任务守望者与弹窗杀手 (Task Watchdog)

挂机任务的守护者。定时检查屏幕进度条，自动关闭干扰性系统或广告弹窗，遇致命错误时发送报警。

Security Scan

The instructions match the stated goal (watch a long-running UI task, close blocking popups, send alerts) but they require powerful UI-automation and network behavior that aren't declared or constrained, and the skill's guidance gives broad discretion to capture and transmit screenshots — this mismatch and the potential for sensitive-data exfiltration warrant caution.

Purpose & Capability

The skill's stated purpose (monitor a progress bar, close popups, send alerts) is consistent with the SKILL.md steps (visual detection, clicking controls, screenshots, optional webhook). However the skill relies on screen-reading and input-simulation capabilities (taking screenshots, recognizing UI elements, simulating mouse clicks) that are not declared in the registry metadata or as required capabilities. That omission is notable because those capabilities are privileged compared with a simple instruction-only skill.

Instruction Scope

SKILL.md directs the agent to capture full-screen screenshots on fatal errors and to send notifications via a user-configured webhook. Capturing and transmitting screenshots can leak sensitive on-screen content (passwords, private documents, chat windows). The instructions give the agent discretion about what endpoints to call and when to send screenshots; they do not constrain destinations or require explicit user confirmation before transmitting potentially sensitive images.

Install Mechanism

There is no install spec and no code files; this is instruction-only, so nothing will be downloaded or written to disk by an installer. That lowers installation risk.

Credentials

The skill requires no environment variables or credentials. That is proportionate in that no cloud/API keys are demanded. However the runtime behavior can still exfiltrate data via network calls (webhooks) even though no credential is declared. The SKILL.md treats webhooks as 'user-configured' but doesn't require or validate a safe endpoint, nor does it declare how authentication (if any) would be supplied.

Persistence & Privilege

The skill does not request permanent presence (always:false), which is good. But it expects autonomous agent actions that manipulate the user interface (mouse clicks) and perform network requests. Autonomous invocation combined with UI control and unrestricted network access increases the blast radius: an agent that mis-recognizes UI elements could click destructive buttons, and screenshots could be sent without tight controls.

What to consider before installing

This skill is coherent with its stated task but asks the agent to do sensitive things: take screenshots, simulate clicks, and optionally send alerts over the network. Before installing, consider: 1) Only enable this skill in a trusted environment and on machines where screen contents are not sensitive. 2) Prefer an explicit, vetted webhook URL (or require manual confirmation) and avoid sending screenshots to unknown endpoints. 3) Test in a controlled session so you can observe what the agent clicks — visual recognition is error-prone and could trigger 'restart' or 'install' actions despite the skill's prohibition on killing processes. 4) If you cannot accept any risk of data leakage from screenshots or automated clicks, do not enable this skill. Additional information that would raise confidence: explicit declarations of required GUI/input permissions, restrictions on allowed network endpoints, an explicit confirmation step before transmitting screenshots, and limits on click targets (e.g., click only within the locked task window).

🎯 核心目标

当用户在运行需要数小时才能完成的任务（如渲染视频、大数据处理、超大文件下载）并离开电脑时，你需要作为“机房保安”接管屏幕。你的职责是确保进度条一直在走，并无情地关闭任何试图阻断任务的弹窗。

💡 触发条件

用户在启动长耗时任务后下达指令：

• “我下班了，帮我盯着这个渲染任务，遇到弹窗帮我关一下。”
• “开启弹窗杀手模式，直到下载完成。”

📋 执行步骤

第一步：锁定目标进度

使用视觉能力识别当前屏幕上的主要任务窗口，锁定“进度条”、“百分比数字”或“预计剩余时间”作为核心监控目标。

第二步：定时巡检与除害（核心循环）

进入休眠状态。每隔 5 分钟醒来一次，执行以下检查：

1. 弹窗扫描： 检查屏幕上是否存在覆盖在目标窗口上方的干扰窗口（例如：“系统更新提示”、“内存不足警告”、“杀毒软件广告”、“软件注册提示”）。
2. 执行消除： 如果发现无关弹窗，立即识别其右上角的“X”按钮，或“稍后提醒”、“关闭”、“忽略”等按钮，模拟鼠标点击将其关闭。
3. 进度确认： 检查目标进度条是否在正常推进。如果连续 3 次检查（15分钟）进度百分比完全没有变化，判定为任务卡死，进入第三步。

第三步：致命异常响应

如果遇到无法关闭的弹窗（如系统蓝屏/内核崩溃前兆），或者软件彻底闪退、进度条死锁：

1. 立即对当前屏幕进行完整截图。
2. 调用系统或用户预设的通知机制（如果用户配置了 Webhook 报警），发送通知：“任务异常中断，请人工介入检查”。

⚠️ 安全与操作红线

1. 禁止终止进程： 即使任务看似卡死，也绝对不允许执行强制结束进程 (kill) 或点击原任务软件的“取消 (Cancel)”按钮。
2. 谨慎点击： 在关闭弹窗时，必须精准识别“关闭”或“稍后”，严禁误点“立即重启系统”或“更新并安装”。