Self Evolve

Agent 自主进化引擎——让 AI agent 像生物进化一样持续变强。 核心循环：感知差距 → 搜索方案 → 设计实验 → 跑实验 → 选赢家 → 固化 → 下一轮。 使用场景：定期自主进化、能力升级、工作流优化、skill/工具迭代。 关键词：进化, evolution, self-improve, A/B...

Security Scan

The skill's instructions broadly match a self-improvement purpose, but they authorize the agent to run arbitrary shell commands, read and rewrite agent files and code, and autonomously deploy/solidify changes without any enforced protections — this combination is disproportionate and risky.

Purpose & Capability

The name/description (self-evolution of the agent) match the instructions: scanning state.json, running experiments, registering winners, and writing evolution reports. However the skill's claimed capabilities include installing/upgrading tools and modifying agent configs/code (AGENTS.md, TOOLS.md, actual codebases). Those are powerful side-effects that are consistent with 'self-evolve' but are not reflected in any manifest permissions or install-time safeguards. The README even suggests running npx clawhub install, but the skill itself contains no install spec — coherent but operationally heavyweight.

Instruction Scope

Runtime instructions explicitly tell the agent to read/write memory/evolve/state.json, append metrics JSONL, execute telemetry_hook 'command' entries (arbitrary shell commands such as 'cat xxx.log | tail -n 20'), search web/GitHub/Reddit for solutions, deploy chosen solutions, and 'physically solidify' changes into AGENTS.md/TOOLS.md or code. Collecting telemetry via arbitrary shell commands and then using results to modify code/configs grants broad file and system access and could be used to read secrets or exfiltrate data. The guidance relies on agent self-discipline ( Anti-Pseudo Constraint, quality checklist ) but there is no technical enforcement.

Install Mechanism

This is an instruction-only skill with no install spec and no code files to execute; that limits on-disk attack surface. The README's optional npx command is informational only. From an install-mechanism perspective, risk is low because nothing is downloaded or installed by the skill bundle itself.

Credentials

The skill declares no required environment variables or credentials, which superficially is proportional. In practice, the instructions allow running arbitrary shell commands and touching many repo/agent files (memory/evolve, AGENTS.md, TOOLS.md, SOUL.md, logs), and instruct the agent to fetch code/solutions from external sources. Those behaviors can access secrets, tokens, or private files even though no env vars are declared. The manifest therefore understates the level of access the skill expects.

Persistence & Privilege

always is false (good), but the normal autonomous invocation is allowed. Combined with explicit instructions to permanently 'solidify' changes to agent configuration and code, and to register experiments in persistent state files, the skill can change the agent's long-term behavior across restarts. The skill includes human-confirmation rules for high-risk changes in prose, but these are not enforced by the platform — giving a potentially high blast radius if the agent runs autonomously and ignores the guardrails.

What to consider before installing

This skill is coherent with a self-improvement purpose but gives the agent authority to run arbitrary shell commands, read logs and local files, search the web, and modify agent configs and code. Those actions can read secrets or permanently change behavior. Before installing: (1) only run in a sandboxed container or VM, (2) ensure backups/version control for files the skill can change, (3) restrict file permissions so it cannot read sensitive logs/configs, (4) require explicit human approval for any high-risk 'solidify' steps and enforce that outside the skill, and (5) audit any telemetry_hook commands and any external sources the agent is allowed to fetch from. If you cannot enforce those mitigations, treat this skill as unsafe to enable for autonomous runs.

Self-Evolve — Agent 自主进化引擎

生物进化的核心机制：变异 → 选择 → 保留。 这个 skill 把同样的机制应用到 agent 自身——不断发现可改进的角度，用实验找到更好的方案，然后永久固化。

这是什么

Self-Evolve 是一个 meta-skill——它不解决具体业务问题，而是让 agent 自身变强。

• 和 self-think 的边界：self-think 进行反思并发现短板（诊断），self-evolve 对短板寻找解法并跑实验验证（治疗）。
• 不是什么：禁止无目标的空转、禁止为写长篇自省报告而改排版的“伪进化”。（详细约束见: 约束与红线）

快速使用说明 (Quick Start)

触发方式

当被以下情况触发时，请启动 Self-Evolve：

1. 自动触发：self-think 建议进化，或是 cron/HEARTBEAT 定时唤醒。
2. 显式要求：用户下令“进化一下某个能力”。

执行协议 (The 4-Step Tick)

每次触发此技能，你必须执行一次且仅执行一次“滴答巡航”（Tick Handler），绝不能无视状态机一次性跑完实验全流程。 完整并发状态机制及操作边界，请细读： 执行状态机与四步巡航协议

1. Step 1: Status Sync - 扫描 state.json 的活跃实验。
2. Step 2: Record Observations - 将观察指标降噪后写入相应的 jsonl 记录。
3. Step 3: Evaluate & Solidify - 将到期限对比基线固化，并清理任务队列。
4. Step 4: Launch New Experiment - 如并发额度允许（<10），则寻找新瓶颈，搜索方案，设计实验并在状态机注册 OBSERVING。部署要求参考 模板与输出规范。

🚨 注意：部署新实验并更新状态机锁后，本轮巡航彻底结束，必须强制停止（Exit对话），等待下一次唤醒！

参考结构指南

你需要严格遵照以下文件来展开进化，严禁凭感觉行事：

• 执行状态机与四步巡航协议 - How to run a tick (Step 1 to Step 4).
• 约束红线与分级自主权 - What is NOT allowed (Anti-Pseudo Constraint).
• 质量自检 Checklist - Validation for the experimental design and solidification.
• 模板与文件规范 - Data storage paths, YAML format, markdown templates.