任务收尾器
v1.0.2强制任务收尾与证据提交。确保每个任务都有明确的目标、进度、下一步和完成证据。
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
The name/description match the instructions: the skill enforces goal/progress/next-step reporting and evidence. It reasonably expects access to agent memory files, workspace artifacts, and the ability to verify URLs/paths. One minor mismatch: the SKILL.md uses the PowerShell 'Select-String' command for verification but does not declare a required shell/platform; that is an operational assumption rather than a functional mismatch.
Instruction Scope
All instructions stay within the task-finishing purpose (emit structured progress, produce evidence, verify artifact existence). The skill directs the agent to read memory/{date}.md, artifact paths, and to verify URLs/paths—this is expected for verification. However, the instructions do not include explicit safeguards or redaction rules for removing sensitive data from evidence, and they assume the agent may run commands or perform network checks to validate artifacts; those behaviors can expose secrets if the agent captures unfiltered command output.
Install Mechanism
Instruction-only skill with no install spec and no code files. Lowest-risk installation surface: nothing is written to disk by the skill itself.
Credentials
The skill requests no environment variables, binaries, or external credentials. It does assume access to the agent's workspace, memory files, ability to run checks (filesystem or HTTP) and to produce artifact paths/URLs; these are proportionate to a task-evidence verifier.
Persistence & Privilege
always=false and no config paths or persistent privileges requested. The skill does not attempt to modify other skills or system-wide settings.
Assessment
This skill appears to do what it says, but before installing or using it: 1) confirm the agent platform provides access to the referenced memory and workspace paths and supports the verification commands (the SKILL.md uses Select-String, which is PowerShell-specific); 2) decide and enforce a policy for redacting sensitive data from 'EVIDENCE' (command outputs, artifact contents, URLs) because the skill asks the agent to surface evidence but does not mandate redaction; 3) restrict the skill to contexts where verifying artifact URLs/paths is safe (avoid running it with access to secrets or production credentials); 4) test the skill on non-sensitive tasks first to confirm it behaves as expected and doesn't leak data.Like a lobster shell, security has layers — review code before you run it.
latest
任务收尾器
强制任务收尾,确保每个任务都有明确的完成证据。防止任务中途停滞、无证据完成。
为什么需要这个技能
问题: AI 代理经常:
- 任务中途停止,无解释
- 只输出计划,不执行
- 缺少明确的完成标准
- 没有证据工件
解决方案: 此技能强制执行:
- 每个实质性步骤输出目标/进度/下一步
- 完成时必须提供证据
- 多步骤任务必须有完成证明格式
工作流程
1. 任务启动
每个任务开始时输出:
**目标**: 完成时是什么样子
**进度**: 已完成什么
**下一步**: 现在执行的一个具体行动
2. 执行中更新
每个实质性步骤后更新进度:
**进度更新**:
- 已完成:步骤 1, 2
- 当前:执行步骤 3
- 阻塞:无(或具体阻塞 + 已尝试 + 最小解锁输入)
3. 完成证明格式
2 步以上的任务必须包含:
**DONE_CHECKLIST**:
- [ ] 项目 1 已完成
- [ ] 项目 2 已完成
**EVIDENCE**:
- 已执行:命令/操作摘要
- 工件:路径/URL/ID
- 已验证:检查命令结果
**NEXT_AUTONOMOUS_STEP**:
- 一个无需用户输入即可执行的后续步骤
4. 反停滞规则
- 仅计划的回复:最多 1 次
- 下一次回复必须包含执行证据
- 永远不要以"我现在将..."结束而不显示工具结果
可执行完成标准
任务完成当且仅当:
| 标准 | 验证 |
|---|---|
| 目标已陈述 | Select-String "目标" memory/{date}.md 匹配 |
| 进度已追踪 | Select-String "进度" memory/{date}.md 匹配 |
| 下一步已定义 | Select-String "下一步" memory/{date}.md 匹配 |
| 证据存在 | 工件路径/URL 存在 |
| 无未解决标记 | Select-String "TODO |
使用示例
示例 1:发布技能到 ClawHub
**目标**: 发布 skill 到 ClawHub,返回 URL 和 skill_id
**进度**:
- 已完成:读取 SKILL.md,打开浏览器
- 当前:填写发布表单
**下一步**: 填写 Slug/Name/Version 字段
---
**进度更新**:
- 已完成:表单填写,文件上传
- 当前:等待发布确认
- 阻塞:无
---
**DONE_CHECKLIST**:
- [x] 技能已发布
- [x] URL 已验证可访问
- [x] skill_id 已记录
**EVIDENCE**:
- 已执行:clawhub publish 命令
- 工件:https://clawhub.ai/Dalomeve/my-skill
- skill_id: k97xxxxx
- 已验证:浏览器导航到 URL,200 OK
**NEXT_AUTONOMOUS_STEP**:
- 更新 INDEX.md 添加新技能条目
示例 2:研究 GitHub 热门项目
**目标**: 研究 12+ GitHub 热门项目,提炼 3-5 个痛点,选择一个方向实现
**进度**:
- 已完成:浏览 GitHub Trending,记录 14 个项目
- 当前:分析痛点,选择方向
**下一步**: 编写研究报告到 outputs/
---
**DONE_CHECKLIST**:
- [x] 14 个项目已分析
- [x] 5 个痛点已识别
- [x] 研究方向已选择(验证空白)
- [x] 新项目已创建并发布
**EVIDENCE**:
- 研究报告:outputs/2026-03-01-github-research.md
- 新项目:https://github.com/Dalomeve/agent-audit-trail
- 测试通过:python test_audit_trail.py (4/4 tests passed)
- Commit: 03307bdc84abab4bd5e78ab51abcd271ca85a4ab
**NEXT_AUTONOMOUS_STEP**:
- 将新项目添加到长期任务追踪
隐私/安全
- 证据中不包含敏感数据
- 工件路径使用相对路径或工作空间路径
- 任务日志中不包含凭据
自触发
当以下情况时使用:
- 启动任何多步骤任务
- 中断后恢复
- 移交给另一个代理
- 用户要求"不要只给计划,直接执行"
与其他技能集成
- task-finish-contract: 英文版本,概念相同
- agent-audit-trail: 使用审计追踪记录证据
- phoenix-loop: 从失败中学习,更新收尾模式
- HEARTBEAT.md: 心跳检查任务收尾状态
限制
- 需要代理自觉遵守
- 证据验证是尽力的(URL 可能过期)
- 不适用于单步骤简单任务
参考
tasks/QUEUE.md- 任务队列格式memory/tasks.md- 任务历史skills/task-finish-contract/- 英文版本
有始有终。用证据证明。
Comments
Loading comments...
