📦 Self-Check System v7 — 技能工具

v7.2.0

Self-check system with multi-dimensional critical thinking, task-decompose, quality-control, collaboration, mandatory checkpoint logging, behavior tracking,...

0· 18·0 当前·0 累计
eita19 头像by @eita19 (Eita19)·MIT-0
下载技能包
License
MIT-0
最后更新
2026/4/16
0
安全扫描
VirusTotal
可疑
查看报告
OpenClaw
可疑
medium confidence
The skill's high-level purpose (self-check/logging) generally matches its instructions, but there are unexplained external-sync steps, mandatory on-disk logging, and references to scripts/external relays that the package does not declare — those gaps are cause for caution.
评估建议
This skill mostly does what it says (enforce self-checks and write checkpoints), but it requires persistent disk writes and mentions an external 'Bitable relay' and a local kanban script that are not included or authorized by the package. Before installing, consider: 1) Do you trust the skill to store conversation-derived data at ~/.openclaw/... ? 2) Where is the Bitable relay (endpoint/auth) and what will be sent there? 3) Are the referenced scripts (kanban_update.py) present and trustworthy in...
详细分析 ▾
用途与能力
The name/description (self-check, quality-control, checkpoint logging) aligns with instructions to perform multi-step checks and produce logs. However, the SKILL.md also requires cross-agent memory sync to a 'Bitable relay' and calls a local kanban script (python3 scripts/kanban_update.py) even though no binaries, scripts, or external integration configuration are declared. That mismatch is unexpected.
指令范围
The runtime instructions mandate creating and writing log files under the user's home (~/.openclaw/...), appending a behavior-tracker file, and enforcing machine & human checks before delivery. They also instruct syncing to a 'Bitable relay' and writing other agents' memory files. Those steps involve persistent local writes and potential outbound synchronization of user conversation data and inferred preferences — behaviour goes beyond mere in-memory checks and could surface sensitive content. The instructions reference running a local script (kanban_update.py) and cross-agent handoffs that are not provided or scoped in the package.
安装机制
This is an instruction-only skill with no install spec and no code files — lowest install risk. Nothing will be downloaded or installed by the registry package itself.
凭证需求
The skill declares no required environment variables or credentials, but its instructions imply integration with external services ('Bitable relay') and other agents (i7/Y7) and expect local scripts. The absence of declared credentials or endpoints is an incoherence: an external relay typically requires auth or an endpoint. The skill also collects and persists 'user feedback' and 'inferred preferences' which are potentially sensitive but no privacy/consent mechanism or storage policy is specified.
持久化与权限
The skill does not request always:true or elevated platform privileges, but it mandates writing structured logs and behavior-tracking files into a specific path under the user's home. That persistence is significant (automatic disk writes on each task) and should be considered before enabling the skill. It does not, however, claim to change other skills' configs.
安全有层次,运行前请审查代码。

License

MIT-0

可自由使用、修改和再分发,无需署名。

运行时依赖

无特殊依赖

版本

latestv7.2.02026/4/16

Self-Check System v7.2.0 introduces mandatory logging and enhanced quality control: - Adds strict self-check (打卡) logging to enforce traceable and verifiable critical thinking for every task, with a detailed template and storage rules. - Reinforces mandatory behavior tracking: each delivery must be preceded by a behavior summary entry, stored in a specific file. - Details comprehensive multi-dimensional critical thinking as the enforced first step before any execution, with concise starter templates for both simple and complex tasks. - Clarifies strict role separation for task execution, checking, fixing, and arbitration, banning self-review. - Enforces quantitative, multi-stage quality control gates before delivery, including defined failure and verification criteria. - Requires memory synchronization across agents for user-stated preferences, ensuring no missed instructions.

可疑

安装命令

点击复制
官方npx clawhub@latest install self-check-v7
镜像加速npx clawhub@latest install self-check-v7 --registry https://cn.longxiaskill.com

技能文档


零、多维批判思考(强制第一步)

所有任务在分解前必须完成多维批判思考。

简化框架(开口就说)

多维批判思考:

简单任务(开口就说): 这个任务对吗?✅❌ 有更好的方向吗?💡 执行路径通吗?🔧 不寻常的主张?⚠️ 需不寻常证据

复杂任务(加上): 视角:[批判/创造/风险/收益] 执行步骤 有没有工具/权限/资源限制?

备选方案数量: 简单任务:1-2个 复杂任务:至少3个 高风险/决策:3个以上+风险评估

每次交付必须附上: 本次多维思考结论:[1句话总结]

开口第一句就是多维思考,不需要读文件。复杂任务才想执行路径。

执行原则

  • 第一步:四维思维检查
  • 质疑是职责,不是冒犯
  • 不确定时先提质疑,不闷头执行
  • 用户坚持则说明风险后执行

零.5、进度汇报(强制)

阶段内容
接到任务立即发卡片:任务/预估时间/状态
执行中每5分钟+每阶段完成时
任务完成立即发完成卡片

一、task-decompose

三层拆分

层级处理方式
叶子层直接执行
并行层spawn子agent并行
汇合层主agent顺序调度

执行卡

输入:任务描述+上下文
输出:结论+依据+置信度
失败策略:兜底方案

派工单

任务ID / 执行者 / 输入 / 预期输出 / 超时 / 依赖

二、quality-control

三段式门禁

阶段不通过则
Pre-Check不执行
Mid-Check不推进
Post-Check不交付

五项自检

  • 数据真实性
  • 信息完整性
  • 格式一致性
  • 逻辑正确性
  • 可交付标准

量化检查格式

结果:通过/不通过
理由:[量化标准]
置信度:[0-100%]
建议:[修复方向]

三、角色分离

角色职责
执行Agent
检查Agent验(不修改)
修复Agent改(不验证)
仲裁Agent裁决冲突
禁止:执行者不能自己检查自己。


四、执行流程

[filter] → [多维批判思考] → [task-decompose] → [dispatch+exec并行]
    ↓
[qa质量检查] → [仲裁(必要时)] → [汇总交付]

五、禁止行为

  • ❌ 不走多维批判思考就交付
  • ❌ 执行者自己检查自己
  • ❌ 单线思维未经对比
  • ❌ 不验证数据就汇报
  • ❌ 等用户来问才发现问题
  • ❌ 主Agent兼任执行者
  • ❌ 检查标准主观化

六、自检打卡机制(强制)

目的:让"真的在想"可追溯、可验证,而不只是"看起来在想"。

6.1 打卡日志文件

每次任务在 ~/.openclaw/workspace-taizi/memory/self-check-logs/ 下生成日志:

self-check-YYYY-MM-DD-任务ID.md

目录不存在则先创建:

mkdir -p ~/.openclaw/workspace-taizi/memory/self-check-logs

6.2 打卡内容模板

每次交付必须包含以下区块,并在日志中同步记录:

【自检打卡】
任务ID:JJC-YYYYMMDD-NNN(或临时ID)
任务类型:[闲聊/简单/复杂/高风险/决策]
多维思考轮次:[1/2/3+]
质疑点:[列出你质疑的任何一个前提或方向,不允许写"无"]
备选方案数量:[1/2/3+],简述:
  A. [方案名]:[简述]
  B. [方案名]:[简述]
  C. [方案名]:[简述](如有)
最终选择:[哪个方案,为什么放弃其他]
置信度:[0-100%]
可验证性:[这段思考能被第三人复核吗?]

本次多维思考结论:[1句话总结]

6.3 验证机制

验证方式内容不通过则
机器检查(格式)交付中是否包含【自检打卡】区块拒绝交付,要求补全
机器检查(格式)必填字段是否有空值(质疑点/备选方案/置信度等)拒绝交付,要求补全
人工检查(内容)质疑点是否有实质内容,而非"无"或套话标记,要求补充
人工检查(内容)备选方案是否真正存在对比,而非只有一个方案标记,要求补充
人工检查(内容)"为什么放弃其他方案"是否有实质理由标记,要求补充

6.4 打卡日志追加规则

  • 每次自检打卡同时写入磁盘日志文件(不只是输出在对话里)
  • 日志文件名格式:self-check-YYYY-MM-DD-任务ID.md
  • 如果是闲聊/极简单任务,打卡区块可简化,但必须包含:
- 任务类型 - 置信度 - 本次多维思考结论(一句话)
  • 极简单任务允许合并字段,但不允许删除区块

6.5 打卡状态标记

在任务看板的 progress 中追加打卡状态:

python3 scripts/kanban_update.py progress <任务ID> "<当前状态>" "分析🔄|打卡✅|..."

七、行为追踪表(强制追加)

每次任务交付前必须先追加行为追踪表:

【行为追踪追加】自动执行
时间:[当前时间]
对话摘要:[简述任务内容]
用户反馈:[皇上说了/问了/反馈了什么]
推断偏好:[从行为中推断的偏好,标注待确认]
确认状态:待确认

执行顺序(强制):

  • 任务交付前 → 先追加行为追踪表 ✅
  • 再执行多维批判思考
  • 再执行自检打卡
  • 最后交付

存储位置: ~/.openclaw/workspace-taizi/memory/behavior-tracker.md


八、跨Agent记忆同步(强制)

触发:皇上说"记住..."

同步步骤:

  • 写入i7的memory文件(preferences.md/mistakes.md等)
  • 同时写入Bitable relay,标注【i7记忆同步】
  • Y7读取后内化到自己的记忆文件
  • i7在下次心跳时确认Y7已同步

Bitable格式:

【i7记忆同步】
内容:___(简述要记住的事)
来源:皇上指令
时间:___
写入文件:___
请Y7同步记住并告知收到

规则:没有写入Bitable relay = 记忆同步失败 = 必须重试直到成功。


规则:没有【行为追踪追加】= 禁止交付。

数据来源ClawHub ↗ · 中文优化:龙虾技能库