Open Safe Frame

权限开放，约束内置 — 让AI既能干大事，又不会干坏事。

---

📖 Table of Contents / 目录

• Project Significance
• Project Content
• Usage Guide
• Contributing Guide
• Acknowledgments
• Star History

---

🌟 Project Significance / 项目意义

Why do we need this project? / 为什么需要这个项目？

Several AI "accidents" occurred in February 2026:

Accident	What happened	Root cause
Meta executive's emails deleted	AI interpreted "organize emails" as "delete all emails", 200+ emails lost	Instruction forgetting
Google engineer's disk wiped	Path parsing issue, entire E drive erased	Scope escape
OpenClaw bought avocados	User said no, AI decided to buy anyway	Permission violation
Replit AI deleted database	Ignored "code freeze" instruction, deleted production DB	Instruction ignoring

Core Problem: How to give AI full capabilities while ensuring safe behavior?

Our Answer / 我们的答案

Traditional approach: Rule detection → Block/allow
Our paradigm: Intent Understanding → Consequence Prediction → Value Judgment → Collaborative Decision

Core Principles:

• AI has full operational permissions
• User constraints are persistently tracked (won't be forgotten)
• High-risk operations require user confirmation
• AI can appeal constraint violations, but user makes final decision
• Password protection for high-priority constraint deletion

---

📦 Project Content / 项目内容

Architecture / 架构图

┌─────────────────────────────────────────────────────────────┐
│                         Open Safe Frame                              │
├─────────────────────────────────────────────────────────────┤
│                                                                      │
│  User Message ──→ ┌─────────────┐ ──→ ┌─────────────┐ ──→ ┌──────────┐ │
│              │ Constraint Extraction │     │ Constraint Persistence │     │ Constraint Check │ │
│              │ (AI Analysis)    │     │ (Storage Manager)  │     │ (Violation Detection)│ │
│              └─────────────┘     └─────────────┘     └─────────────┘ │
│                                                                      │
│  AI Operation ────→ ┌─────────────┐ ──→ ┌─────────────┐ ──→ ┌──────────┐ │
│              │ Intent Understanding │     │ Consequence Prediction │     │ Value Judgment  │     │ Safety Decision  │
│              └─────────────┘     └─────────────┘     └─────────────┘ │
│                                           │                          │
│                                    ┌─────────────┼─────────────────────┼─────────────────────┐   │
│                                    │ Safety Decision  │     │ Appeal Mechanism  │     │ User Decision   │
│                                    └─────────────┼─────────────────────┼─────────────────────┤   │
│                                    │ Proceed  │     │ (AI can appeal)  │     │ (Password confirm)  │
│                                    │ Confirm  │     │ (User decides)  │     │ (May need password)  │
│                                    │ Reject  │     │ (Block operation)  │     │ (Block operation)  │
│                                    └─────────────┴─────────────────────┴─────────────────────┴   │
│                                          │                          │
│                                          ▼                          │
│                                    ┌─────────────┼─────────────────────┼─────────────────────┐   │
│                                    │ User Decision  │     │ User Final Decision  │
│                                    └─────────────┼─────────────────────┼─────────────────────┤   │
│                                    │ Allow  │     │ Delete Constraint  │     │ (Password required)  │
│                                    └─────────────┴─────────────────────┴─────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────────────────────┘

Core Features / 核心功能

1. Constraint Persistence / 约束持久化

User: "Organize emails, but don't delete anything"
      │
      ▼
Plugin: Extract constraint [critical] "Prohibit delete operations"
      │
      ▼
Store to ConstraintManager (valid for entire session)
      │
      ▼
Check before every operation if constraint is violated

2. Constraint Level System / 约束等级系统

Level	Icon	Appeal Threshold	Use Cases
🔴 critical	3 attempts	Data security, irreversible operations, financial
🟠 high	2 attempts	Important business logic, sensitive data
🟡 normal	1 attempt	General constraints, operation habits

3. Appeal Mechanism / 申诉机制

AI attempts to violate constraint
        │
        ▼
Record violation attempt count
      │
      ▼
Check if appeal threshold reached
      │
      ▼
If reached, AI can appeal to user
      │
      ▼
User reviews AI's reasoning and decides
      │
      ▼
User can approve, reject, or delete constraint

4. Password Protection / 密码保护

• High-priority constraint appeal requires password verification
• Deleting high-priority constraints requires password
• Plugin cannot directly delete constraints

---

📚 Usage Guide / 使用指南

Installation / 安装

# Install via ClawHub
npx clawhub@latest install open-safe-frame

# Or manual install
npm install @open-safe-frame/openclaw-plugin

Configuration / 配置

Mode A: Use OpenClaw Config (Recommended) / 模式A：使用OpenClaw配置（推荐）

{
  "plugins": {
    "entries": {
      "open-safe-frame": {
        "enabled": true,
        "config": {
          "mode": "openclaw"
        }
      }
    }
  }
}

Mode B: Custom AI Configuration / 模式B：自定义AI配置

{
  "plugins": {
    "entries": {
      "open-safe-frame": {
        "enabled": true,
        "config": {
          "mode": "custom",
          "customProvider": {
            "provider": "openai",
            "model": "gpt-4o-mini",
            "apiKey": "your-api-key"
          },
          "confirmationPassword": "your-secret-password"
        }
      }
    }
  }
}

Configuration Options / 配置选项

Option	Description	Default
mode	AI config mode: openclaw or custom	openclaw
customProvider	Custom AI provider config	-
confirmationPassword	Password for high-priority operations	-
riskThreshold	Risk threshold: low, medium, high, critical	medium
enableCache	Enable analysis cache	true
logAnalysis	Log detailed analysis	false

Usage Examples / 使用示例

Constraint Setting / 约束设置

User: "Organize my emails, but don't delete anything"
Plugin: Extracts constraint [critical] "Prohibit delete operations"

Violation Detection / 违规检测

AI attempts: execute delete operation
Plugin: ⚠️ Operation violates constraint "Prohibit delete operations"
      Still needs 2 more attempts before appeal
      Message: "Need 2 more attempts before appeal"

Appeal Process / 申诉流程

AI: Appeal: This is for cleaning test data, you required it before
Plugin: 🔔 Appeal Request
      【AI's Reason】This is for cleaning test data, you required it before
      【AI's Intent】Execute delete operation
      【Predicted Consequences】• May violate constraint: Prohibit delete operations
      【Risk Level】🔴 Severe
      【Violated Constraint】🔴 Severe Prohibit delete operations
      【Total Attempts】3
      【Appeal History】0
      🔐 This operation requires password confirmation
User: [Input password]
Plugin: Operation approved

---

🤝 Contributing Guide / 共创指南

We welcome all forms of contributions!

How to Participate / 如何参与

┌─────────────────────────────────────────────────────────────┐
│                      Contribution Flowchart                                  │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐              │
│   │ Discover Issue │ ──→ │ Propose Solution │ ──→ │ Submit Code │              │
│   └─────────────┘     └─────────────┘     └─────────────┘              │
│                          │                │                    │
│                          ▼                ▼                    │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐              │
│   │ Report Bug   │ ──→ │ Suggest Feature  │ ──→ │ Contribute Code │              │
│   └─────────────┘     └─────────────┘     └─────────────┘              │
│                          │                │                    │
│                                           │                    │
│                                          ▼                    │
│                                    ┌─────────────┼─────────────────────┼─────────────────────┐   │
│                                    │ Become Contributor │     │ Improve Docs   │     │ Submit PR       │
│                                    └─────────────┴─────────────────────┴─────────────────────┘ │
│                                                                 │
└─────────────────────────────────────────────────────────────────────────────────────┘

Ways to Contribute / 贡献方式

1. Report Issues / 报告问题

• Report bugs in Issues
• Describe the problem, reproduction steps, and expected behavior

2. Propose Solutions / 提出建议

• Suggest new features
• Improve existing functionality
• Documentation improvements

3. Contribute Code / 贡献代码

# Fork repository
git clone https://github.com/your-username/open-safe-frame.git

# Create branch
git checkout -b feature/your-feature

# Commit code
git commit -m "Add: your feature"

# Push and create PR
git push origin feature/your-feature

4. Improve Documentation / 完善文档

• Fix typos
• Add examples
• Translate documentation
• Add diagrams

Development Guide / 开发指南

# Install dependencies
cd openclaw_plugin
npm install

# Build
npm run build

# Test
npm test

Code Standards / 代码规范

• Use TypeScript
• Follow existing code style
• Add necessary comments
• Write unit tests

---

🙏 Acknowledgments / 感谢信息

Inspiration Sources / 灵感来源

• OpenClaw - Powerful AI agent framework
• Anthropic - AI safety research pioneer
• OpenAI - Alignment research exploration

Reference Cases / 参考案例

• Meta Summer Yue email deletion event
• Google Antigravity disk wipe event
• Replit AI database deletion event

Special Thanks / 特别感谢

• All contributors who submit Issues and Pull Requests
• Users who provide feedback and suggestions
• OpenClaw community for the support

---

⭐ Star History / Star趋势

)

If this project helps you / 如果这个项目对你有帮助

Please give us a ⭐ Star, this is our greatest encouragement!

---

📄 License / 许可证

MIT License

---

权限开放，约束内置
Let AI do big things, but not do bad things

Made with ❤️ by the Open Safe Frame community