Self-Evolution Cognition

核心自进化认知框架 —— 基于 SOUL 五律的 Agent 自我进化操作系统

Security Scan

The skill's documentation describes a Python API that writes persistent files and runs verification commands, but the published bundle contains only markdown docs and no executable code—this mismatch is suspicious and you should not run the suggested commands or install it without further verification.

Purpose & Capability

The SKILL.md and README claim a runnable Python class (SelfEvolutionCognition) that persists memory files, performs redline checks, and exposes methods like evolve() and inject_human_feedback(). However, the package contains only markdown files (no .py or executable code). That is inconsistent: either the implementation is missing from the bundle or the documentation is misleading.

Instruction Scope

The instructions and verification protocol instruct running shell/python commands (e.g., grep on skills/*.py, python import of SelfEvolutionCognition, asserting presence of VERIFICATION_LOG.md) and expect the skill to write/read files like VERIFICATION_LOG.md and known_unknown_cache.json. Because there is no code included, these runtime instructions are not backed by implementation and could cause an agent or user to execute commands that do things on disk without an actual contained implementation to audit.

Install Mechanism

No install spec or code is included—this lowers supply-chain risk in that nothing will be automatically downloaded/installed by the skill itself. However, the README suggests using 'clawhub install' or copying a skills directory; those steps assume a codebase exists elsewhere (e.g., GitHub). The absence of an install artifact means the documentation is incomplete or the implementation resides off-bundle and must be fetched—this fetch step is not specified and would be higher-risk if performed automatically.

Credentials

The skill declares no required environment variables, binaries, or credentials. That is proportionate to the documentation. Nevertheless, the skill expects to read/write files in the user's workspace and run local verification commands, so filesystem access is required in practice even if not declared as env/config requirements.

Persistence & Privilege

The design explicitly requires persistent file writes (VERIFICATION_LOG.md, known_unknown_cache.json, heartbeat logs). That persistence is coherent with the 'file as memory' claim but is a material security/privacy consideration: persistent files can contain sensitive intermediate reasoning and may be read by other processes. The skill does not request elevated platform privileges (always:false), but it does assume the ability to write into the workspace.

What to consider before installing

This bundle contains only documentation describing a Python skill that will write persistent files and run verification shell/python commands, but it does not include the implementation code. Do not run the suggested grep/python commands or copy/install code from unknown sources until you verify the real implementation repository. Before installing or running anything: 1) Inspect the GitHub repo linked in the homepage (confirm it actually contains the SelfEvolutionCognition .py files and review them). 2) If you must test, do so in an isolated sandbox/container and monitor file writes and network activity. 3) Audit any code that writes VERIFICATION_LOG.md, known_unknown_cache.json, or similar for sensitive data leakage and for safe handling of shell commands (ensure no use of unsanitized shell.exec). 4) Prefer a release with actual code included in the skill bundle or a clearly specified, trustworthy install source (GitHub releases) rather than only docs. If you want, provide the linked repository contents and I can re-evaluate the implementation for coherence and risk.

Self-Evolution Cognition Skill

元数据

字段	值
名称	self-evolution-cognition
版本	1.0.0
作者	KingOfZhao
发布日期	2026-03-31
置信度	97%

学术参考文献

本框架的理论基础受以下前沿研究启发：

1. A Survey of Self-Evolving Agents — 自进化Agent综述，涵盖What/When/How三维度
2. SAGE: Multi-Agent Self-Evolution for LLM Reasoning — Challenger+Planner+Solver+Critic四Agent闭环（←四向碰撞的理论对应）
3. Group-Evolving Agents — 群体进化+经验共享（←人机闭环的扩展方向）
4. Self-evolving Embodied AI — 具身智能自进化（←向长期方向的演进路径）
5. Memory in the Age of AI Agents — Agent记忆体系（←"文件即记忆"的学术支撑）
6. Beyond RAG for Agent Memory — 超越RAG的长时序记忆（←记忆模块升级方向）

核心能力

实现 Agent 的 SOUL 五律自进化循环：

1. 已知 vs 未知分离：推理前强制声明已知集合与未知集合
2. 文件即记忆：所有中间状态写入持久化文件，不依赖短期上下文
3. 四向碰撞推理：正面→反面→侧面→整体，禁止过早收敛
4. 人机闭环注入：接受人类证伪输入，结构化吸收并更新世界模型
5. 置信度标注 + 红线守护：每个输出附置信度，红线永不触碰

安装命令

clawhub install self-evolution-cognition
# 或手动安装
cp -r skills/self-evolution-cognition ~/.openclaw/skills/

调用方式

from skills.self_evolution_cognition import SelfEvolutionCognition

cog = SelfEvolutionCognition(workspace="/path/to/project")
result = cog.evolve(task="描述你的任务", context={})
print(result.confidence, result.output)