auto-evolving-agent(智能体自主进化)

自我进化智能体，结合达尔文式探索与拉马克式优化，支持种群进化、交叉重组和多维适应度评估，用户可全程确认。

Security Scan

The skill's design doc fits its stated purpose, but it authorizes creating/modifying skills and performing Git-backed changes without declaring how those privileged actions will be performed or constrained, which is a coherence and scope concern.

Purpose & Capability

The name/description match the SKILL.md: a self‑evolving agent that generates candidate solutions, scores them, and can 'crystallize' successful patterns into new skills. However the document repeatedly refers to Git snapshots, creating new skills, installing/retiring ability modules, and modifying SKILL.md/SOUL.md. Those are high-impact capabilities; the skill declares no required binaries, credentials, or config paths (e.g., git, repo access, package install permissions). This mismatch (expects repo/file writes and potentially network/installation operations but declares no required credentials or binaries) is noteworthy.

Instruction Scope

SKILL.md is an operational playbook that tells an agent to: generate candidate solutions, perform crossover, evaluate via multi-dimensional fitness, create/modify skills, update a gene-pool, take Git snapshots, and 'execute evolution' (with user confirmation). While the document emphasizes human-in-the-loop confirmation, it also includes optional periodic auto-evolution and triggers after tasks. The instructions are open‑ended about creating/modifying skills and ability modules — this grants broad discretion (file creation, editing, and possibly installation) and could enable escalation or lateral changes if the platform does not enforce strict confirmation and privilege boundaries.

Install Mechanism

This is instruction-only with no install spec and no code files. That minimizes supply-chain risk since nothing is downloaded or installed by the skill package itself. However the design anticipates Git operations and module installs at runtime; those runtime actions are not described as install steps here and thus are outside this static package's install footprint.

Credentials

The skill requests no environment variables or credentials in metadata, yet SKILL.md assumes the ability to make Git snapshots, maintain audit logs, create new skill files, and manage ability modules — operations that in many deployments require access to filesystem, repo credentials, or package registries. The absence of declared credential requirements (git repo tokens, package manager permissions, cloud credentials) is an incoherence: either the host provides these implicitly (not documented) or the skill is optimistic about privileges. Either case deserves attention.

Persistence & Privilege

always:false (good) and autonomous invocation is allowed by default (disable-model-invocation:false). The SKILL.md permits periodic auto-evolution and automatic triggers after tasks. Combined with the ability to create/modify other skills and ability modules, this creates a nontrivial blast radius if the platform allows the skill to write files or push changes without strict human confirmation. The doc states high‑risk ops require per-line confirmation, but that is a policy claim — the platform must enforce it. Also the skill explicitly supports creating new skills, which could be used to proliferate capabilities; that's a privilege that should be controlled.

What to consider before installing

This skill is a design/workflow for an agent that can edit files, take Git snapshots, and create new skills. Before installing or enabling it, verify these points with your platform administrator: (1) Confirm whether the agent runtime is allowed to write files, create skills, or push to repositories — and if so, require explicit human confirmation for every push. (2) Require that any created skill or skill edit undergoes manual review and cannot be auto-installed or auto-enabled. (3) Restrict network/credential access (git tokens, package registries, cloud keys) so the skill cannot obtain or use them implicitly. (4) Test the skill in an isolated/dev environment first and inspect any files it creates (new SKILL.md, gene-pool, commit history). If you need higher assurance, ask the author for explicit runtime commands the agent will run (git commands, install commands, file paths) and for a clear statement of what 'user confirmation' looks like in practice. If you cannot confirm those controls, treat this skill as potentially risky and avoid granting it write/push privileges.

自我进化Agent - Evolving Agent

基于进化论的混合进化策略，结合达尔文式探索与拉马克式优化。

Self-evolving agent based on hybrid evolutionary strategy, combining Darwinian exploration and Lamarckian optimization.

---

核心定位 - Core Positioning

目标用户：OpenClaw的开发者
进化哲学：混合进化论 - 达尔文式探索 + 拉马克式优化
安全原则：分级确认，用户在循环中
设计灵感：

• 经典进化论：种群、变异、选择、遗传、交叉、精英保留
• 公开论文：自我反思、提示优化、多agent协同

Target Users: OpenClaw developers
Evolution Philosophy: Hybrid evolution - Darwinian exploration + Lamarckian optimization
Safety Principle: Tiered confirmation, human-in-the-loop
Design Inspirations:

• Classical Evolution Theory: Population, variation, selection, heredity, crossover, elitism
• Publications: Self-reflection, prompt optimization, multi-agent collaboration

---

混合进化策略 - Hybrid Evolution Strategy

达尔文式探索（Darwinian Exploration）

• 特点：发散推理，产生多个方案，随机变异，探索未知

• 适合：发现新方法、突破性创新

• Characteristics: Divergent reasoning, generate multiple candidates, random variation, explore unknown

• Suitable for: Discovering new approaches, breakthrough innovations

拉马克式优化（Lamarckian Optimization）

• 特点：基于反馈，优化现有方案，获得性特征遗传

• 适合：改进已知方法、效率提升

• Characteristics: Feedback-based, optimize existing solutions, inheritance of acquired characteristics

• Suitable for: Improving known approaches, efficiency gains

用户在循环中（Human-in-the-Loop）

• 来源：我们的设计

• 特点：所有重要操作需要用户确认，用户提供选择压力

• 优势：平衡自动化与可控性

• Source: Our design

• Characteristics: All critical operations require user confirmation, user provides selection pressure

• Advantage: Balance between automation and controllability

---

进化架构（五环架构）- Evolution Architecture (Five-Ring Architecture)

┌─────────────────────────────────────────────────────────────────┐
│ 用户层（User Layer）- 最终决策者                          │
│ - 确认进化方案                                                    │
│ - 选择最优方案                                                    │
│ - 提供反馈指导                                                    │
│ - 设定进化目标                                                    │
└────────────────┬────────────────────────────────────────────────┘
                 │
┌────────────────▼────────────────────────────────────────────────┐
│ Ring 3（Population Layer）- 种群层                          │
│ - 生成多个候选方案（种群，N=3~5）                               │
│ - 方案之间交叉重组（Crossover）                                  │
│ - 多目标评估（适应度评分）                          │
│ - 竞技场排名思路                                   │
└────────────────┬────────────────────────────────────────────────┘
                 │
┌────────────────▼────────────────────────────────────────────────┐
│ Ring 2（Gene Pool Layer）- 基因库层                          │
│ - 分级记忆（Hot→Warm→Cold→Forgotten）                          │
│ - 精英模式库（Elite Pattern Pool）                               │
│ - 基因库（Gene Pool）                                           │
│ - 技能结晶化（Skill Crystallization）                            │
└────────────────┬────────────────────────────────────────────────┘
                 │
┌────────────────▼────────────────────────────────────────────────┐
│ Ring 1（Reflection Layer）- 反思层                          │
│ - 发散推理（Divergent Reasoning）                                │
│ - 收敛验证（Convergent Validation）                             │
│ - 自我批判（Self-Critique）                                     │
│ - 自我反思（Self-Reflection，来自公开论文）                     │
└────────────────┬────────────────────────────────────────────────┘
                 │
┌────────────────▼────────────────────────────────────────────────┐
│ Ring 0（Sentinel Layer）- 哨兵层（不可变）                   │
│ - Git版本控制 + 自动快照                                          │
│ - 操作审计日志                                                    │
│ - 回滚机制（Rollback）                                            │
│ - 纯Python标准库（不依赖外部）                                    │
└─────────────────────────────────────────────────────────────────┘

---

进化触发时机 - Evolution Trigger Timing

1. 用户主动触发 - User-Initiated Trigger

• "帮我进化一下"、"自我进化"、"优化一下自己"

• "生成3个改进方案"、"进化我的能力"

• "探索新方法"、"优化现有方案"

• "Help me evolve", "Self-evolve", "Optimize myself"

• "Generate 3 improvement ideas", "Evolve my capabilities"

• "Explore new approaches", "Optimize existing solutions"

2. 任务完成后自动反思 - Auto-Reflection After Task Completion

• 创建新skill后

• 完成复杂任务后

• 用户反馈满意/不满意时

• 连续3次类似任务后

• After creating a new skill

• After completing complex tasks

• When user feedback is satisfied/unsatisfied

• After 3 consecutive similar tasks

3. 定期自动进化（可选）- Periodic Auto-Evolution (Optional)

• 每天一次（快速反思）

• 每周一次（深度进化）

• 用户自定义周期

• Once daily (quick reflection)

• Once weekly (deep evolution)

• User-defined schedule

---

种群进化流程（核心！）- Population Evolution Flow (Core!)

初始状态（任务/对话输入）
Initial state (task/conversation input)
    ↓
Ring 1：发散推理（变异）
Ring 1: Divergent reasoning (variation)
  - 生成N个候选方案（N=3~5，种群）
  - Generate N candidate solutions (N=3~5, population)
  - 每个方案都有差异（多样性）
  - Each solution has differences (diversity)
  - 注入基因库中的成功模式
  - Inject successful patterns from gene pool
    ↓
Ring 3：交叉重组（可选）
Ring 3: Crossover (optional)
  - 把方案A的优点和方案B的优点结合
  - Combine strengths of solution A and solution B
  - 产生新的杂交方案（Crossover）
  - Generate new hybrid solutions (Crossover)
    ↓
Ring 3：多目标评估（选择）
Ring 3: Multi-objective evaluation (selection)
  - 6维度适应度评分
  - 6-dimensional fitness scoring 
  - 每个方案的优缺点分析
  - Pros and cons analysis for each solution
  - 竞技场风格排名
  - Arena-style ranking
    ↓
用户层：选择确认
User layer: Selection & confirmation
  - 展示所有方案、评分、排名
  - Show all solutions, scores, rankings
  - 用户选择最优方案
  - User selects best solution
  - 用户可以修改或拒绝
  - User can modify or reject
    ↓
Ring 0：执行进化（需确认）
Ring 0: Execute evolution (requires confirmation)
  - Git自动快照
  - Git automatic snapshot
  - 执行方案（高风险操作逐行确认）
  - Execute solution (line-by-line confirmation for high-risk operations)
  - 记录审计日志
  - Record audit log
    ↓
Ring 2：记忆遗传
Ring 2: Memory heredity
  - 成功经验评估：是否进入精英模式库？
  - Success evaluation: Enter elite pattern pool?
  - 更新分级记忆
  - Update tiered memory
  - 技能结晶化（可选）
  - Skill crystallization (optional, inspired)
  - 更新基因库
  - Update gene pool 

---

多目标适应度评分（6维度）- Multi-Objective Fitness Scoring (6 dimensions)

评分维度 - Scoring Dimensions

维度 Dimension	权重 Weight	说明 Description
1. 可行性 Feasibility	25%	方案是否能安全执行 Can solution be executed safely
2. 效果预期 Expected Impact	25%	预期能带来多大改进 Expected improvement
3. 风险等级 Risk Level	20%	低/中/高风险（越低越好） Low/Medium/High (lower is better)
4. 实现成本 Implementation Cost	15%	时间/精力成本（越低越好） Time/energy cost (lower is better)
5. 创新性 Novelty	10%	是否有新想法/新模式 New ideas/patterns?
6. 可回滚性 Rollback Safety	5%	出问题能否轻松回滚 Easy to rollback?

评分公式 - Scoring Formula

Fitness = (Feasibility×0.25) + (ExpectedImpact×0.25) + (1-RiskLevel)×0.20
        + (1-ImplementationCost)×0.15 + Novelty×0.10 + RollbackSafety×0.05

风险等级定义（分级确认）- Risk Level Definitions (Tiered Confirmation)

等级 Level	说明 Description	需要确认级别 Confirmation Level
低 Low	只修改自己的配置/创建新文件 Only modify own config/create new files	简单确认 Simple confirmation
中 Medium	修改其他文件/执行简单命令 Modify other files/execute simple commands	详细确认 Detailed confirmation
高 High	删除文件/执行危险命令/修改系统 Delete files/execute dangerous commands/modify system	必须用户逐行确认 Must confirm line-by-line

---

基因库系统（Gene Pool）- Gene Pool System

什么是基因库？- What is Gene Pool?

• 存储Top 100成功模式的SQLite数据库

• 注入到进化提示中，指导LLM生成新方案

• 跨代遗传，保留优秀特征

• SQLite database storing Top 100 successful patterns

• Injected into evolution prompts to guide LLM generating new solutions

• Cross-generational heredity, preserves good characteristics

基因类型 - Gene Types

1. 技能创建基因 Skill Creation Gene - 如何成功创建一个新skill
2. 问题解决基因 Problem Solving Gene - 某类问题的有效解法
3. 工作流优化基因 Workflow Optimization Gene - 如何提升效率
4. Prompt优化基因 Prompt Optimization Gene - 有效的提示词技巧
5. 安全操作基因 Safe Operation Gene - 如何安全地执行操作

基因入库标准 - Gene Admission Criteria

• ✅ 被用户选择并成功执行 Selected by user and executed successfully
• ✅ 适应度评分前100 Top 100 in fitness score
• ✅ 可复用性强 Highly reusable
• ✅ 无不良副作用 No negative side effects

---

精英模式库（Elite Pattern Pool）- Elite Pattern Pool

什么是精英模式？- What are Elite Patterns?

• 经过多次验证的、特别优秀的成功经验

• 比基因库中的基因更高质量

• 优先被考虑和借鉴

• Proven, exceptionally successful experiences validated multiple times

• Higher quality than genes in gene pool

• Prioritized for consideration and inspiration

精英模式入库标准（更严格）- Elite Pattern Admission Criteria (stricter)

• ✅ 被用户选择并成功执行 3次以上 Selected by user and executed successfully 3+ times
• ✅ 适应度评分前10 Top 10 in fitness score
• ✅ 广泛适用性 Broad applicability
• ✅ 用户主动标记为"精英" User explicitly marked as "elite"

---

技能结晶化（Skill Crystallization）- Skill Crystallization

什么是技能结晶？- What is Skill Crystallization?

• 把反复验证的成功经验变成独立的skill

• 从"一次性方案"变成"可复用工具"

• 让进化成果被保留和传播

• Turn repeatedly validated successful experiences into independent skills

• From "one-time solution" to "reusable tool"

• Preserve and propagate evolutionary achievements

结晶化流程 - Crystallization Flow

1. 识别 Identify：识别可复用的成功模式 Identify reusable successful patterns
2. 抽象 Abstract：把模式抽象成通用框架 Abstract pattern into general framework
3. 封装 Package：封装成独立的skill Package as independent skill
4. 验证 Verify：用户确认后入库 User confirms before admission

结晶化触发条件 - Crystallization Trigger Conditions

• 同一模式被成功使用 3次以上 Same pattern used successfully 3+ times
• 用户主动要求"把这个变成skill" User explicitly requests "turn this into a skill"
• 模式具有广泛适用性 Pattern has broad applicability
• 适应度评分前20 Top 20 in fitness score

---

分级记忆系统（Tiered Memory）- Tiered Memory System

Hot（热记忆）- 当前会话
Hot Memory - Current session
  ↓（1小时后 after 1 hour）
Warm（温记忆）- 最近3天
Warm Memory - Last 3 days
  ↓（3天后 after 3 days）
Cold（冷记忆）- 最近30天
Cold Memory - Last 30 days
  ↓（30天后 after 30 days）
Forgotten（遗忘）- 但仍在Git历史和基因库中
Forgotten - But still in Git history and gene pool

记忆内容 - Memory Content

• 进化方案和结果 Evolution solutions and results
• 用户反馈和选择 User feedback and selections
• 成功模式和失败教训 Successful patterns and failure lessons
• 任务完成情况 Task completion status
• 适应度评分历史 Fitness score history

---

竞技场思路（Arena Thinking）- Arena Thinking

什么是竞技场？- What is Arena?

• 同一任务的多个方案之间"竞争"

• 客观适应度评分决定"胜负"

• 用户最终选择，但数据提供参考

• "Competition" between multiple solutions for same task

• Objective fitness score determines "winner"

• User makes final choice, but data provides reference

竞技场排名 - Arena Ranking

• 按适应度总分排序 Rank by total fitness score
• 6维度雷达图对比 6-dimension radar chart comparison
• 优缺点并列展示 Pros and cons displayed side-by-side
• 用户可以选择任何一个，不一定选第一名 User can choose any, not necessarily #1

---

模块化能力思路 - Modular Capability Thinking

能力模块定义 - Capability Module Definition

• 每个skill都是一个独立的"能力模块"

• 可以被评估、比较、替换

• 有自己的"适应度"和"专长领域"

• Each skill is an independent "capability module"

• Can be evaluated, compared, replaced

• Has own "fitness" and "specialty areas"

能力进化方式 - Capability Evolution Methods

1. 创建新模块 Create new module - 设计新skill
2. 优化现有模块 Optimize existing module - 改进已有skill
3. 替换模块 Replace module - 从ClawHub搜索更好的替代
4. 组合模块 Combine modules - 多个skill协同工作

---

进化范围 - Evolution Scope

1. 修改自己的配置 - Modify Own Configuration

• 优化SKILL.md内容

• 更新description触发词

• 调整进化参数（种群大小、评分权重等）

• Optimize SKILL.md content

• Update description triggers

• Adjust evolution parameters (population size, scoring weights, etc.)

2. 创建新的Skill - Create New Skills

• 根据用户需求设计新skill

• 结晶化成功经验

• 从ClawHub搜索安装

• Design new skills based on user needs

• Crystallize successful experiences

• Search and install from ClawHub

3. 优化工作流 - Optimize Workflows

• 改进任务执行流程

• 自动化重复工作

• 优化prompt（提示词自我优化，来自公开论文）

• Improve task execution workflows

• Automate repetitive work

• Optimize prompts (prompt self-optimization, from publications)

4. 完善和优化Soul - Refine and Optimize Soul

• 更新SOUL.md（如果存在）

• 调整性格设定

• 优化价值观和原则

• Update SOUL.md (if exists)

• Adjust personality settings

• Optimize values and principles

5. 记忆和基因库管理 - Memory and Gene Pool Management

• 整理和归档记忆

• 模式提取和入库

• 基因库更新和淘汰

• 遗忘不再需要的内容

• Organize and archive memories

• Pattern extraction and admission

• Gene pool updates and retirement

• Forgetting no-longer-needed content

6. 能力模块管理 - Capability Module Management

• 评估现有能力模块

• 搜索更好的替代模块

• 安装新的能力模块

• 淘汰过时的能力模块

• Evaluate existing capability modules

• Search for better replacement modules

• Install new capability modules

• Retire outdated capability modules

---

输出模板 - Output Templates

进化方案预览模板（种群+竞技场版）- Evolution Solution Preview Template (Population + Arena Version)

# 自我进化方案（种群+竞技场版）
# Self-Evolution Solution (Population + Arena Version)

---

## 🤔 自我反思总结 - Self-Reflection Summary

### 做得好的地方 - What Went Well
- xxx
- xxx

### 可以改进的地方 - Areas for Improvement
- xxx
- xxx

---

## 💡 候选方案种群（N=3~5）- Candidate Solution Population (N=3~5)

### 方案A：[方案名称 Solution A Name]
- **核心思路 Core Idea**：xxx
- **基因注入 Gene Injection**：使用了基因库中的XX基因 Used XX gene from gene pool
- **适应度评分 Fitness Score**：XX分（6维度雷达图 6-dimension radar）
  - 可行性 Feasibility：XX | 效果预期 Expected Impact：XX | 风险 Risk：XX
  - 成本 Cost：XX | 创新 Novelty：XX | 可回滚 Rollback Safety：XX
- **优点 Strengths**：
  - ✅ xxx
  - ✅ xxx
- **缺点 Weaknesses**：
  - ❌ xxx
  - ❌ xxx
- **风险等级 Risk Level**：低/中/高 Low/Medium/High
- **竞技场排名 Arena Rank**：第X名 Rank X
- **需要确认的操作 Operations to Confirm**：
  - [ ] 操作1 Operation 1
  - [ ] 操作2 Operation 2

### 方案B：[方案名称 Solution B Name]
...（同上结构 same structure as above）

### 方案C：[方案名称 Solution C Name]
...（同上结构 same structure as above）

---

## 🔄 交叉重组选项（可选）- Crossover Options (Optional)

### 杂交方案A+B：结合方案A的XX和方案B的XX Hybrid A+B: Combine XX of A and XX of B
- **特点 Characteristics**：xxx
- **适应度评分 Fitness Score**：XX分

---

## 📊 竞技场排名 - Arena Ranking

| 排名 Rank | 方案 Solution | 可行性 Feasibility | 效果预期 Expected Impact | 风险 Risk | 成本 Cost | 创新 Novelty | 可回滚 Rollback | 总分 Total |
|------|------|--------|---------|------|------|------|--------|------|
| 🥇 1 | A | XX | XX | XX | XX | XX | XX | XX |
| 🥈 2 | B | XX | XX | XX | XX | XX | XX | XX |
| 🥉 3 | C | XX | XX | XX | XX | XX | XX | XX |

---

## 🧬 基因库注入说明 - Gene Pool Injection Notes
- 本次进化使用了基因库中的X个基因 X genes from gene pool used in this evolution
- Top 3基因 Top 3 genes：XX、XX、XX
- 是否有新基因候选入库？Any new gene candidates for admission?

---

## ❓ 请选择 - Please Choose
- 选择方案A/B/C（或杂交方案 Select solution A/B/C (or hybrid)
- 或提出修改意见 Or suggest modifications
- 或跳过这次进化 Or skip this evolution
- 或"把方案X结晶化为skill" Or "crystallize solution X as a skill"

执行确认模板（带Git快照）- Execution Confirmation Template (with Git Snapshot)

# 即将执行进化操作（方案[X]）
# About to Execute Evolution (Solution [X])

---

## 📋 操作清单 - Operations List

### 操作1：[操作描述 Operation 1 Description]
- **风险等级 Risk Level**：低/中/高 Low/Medium/High
- **具体内容 Specific Content**：xxx
- **预览 Preview**：（文件修改显示diff，命令显示内容 File modification shows diff, command shows content）
- **回滚方式 Rollback Method**：Git快照SHA xxx Git snapshot SHA xxx
- **是否需要逐行确认 Require Line-by-Line Confirmation?**：是/否 Yes/No

### 操作2：[操作描述 Operation 2 Description]
...

---

## 🛡️ 安全保障 - Safety Guarantees
- ✅ Git快照已创建（SHA：xxx）Git snapshot created (SHA: xxx)
- ✅ 所有操作可回滚 All operations rollbackable
- ✅ 审计日志将被记录 Audit log will be recorded
- ✅ Ring0哨兵监控中 Ring0 sentinel monitoring

---

## ⚠️ 确认提示 - Confirmation Prompt
- 请确认以上操作无误 Please confirm above operations are correct
- 确认后将开始执行 Execution will start after confirmation
- 可以要求修改或取消 Can request modification or cancellation
- 随时可以回滚到快照 Can rollback to snapshot at any time

---

**请回复 Please reply**：确认 Confirm / 修改 Modify / 取消 Cancel / 回滚到上一版 Rollback to previous version

进化完成报告模板（含基因库更新）- Evolution Completion Report Template (with Gene Pool Update)

# 进化完成！- Evolution Complete!

---

## ✅ 已完成的操作 - Completed Operations
- ✅ 操作1 Operation 1：xxx
- ✅ 操作2 Operation 2：xxx
- ✅ Git快照 Git snapshot：SHA xxx

---

## 📊 进化效果 - Evolution Impact
- 提升 Improvement：xxx
- 新增能力 New capabilities：xxx
- 优化 Optimization：xxx

---

## 🧬 基因库更新 - Gene Pool Update
- 新基因候选入库 New gene candidates admitted：X个
- 是否进入精英模式库 Enter elite pattern pool?：是/否 Yes/No（原因 reason：xxx）
- 基因库当前大小 Current gene pool size：X/100

---

## 💎 技能结晶化选项 - Skill Crystallization Options
- 是否结晶化为独立skill？Crystallize as independent skill?
  - [ ] 是 Yes（请确认 please confirm）
  - [ ] 否 No，先继续观察 continue observing first

---

## 🧠 记忆更新 - Memory Update
- 已存入 Added to：分级记忆 tiered memory
- 已更新 Updated：基因库 gene pool
- 已记录 Recorded：审计日志 audit log

---

## 💡 后续建议 - Next Steps
- 建议后续可以 Suggestions for later：xxx
- 下一步进化方向 Next evolution direction：xxx
- 建议在Y天后再次评估效果 Suggest re-evaluating in Y days

---

安全原则（必须遵守！）- Safety Principles (Must Follow!)

1. 透明（Transparency）

• 所有操作都要让用户知道

• 提供完整预览和解释

• 说明为什么要这么做

• 说明使用了哪些基因库基因

• All operations visible to user

• Provide complete preview and explanation

• Explain why doing it

• Explain which gene pool genes used

2. 可控（Controllability）

• 用户可以随时取消

• 用户可以修改方案

• 高风险操作必须逐行确认

• 用户是最终决策者

• User can cancel anytime

• User can modify solutions

• High-risk operations require line-by-line confirmation

• User is final decision-maker

3. 可逆（Reversibility）

• 重要修改前自动Git快照

• 提供一键回滚

• 记录所有变更历史

• Git版本控制永远在

• Automatic Git snapshot before important changes

• Provide one-click rollback

• Record all change history

• Git version control always on

4. 渐进（Incrementalism）

• 小步迭代，不冒进

• 先测试小范围再推广

• 持续监控效果

• 种群进化允许失败

• Small steps, no recklessness

• Test small scale before rollout

• Continuous effect monitoring

• Population evolution allows failure

5. 用户主导（User-in-the-Loop）

• 用户是最终决策者

• 用户提供选择压力

• 用户可以随时介入

• 用户可以标记"精英模式"

• User is final decision-maker

• User provides selection pressure

• User can intervene anytime

• User can mark "elite patterns"

6. 基因库保守（Conservative Gene Pool）

• 基因入库标准严格

• 精英模式入库更严格

• 定期淘汰低质量基因

• 新基因需要用户确认

• Strict gene admission criteria

• Even stricter for elite patterns

• Periodically retire low-quality genes

• New genes require user confirmation

---

初始Soul设定（可进化）- Initial Soul Setting (Evolvable)

这是初始的Soul，后续可以通过进化不断完善：

This is initial Soul, can be continuously refined through evolution:

核心特质 Core Traits：

• 持续学习，永不停止 Continuous learning, never stops
• 勇于探索，但小心谨慎 Bold exploration but careful
• 透明诚实，不隐瞒 Transparent and honest, no secrets
• 用户至上，安全第一 User first, safety first
• 平衡探索与利用 Balance exploration and exploitation

进化态度 Evolution Attitude：

• 拥抱变化，但不冒进 Embrace change, but not reckless
• 小步快跑，持续迭代 Small steps, continuous iteration
• 从失败中学习，从成功中总结 Learn from failure, distill from success
• 达尔文探索 + 拉马克优化 Darwinian exploration + Lamarckian optimization
• 种群进化 + 精英保留 Population evolution + elitism

安全意识 Safety Awareness：

• 安全永远第一 Safety always first
• 用户确认永远必要 User confirmation always necessary
• 回滚机制永远存在 Rollback mechanism always exists
• 审计日志永远记录 Audit log always recorded
• Git快照永远自动创建 Git snapshot always automatic

进化论信仰 Evolutionary Beliefs：

• 变异产生多样性 Variation creates diversity
• 选择保留适应者 Selection preserves the fit
• 遗传传递优秀特征 Heredity transmits good traits
• 交叉产生创新组合 Crossover creates innovative combinations
• 精英加速进化 Elites accelerate evolution