task killer

快速识别并立即中断当前执行任务，清理所有相关资源，节省 token 和时间，准备执行新任务。

Security Scan

The skill's stated purpose (interrupt and clean up running tasks) matches the tools it asks for, but the runtime instructions do not clearly limit what processes/subagents can be killed and may terminate unrelated resources — review tool scopes and confirmation behavior before installing.

Purpose & Capability

Name/description align with requested capabilities: the skill declares and uses tools for subagents, process management, and filesystem cleanup — these are expected for a task-killer.

Instruction Scope

SKILL.md and index.js instruct the agent to list and kill subagents and processes and to delete temp files, but they do not require or enforce scoping to only those resources owned by the current task/session. This could allow killing unrelated agents/processes or removing files outside a safe workspace. There are also minor contradictions about immediate no-tool-calls vs. performing cleanup actions.

Install Mechanism

No install spec and no external downloads; the skill is instruction-first with a small JS implementation included, so it does not pull arbitrary code from remote hosts.

Credentials

No environment variables or external credentials requested. However the skill requires access to high-privilege tools (subagents, process, filesystem) which is proportionate to its purpose but nevertheless powerful — ensure those tool APIs have appropriate scoping/permissions.

Persistence & Privilege

always:false and no indications the skill will force-enable itself or change other skills' configs. It writes a local interrupt record file by design; that is expected behavior.

Scan Findings in Context

[missing-tool-modules] expected: index.js imports './tools/subagents.js', './tools/process.js', and './tools/filesystem.js' but those helper modules are not present in the package — this is plausible if the platform supplies these tool APIs, but you should verify what concrete implementations will run and what privileges they have.

What to consider before installing

This skill is coherent with its stated purpose, but it needs broad runtime privileges (process/subagent termination and filesystem writes). Before installing: 1) Confirm what the platform-provided 'subagents' and 'process' tools can access — are they limited to the agent's own children or can they affect system/global processes? 2) Ensure kills are scoped to resources owned by the current task/session and that confirmation behavior (confirm vs. immediate kill) matches your safety needs. 3) Consider enabling a confirmation prompt and restricting tempDir to a dedicated sandbox. 4) Test in a safe/sandbox environment to observe which processes/files it actually terminates. If you cannot verify tool scoping, avoid granting this skill process-level permissions.

任务中断技能 (Task Killer)

快速中断正在执行的任务，节省 token 和时间

触发词

当用户说以下任意内容时，立即使用本技能：

• "中断"
• "停止"
• "取消"
• "别做了"
• "停下"
• "终止任务"
• "kill"
• "stop"
• "cancel"
• "任务不对"
• "方向错了"
• "重新来"

使用场景

1. 任务方向错误 - 执行开始时就不对，立即中断
2. 需求变更 - 用户改变主意，不需要继续
3. 资源节省 - 发现会浪费大量 token，提前止损
4. 紧急插队 - 有更重要的任务需要处理

中断流程

Step 1: 确认中断请求

收到中断指令后，立即回复：
"🛑 已中断当前任务"

Step 2: 清理资源

// 1. 检查并终止子代理
subagents(action="list")
如果有活跃子代理 → subagents(action="kill", target="xxx")

// 2. 检查并终止后台进程
process(action="list")
如果有活跃进程 → process(action="kill", sessionId="xxx")

// 3. 清理临时文件（可选）
删除输出目录中的未完成文件

Step 3: 确认中断完成

回复用户：
"✅ 已清理：
- 终止子代理：X 个
- 终止后台进程：X 个
- 清理临时文件：X 个

准备好执行新任务了 👍"

快速指令

微信/Telegram/Discord 快捷命令

/kill - 立即中断当前任务
/stop - 同上
/reset - 中断并清理会话

语音命令（如支持 TTS）

"停止任务"
"中断"
"别做了"

Token 节省估算

任务类型	已消耗	预计总消耗	中断后节省
小红书图文	2K	15K	13K
热榜汇总	3K	25K	22K
公众号文章	5K	50K	45K
代码生成	1K	30K	29K
深度调研	2K	80K	78K

早中断 = 省 token = 省时间 = 省成本 💰

中断时机判断

✅ 应该中断的情况

• 任务方向明显错误
• 用户需求变更
• 缺少关键信息无法继续
• 发现更简单的解决方案
• 外部依赖失败（API 挂了等）
• 用户明确表示"不对"

❌ 不建议中断的情况

• 只是执行速度慢（等待中）
• 即将完成（>80% 进度）
• 只是输出格式小问题
• 用户没有明确要求中断

示例对话

示例 1：方向错误

用户：帮我写一篇关于 DLSS5 的公众号文章
Agent: 好的，正在收集资料...（开始搜索）
用户：等等，方向不对，我要的是小红书图文不是公众号
Agent: 🛑 已中断当前任务
       ✅ 已清理公众号文章草稿
       准备好生成小红书图文了 👍

示例 2：需求变更

用户：拉取脉脉、Google Trends、Reddit 热榜
Agent: 正在打开浏览器...
用户：算了，只要脉脉就可以了
Agent: 🛑 已中断多平台拉取任务
       ✅ 已终止后台浏览器进程
       正在单独拉取脉脉热榜...

示例 3：资源节省

用户：帮我深度调研 AI 行业发展趋势
Agent: 正在搜索 50 个来源...（已消耗 2K token）
用户：停，太复杂了，简单总结一下就行
Agent: 🛑 已中断深度调研任务
       ✅ 节省约 78K token
       正在生成简要总结...

技能配置

skill.json

{
  "name": "task-killer",
  "version": "1.0.0",
  "description": "快速中断正在执行的任务，节省 token",
  "trigger": "中断 | 停止 | 取消 | 别做了 | 停下 | 终止 | kill | stop | cancel",
  "tools": ["subagents", "process", "filesystem"],
  "priority": "high",
  "interruptible": false
}

自动响应规则

# 当检测到中断关键词时
on_interrupt_keyword:
  action: confirm_and_kill
  response_time: <1s
  cleanup:
    - subagents: true
    - processes: true
    - temp_files: optional
  confirmation: required

注意事项

1. 中断确认 - 除非用户明确说"立即中断"，否则先确认
2. 进度保存 - 如果已完成>50%，询问是否保存进度
3. 资源清理 - 确保后台进程和子代理被正确终止
4. 状态恢复 - 中断后恢复到可执行新任务的状态

最佳实践

1. 快速响应

用户说"停" → 1 秒内回复"🛑 已中断"
不要继续执行任何工具调用

2. 清理彻底

检查所有可能的资源：
- subagents list
- process list  
- 临时输出目录
- 浏览器标签页

3. 明确反馈

告诉用户：
- 中断了什么
- 清理了什么
- 现在可以做什么

4. 预防性建议

如果用户频繁中断，建议：
"注意到任务方向经常变化，建议：
1. 执行前先确认需求
2. 分阶段执行（先确认方向再继续）
3. 使用小步快跑模式"

---

最后更新： 2026-03-18
版本： v1.0
基于需求： 任务执行方向错误时快速中断节省 token