv  vis-act  embodied
self-evolution       -       ✓(已碰撞)    ×        ×      ×         ×
diepre-vision        ✓          -         ×        ×      ×         ×
human-ai-closed      ×          ×         -        ×      ×         ×
arxiv-collision      ×          ×         ×        -      ×         ×
vis-act-loop         ×          ×         ×        ×      -         ×
diepre-embodied      ×          ×         ×        ×      ×         -

✓ = 已碰撞产生新Skill
× = 待碰撞（潜在新Skill方向）

学术参考文献

1. A Survey of Self-Evolving Agents — 自进化Agent综述，元进化的理论基础
2. SAGE: Multi-Agent Self-Evolution — 多Agent碰撞产生新认知
3. Group-Evolving Agents — 群体进化+经验共享（←Skill碰撞的经验来源）
4. Self-evolving Embodied AI — 元进化+自改进循环
5. Beyond RAG for Agent Memory — 记忆聚合（←碰撞结果的知识存储）
6. Memory in the Age of AI Agents — 长时序记忆（←进化历史持久化）