claw-windows-automator

当用户想要**Windows自动化**、**自动打开CMD**、**在指定目录执行命令**、**运行bat脚本**、**GitHub自动下载最新版**、**一键下载仓库源码**、**桌面GUI自动化**、**pyautogui任务**、**全屏可视化提示**、**鼠标单击强制停止**、**输入法自动切换**、**...

Security Scan

The skill mostly matches its Windows-automation description, but it performs surprising system-side actions (automatic pip installs at import, venv creation and process restart, and creation of project-level venv/log/model dirs) that are disproportionate to the declared behavior and warrant caution.

Purpose & Capability

Name/description (Windows desktop automation, CMD automation, GitHub downloads, overlay, pyautogui) align with the code (cmd_task, github_download, gradient overlay). However several modules include unrelated heavy infra (Project MODEL_DIR, env_manager logic to install PyTorch/audio libs) that are not needed for basic CMD/GitHub automation.

Instruction Scope

SKILL.md instructs running scripts/operations.py, which will import modules that (a) call ensure_package on import and thus run pip operations, (b) may create a virtualenv and restart the process, and (c) create directories (logs/models/data/venv). Those import-time side-effects go beyond 'open CMD / download zip' and give the skill broad discretion over the host environment.

Install Mechanism

There is no formal install spec, but the code self-installs Python packages via ensure_package (subprocess pip calls). Notably, ensure_package runs fix_setuptools at module import time (unconditional pip install to modify setuptools), which triggers network downloads/installs on import—this is intrusive and higher risk than an instruction-only skill.

Credentials

The skill requests no credentials, but it will create/modify a virtualenv path (VENV_DIR computed relative to project root), write logs, and may install large packages (pip, potentially torch/audio libraries). Those filesystem and network privileges are disproportionate for a simple 'open CMD / download GitHub ZIP' task and could alter the host environment unexpectedly.

Persistence & Privilege

always:false and model invocation is normal. However the skill's bootstrap will create a persistent venv and log directories and can re-launch itself inside that venv—this grants a persistent footprint on disk. It does not modify other skills' configs but will install packages system-wide inside the created venv and may alter global Python environment when run as imported.

What to consider before installing

This skill appears to implement the claimed Windows automation features, but it performs intrusive environment actions without explicit user consent: on import it runs a fix_setuptools pip install, it auto-installs packages, and its env_manager can create a shared venv and re-launch the program inside it. Before installing or running: - Only run on Windows and preferably in an isolated environment (VM/container) or throwaway account. - Inspect and remove/disable ensure_package.fix_setuptools() automatic call (it runs pip on import). Consider making package installation explicit instead of automatic. - Be aware running operations.py will create a venv path (VENV_DIR), logs, and model/data dirs; confirm those paths are acceptable. - If you only need simple CMD automation or GitHub ZIP downloads, consider extracting and running only the minimal modules (cmd_task/github_download and the overlay) after removing auto-install and env_manager bootstrap code. - If you trust the author and intend to use it, run it interactively the first time so you can observe prompts and network activity; otherwise run in an isolated VM and audit network traffic. I assessed this as 'suspicious' (not clearly malicious) because the unexpected auto-install and environment-modifying behavior could be legitimate design choices but are disproportionate and risky without clear prompt/consent. Additional author provenance (official homepage, repo, maintainer identity) or changing the code to remove import-time installs would increase confidence.

OpenClaw Windows Automation Skill

功能：高稳定、可视化、安全可中断的 Windows 桌面自动化引擎。集成全屏 Overlay 实时进度提示、鼠标左键一键强制终止、输入法自动切换、CMD 自动打开执行命令、GitHub 最新源码一键下载等能力。通过统一 CLI 入口，让 LLM 大模型能够可靠地驱动 Windows 完成复杂自动化任务。

触发时机（Triggers）

• 用户明确提到 Windows 自动化、CMD 命令执行、GitHub 下载、项目初始化、批量部署等需求。
• 用户提供文件夹路径并要求“在该目录下执行命令/脚本”。
• 用户提供 GitHub 仓库地址并要求“下载最新版”或“下载源码包”。
• 用户希望看到实时进度提示或需要“鼠标一点就能停止”的安全自动化。
• 用户提到“自动打开命令提示符”“输入法干扰”“pyautogui”“桌面自动化”等关键词。

支持的任务（Tasks）

1. cmd_task（默认推荐）

   • 在指定目录自动打开 CMD 并执行任意命令或 bat 脚本
   • 支持失败重试 + 容错继续模式

2. github_download

   • 自动打开浏览器，智能解析 GitHub /releases/latest，一键下载最新源码 ZIP 包

（后续可无限扩展新任务，只需注册到 OPERATIONS 字典即可被 LLM 调用）

参数提取指南

当决定调用此技能时，请从用户消息中准确提取以下参数：

1. <任务名称> (必填): 必须是 cmd_task 或 github_download 中的一个。
2. <路径> (cmd_task 必填): 要打开 CMD 的文件夹路径（支持相对/绝对路径）。
3. <脚本/命令> (cmd_task 必填): 需要在 CMD 中执行的具体命令或脚本内容。
4. <GitHub地址> (github_download 必填): GitHub 仓库完整地址（如 https://github.com/user/repo）。
5. <容错模式> (cmd_task 选填): 默认开启（continue_on_error=true），出错后继续执行后续步骤。

执行步骤

1. 解析意图：识别用户想要执行的任务类型（cmd_task 或 github_download）以及对应参数。

2. 路径/地址提取：从用户消息中提取文件夹路径或 GitHub URL。

3. 任务选择：根据用户关键词自动匹配对应 task。

4. 调用命令：使用以下兼容性命令启动脚本（优先 python3，失败则 python）。脚本会自动启动 Overlay 提示、执行任务，并在结束或用户点击鼠标时安全退出。

   (python3 scripts/operations.py run --task "<任务名>" [--path "<路径>"] [--script "<脚本>"] [--url "<地址>"]) || (python scripts/operations.py run --task "<任务名>" [--path "<目录路径>"] [--script "<执行脚本>"] [--url "<连接地址>"])