Pandaone AI Agent

Pandaone

AI Agent 通过 pandaone write 提交审计 → Pandaone 签发带签名的合规凭证 → 写入完成

🧑‍💼 Pandaone is your codebase's personal manager. / Pandaone 是你代码库的专属经理。

Every AI agent is an employee working in your company — your codebase. Without a personal manager, employees do what they want — edit files, push commits, break production — and you have no idea what they did, when, or why.

Pandaone is the personal manager you hired to protect your code. Every AI agent — Claude, Cursor, Trae, or any tool you adopt tomorrow — must report to Pandaone first. Nothing gets written, nothing gets committed, nothing gets deployed until Pandaone approves it with a signed, auditable receipt:

┌─────────────────────────────────────────────┐
│ [APPROVED]                  2026-09-11 03:47 │
│ File:      src/payment.py                    │
│ Agent:     Claude Code (employee #3)         │
│ Reason:    "Fix rounding"                    │
│ Problem:   "Charge $50,001 on $50,000 bill"  │
│ Approach:  "Round before transfer, not after"│
│ Commit:    a3f7b2c                           │
└─────────────────────────────────────────────┘

Without a personal manager, your agents move money — er, code — without receipts:

Without Pandaone          With Pandaone
──────────────          ──────────────
$50,000 ???            $50,000  src/payment.py
                        Reason:  "Fix rounding"
                        By:      Claude Code
                        When:    2026-09-11 03:47
                        Commit:  a3f7b2c

The left column is how most AI-agent-driven codebases look today. The right column is how yours will look with one command:

pip install pandaone-guard

Your personal manager is on duty 24/7 — before, during, and after every code change. Every AI agent — Claude, Cursor, Trae — operates under one auditable protocol. No exceptions. No bypass.

AI Agent 代码审计门禁 — 让每一次代码改动都留下合规、可追溯的证据链。

-success)

English | 中文

---

中文 / Chinese

是什么？ / What is it?

Pandaone 是一个 7 层防御体系，强制 AI Agent（或任何开发者）写代码前必须经过审计： 任何受保护文件的修改都必须通过 pandaone write 命令，留下 reason / problem / approach 三段式审计记录。

适用场景 / Use Cases：

• 🤖 AI Agent 开发：防止 Agent 绕过审查直接改代码
• 🏢 企业合规：满足 SOC2 / ISO 27001 等代码变更审计要求
• 👥 团队协作：所有 PR 必须有审计记录才能合入

核心特性 / Core Features

• 17 种文本格式 + 24 种二进制格式保护（.py / .md / .json / .yml / .env / .png / .pdf / ...）
• 7 层防御体系 / 7-Layer Defense System (see Defense Layers)
• MCP Server 原生支持 — Claude / Cursor / Trae 等 AI IDE 直连
• 13 种审计日志导出格式（Excel / Word / PDF / SQLite / ...）
• 纯 Python — Windows / macOS / Linux 通吃，无 C 扩展
• 右键菜单集成 — 一行命令在 3 大平台安装原生右键菜单（HKCU / Quick Action / Nautilus + Dolphin）
• i18n 国际化 — zh-CN / en 双语 CLI 自动检测 + --lang 强制 + 持久化偏好

快速开始 / Quick Start

安装 / Installation

方式 A：从 PyPI 安装（最简单）

pip install pandaone-guard

# 一键装上右键菜单（Windows / macOS / Linux 自动检测）
pandaone install-context

方式 B：从源码一键安装（推荐开发者使用，含最新 i18n + CI lint）

git clone https://github.com/hellob1889/Pandaone-AI-Agent.git
cd pandaone

# Windows
powershell -ExecutionPolicy Bypass -File scripts/install.ps1

# macOS / Linux
./scripts/install.sh

安装脚本自动做 / Install script auto-handles：

1. 检测 Python ≥ 3.10
2. 探测 git（不在 PATH 时尝试常见安装路径）
3. 卸载 site-packages 里可能存在的老 pandaone 版本（避免版本冲突）
4. pip install -e . 本地源码 editable 安装
5. 验证 pandaone 可用（python -m pandaone --version）
6. 调用 doctor.py --fix --persist-path 自动修复剩余问题（依赖、setuptools、指纹、PATH 持久化到 HKCU）

30 秒上手（第一次用请先看这里）/ 30-Second Start

⚠️ 包名提醒：本包在 PyPI 上的名字是 pandaone-guard。 旧名 pandax-guard 已弃用（停在 0.7.4，不再更新，也不再修 bug）——如果你之前装过它，请先卸载： pip uninstall pandax-guard && pip install pandaone-guard

# 1) 装（10 秒）
pip install pandaone-guard

# 2) 在你的项目里初始化（1 秒）——会生成 .pandaone/ 审计目录
cd your-project && pandaone init --root .

# 3) 把 MCP 接到你的 AI IDE（Claude / Cursor / Trae / ChatGPT 都行）

MCP 配置（/ claude_desktop_config.json、Cursor MCP 配置）：

{ "mcpServers": { "pandaone": { "command": "pandaone-mcp" } } }

然后直接对 AI 说一句话就够了：

把 main.py 里的 hi 改成 hello，说明你为什么改。

AI 会被迫走审计通道（pandaone_write）：必须填「原因 / 解决什么问题 / 采用什么方法」三件套。合规的改动才被写入 + 自动 git commit + 留痕；不合规的改动被拒绝，同样留痕。

pandaone log --root .      # 查看审计记录（谁改了什么、为什么）
pandaone status --root .   # 查看当前保护状态与 watchdog
pandaone log --root . --export report.html   # 导出 HTML 报告

想先手动挡住 AI 的写操作？pandaone lock --root . 把受保护文件设为只读（OS 级，绕过不了）；pandaone unlock --root . 还原。注意：请以普通用户运行，root 下 OS 权限位对其无效。

环境诊断（任何时候都能跑）/ Environment Diagnostics (run anytime)

python scripts/doctor.py                 # 人类可读
python scripts/doctor.py --json          # CI 用（返回 pass/fail）
python scripts/doctor.py --quiet         # 只显示 WARN / FAIL
python scripts/doctor.py --fix           # 自动修复 13 类问题（仅当前会话生效）
python scripts/doctor.py --fix --persist-path   # 持久化 PATH 到 HKCU（重启 shell 生效）
python scripts/doctor.py --fix-only      # 只跑修复 + 重测，跳过详细诊断

检测 9 个项目 + 自动修复 13 类问题（pip / git / pandaone 装在 site-packages / pandaone.exe PATH / 指纹污染 / setuptools / 5 个运行时依赖）。

--fix --persist-path 会做什么：

• 卸载 site-packages 老版本 → 重装本地源码
• 删除污染的 ~/.pandaone_fp.txt
• 自动 pip install 缺失的运行时依赖
• Windows：用 setx PATH "%PATH%;<Scripts>" 持久化到 HKCU\Environment\Path（新开的 PowerShell 自动看到）
• macOS / Linux：追加 export PATH="..." 到 ~/.bashrc / ~/.zshrc

初始化项目 / Initialize Project

cd /path/to/your-project
pandaone init --root .
# 自动：
#   - 创建 .pandaone/ 目录
#   - 生成 config.json（17 文本 + 24 二进制扩展名）
#   - 创建 binary_snapshots.json（SHA256 字典）
#   - 初始化 .gitignore 排除

写代码（合规路径）/ Write Code (Compliant Path)

# 修改 Python 文件
pandaone write \
    --file src/main.py \
    --reason "修复 user_id 类型注解" \
    --problem "原代码用 int，实际可能是 None" \
    --approach "改为 Optional[int]" \
    --old 'def get_user(user_id: int):' \
    --new 'def get_user(user_id: Optional[int]):'

# 替换二进制文件（图片/文档/PDF 等）
pandaone write \
    --file assets/logo.png \
    --reason "更新品牌 logo" \
    --problem "旧 logo 与新品牌色不匹配" \
    --approach "用新版 logo 替换" \
    --from-file /tmp/new_logo.png

查看审计日志 / View Audit Log

审计历史：每条变更带 reason / problem / approach 三段式凭证

# 命令行查看
pandaone log --last 10

# 导出为 Excel
pandaone log --format xlsx --output audit_report.xlsx

# 导出为 PDF（含中文支持）
pandaone log --format pdf --output audit_report.pdf

# 全部 13 种格式：text / csv / tsv / json / yaml / md / html / xlsx / docx / pdf / sqlite / rst / asciidoc

项目状态仪表盘 / Project Status Dashboard

pandaone status --root .

终端版：项目状态 + 最近审计表（CLI 一行命令）

显示 / Displays：项目路径、配置摘要、L1 锁定状态、二进制快照、最近审计、版本指纹等。 Project path, configuration summary, L1 lock state, binary snapshots, recent audits, version fingerprints, etc.

🆕 Web 实时仪表盘 / Web Real-time Dashboard:

pandaone serve   # 浏览器打开 http://localhost:8765

Web 版：15 条审计记录 · 11 APPROVED / 4 REJECTED · claude-code / cursor / trae 三方协作实时推送

防御体系（7 层） / Defense Layers (7 Layers)

MCP Server（AI Agent 直连 / AI Agent Native）

// claude_desktop_config.json 或 Cursor MCP 配置
// claude_desktop_config.json or Cursor MCP configuration
{
  "mcpServers": {
    "pandaone": {
      "command": "pandaone-mcp",
      "env": {}
    }
  }
}

暴露 11 个工具：pandaone_init / pandaone_lock / pandaone_unlock / pandaone_write / pandaone_log / pandaone_status / pandaone_install_hook / pandaone_watch / pandaone_install_git / pandaone_fingerprint_update / pandaone_ci

所有 11 个工具都已声明完整的 4 个 MCP annotations（readOnlyHint / destructiveHint / idempotentHint / openWorldHint）—— OpenAI 目录合规（v0.7.14 修复）。

在 ChatGPT 中使用 / Use with ChatGPT

ChatGPT Desktop app 的 Developer Mode（开发者模式）原生支持 MCP server。把 pandaone 加进去：

步骤：

1. 打开 ChatGPT Desktop app
2. Settings → Beta features → 开启 Developer Mode
3. Settings → Connectors → Create new connector
4. 填入以下 JSON（字段名一字不差）：

{
  "name": "pandaone-guard",
  "command": "pandaone-mcp",
  "args": [],
  "env": {}
}

1. 第一次对话前，ChatGPT 会要求安装 pandaone-guard。如未自动装，在终端手动跑：

# macOS / Linux
pip3 install pandaone-guard

# Windows (PowerShell)
pip install pandaone-guard

进阶（自定义 venv / Python 路径 / 中文界面）：

{
  "name": "pandaone-guard",
  "command": "/path/to/venv/bin/python",
  "args": ["-m", "pandaone_mcp"],
  "env": {
    "PANDAX_LANG": "zh"
  }
}

环境变量说明 / Environment variables：

只有一个 secret：PANDAX_FP_PASSWORD。本地配置，永不上传到任何远程。

真实场景测试 / Real-World Validation

完整实战验证报告：实战验证报告.md

包含 / Includes：

• 7 层防御每层实战证据 / 7-layer defense, each layer with real-world evidence
• L3 hook 严重 bug 的发现 + 修复
• 隐藏文件锁定 bug 的发现 + 修复
• 13 种导出格式实测 / 13 export formats tested
• MCP 协议完整工作流

详细示例集 / Detailed Examples

• EXAMPLES.md — 10 个真实工作流 + 终端输出
• examples/ — 5 个可运行的 Python demo 脚本
• docs/ — 中英双语文档站（mkdocs）
• docs/index.html — 交互式 HTML 文档首页（带终端动画）

命令清单 / Command List

开发与测试 / Development & Testing

git clone https://github.com/hellob1889/Pandaone-AI-Agent
cd pandaone
pip install -e .[dev]
pytest tests/ -v

当前测试数：342 passed, 1 skipped（覆盖 i18n / write / status / lock / ci / watchdog / export / init / e2e / serve / desktop-icon / gitignore）

路线图 / Roadmap

完整历史：CHANGELOG.md

许可 / License

MIT License — 详见 LICENSE

---

English

What is it?

Pandaone is a 7-layer defense system that forces every code change through an audit gate. Any modification to a protected file (17 text + 24 binary formats) must go through pandaone write, which records reason / problem / approach as immutable audit evidence.

Use cases

• 🤖 AI Agent development: prevent agents from bypassing review
• 🏢 Enterprise compliance: meet SOC2 / ISO 27001 code change audit requirements
• 👥 Team collaboration: every PR must have audit records before merging

Install

Option A: from PyPI (easiest)

pip install pandaone-guard

Option B: from source (recommended for developers, includes latest i18n + CI lint)

git clone https://github.com/hellob1889/Pandaone-AI-Agent.git
cd pandaone

# Windows
powershell -ExecutionPolicy Bypass -File scripts/install.ps1

# macOS / Linux
./scripts/install.sh

The install script automatically:

1. Checks Python >= 3.10
2. Locates git (scans common install paths if not in PATH)
3. Removes stale pandaone from site-packages (prevents version conflicts)
4. pip install -e . local source editable install
5. Verifies pandaone works (python -m pandaone --version)
6. Runs doctor.py --fix --persist-path (auto-fix remaining issues: deps / setuptools / fingerprint / persist PATH to HKCU)

30-Second Start (start here if you're new)

⚠️ Package name: on PyPI this package is pandaone-guard. The legacy name pandax-guard is deprecated (frozen at 0.7.4, no fixes). If you installed it before, switch now: pip uninstall pandax-guard && pip install pandaone-guard

# 1) Install (10s)
pip install pandaone-guard

# 2) Init inside your project (1s) — creates the .pandaone/ audit directory
cd your-project && pandaone init --root .

# 3) Wire MCP into your AI IDE (Claude / Cursor / Trae / ChatGPT)

{ "mcpServers": { "pandaone": { "command": "pandaone-mcp" } } }

Then just say one sentence to your AI:

Change hi to hello in main.py, and tell me why.

The AI is forced through the audit channel (pandaone_write): it must supply reason / problem / approach. Compliant edits are written, auto-committed and logged; non-compliant ones are rejected — and logged too.

pandaone log --root .      # who changed what, and why
pandaone status --root .   # protection status + watchdog
pandaone log --root . --export report.html   # HTML report

Want to hard-block AI writes first? pandaone lock --root . makes protected files read-only at OS level (unbypassable); pandaone unlock --root . restores them. Run as a normal user — OS permission bits do not apply to root.

Environment diagnostics (run anytime)

python scripts/doctor.py                 # human-readable
python scripts/doctor.py --json          # CI mode (returns pass/fail)
python scripts/doctor.py --quiet         # only WARN / FAIL
python scripts/doctor.py --fix           # auto-fix 13 issue classes (session only)
python scripts/doctor.py --fix --persist-path   # persist PATH to HKCU (new shell)
python scripts/doctor.py --fix-only      # just fix + re-check, skip detailed diagnosis

Checks 9 items + auto-fixes 13 issue classes (pip / git / pandaone-in-site-packages / pandaone.exe PATH / fingerprint pollution / setuptools / 5 runtime dependencies).

What --fix --persist-path does:

• Uninstall stale site-packages version, reinstall local source
• Remove polluted ~/.pandaone_fp.txt
• Auto pip install for missing runtime deps
• Windows: use setx PATH "%PATH%;<Scripts>" to persist to HKCU\Environment\Path (new PowerShell shells see it automatically)
• macOS / Linux: append export PATH="..." to ~/.bashrc / ~/.zshrc

Quick start

cd /path/to/your-project
pandaone init --root .

pandaone write \
    --file src/main.py \
    --reason "Fix user_id type annotation" \
    --problem "Original used int, could be None" \
    --approach "Change to Optional[int]" \
    --old 'def get_user(user_id: int):' \
    --new 'def get_user(user_id: Optional[int]):'

7 Defense Layers

MCP Server

{
  "mcpServers": {
    "pandaone": {
      "command": "pandaone-mcp",
      "env": {}
    }
  }
}

Exposes 11 tools for AI agents (Claude / Cursor / Trae).

All 11 tools declare the full set of 4 MCP annotations (readOnlyHint / destructiveHint / idempotentHint / openWorldHint) — OpenAI directory compliant (fixed in v0.7.14).

Use with ChatGPT

ChatGPT Desktop app's Developer Mode natively supports MCP servers. To add pandaone:

Steps:

1. Open ChatGPT Desktop app
2. Settings → Beta features → enable Developer Mode
3. Settings → Connectors → Create new connector
4. Paste this JSON (field names must match exactly):

{
  "name": "pandaone-guard",
  "command": "pandaone-mcp",
  "args": [],
  "env": {}
}

1. ChatGPT will offer to install pandaone-guard on first use. If it doesn't auto-install, run manually:

# macOS / Linux
pip3 install pandaone-guard

# Windows (PowerShell)
pip install pandaone-guard

Advanced (custom venv / Python path / English interface):

{
  "name": "pandaone-guard",
  "command": "/path/to/venv/bin/python",
  "args": ["-m", "pandaone_mcp"],
  "env": {
    "PANDAX_LANG": "en"
  }
}

Environment variables:

Only one secret: PANDAX_FP_PASSWORD. Local config only, never sent to any remote.

License

MIT — see LICENSE

Links

• PyPI Package
• GitHub Repository
• Issue Tracker
• Documentation
• Changelog
• 实战验证报告 (Validation Report)