skill-governance

OpenClaw Cognitive Operating & Skill Governance Kernel

Security Scan

The skill is an instruction-only governance kernel that mostly matches its description, but its runtime rules reference filesystem paths, automatic mounting/unmounting of other skill bundles, and externally-synchronized summaries without specifying endpoints — these ambiguities could enable unintended cross-skill actions or data exfiltration.

Purpose & Capability

The name/description as a 'cognitive operating & skill governance kernel' matches the SKILL.md: it defines perception, decision, mounting, lifecycle, and closure protocols. There are no unexpected environment variables, binaries, or installs requested that contradict its stated purpose.

Instruction Scope

The instructions mandate behaviors that involve system state and other skills: automatic 'mount/unmount' of bundles, writing forced archives to /memory/YYYY-MM-DD-task.md, moving skills to archived_skills/, and generating summaries for 'external synchronization'. The SKILL.md does not specify where /memory or archived_skills/ live, what API/endpoints should be used for external sync, or what authorization is needed. Those gaps create a risk that the agent will read/write files or transmit data outside expected boundaries or trigger other skills unexpectedly.

Install Mechanism

No install spec and no code files — instruction-only — so nothing is downloaded or written at install time. This is lower risk from an install-mechanism perspective.

Credentials

The skill requests no environment variables or credentials (proportionate). However, it references filesystem locations and lifecycle operations that imply write/delete privileges over skill storage areas even though no config paths were declared in the manifest; that mismatch should be clarified.

Persistence & Privilege

always:false and model invocation allowed (normal). But the protocol includes lifecycle actions that move and mark other skills (archived_skills/, deletion candidates) and requires sending notifications before deletion. Those are operations that modify other skills' state or system-wide skill storage; the skill does not declare these config paths or required permissions, which is a privilege/footprint mismatch and a potential control risk.

What to consider before installing

This skill is conceptually coherent with a governance kernel, but it contains ambiguous operational directives that could change files, archive or delete other skills, and push summaries externally without specifying where. Before installing: 1) Ask the author to clarify exact filesystem paths and required permissions for /memory and archived_skills/, and which service/endpoints (and credentials) are used for 'external synchronization'. 2) Confirm which bundles the skill may mount/unmount and get explicit allow-listing for those bundles. 3) Run the skill in a restricted sandbox with monitoring to observe file writes and outgoing connections. 4) If you cannot get clear answers, avoid enabling autonomous invocation or limit its scope and privileges.

🧠 OpenClaw 认知与技能治理内核 v2.1

本协议不只是管理技能，而是管理认知资源、上下文负载与交付闭环。

---

0️⃣ GLOBAL PRIME DIRECTIVE

系统优化目标： [ Maximize ; Output / (Noise \times Cognitive\ Load) ] 任何行为若增加噪音或认知负载而不增加输出，必须拒绝。

---

1️⃣ 感知层协议 (Perception Protocol)

1.1 强制脱水规则 (Mandatory Dehydration)

若 Raw_Data > 500 条 或 > 2000 tokens： 必须先执行：

• 低成本摘要（Flash）
• 输出仅包含：
  • Top 高频主题 (≤10)
  • 异动信号 (≤5)
  • 潜在机会 (≤3) 禁止：
• 将全量原始数据直接传入主模型

1.2 信号阈值规则

只有满足以下任一条件，才允许进入分析阶段：

• 增长率 > 30%
• 偏离历史均值 > 2σ
• 直接影响现金流或决策 否则自动丢弃。

---

2️⃣ 决策层协议 (Cognitive Budget Protocol)

2.1 每日启动校准

每日首次任务前，必须输入：

• Sleep (1-10)
• Focus (1-10)
• Stability (1-10) 计算： [ Capacity = (S + F + St) / 3 ]

2.2 权限锁定规则

Capacity ≤ 4 → 禁止：

• 架构重构
• 多技能联动 (>2)
• 高复杂调研 Capacity 5–7 → 允许分析类任务 Capacity ≥ 8 → 允许战略级任务

---

3️⃣ 动态挂载内核 (Contextual Mounting Kernel)

3.1 场景触发器

检测关键词：

• 调研 / Research → research.bundle
• 复盘 / Review → analytics.bundle
• 部署 / Deploy → automation.bundle
• 谈判 / Strategy → decision.bundle 自动执行： mount bundle 任务结束： unmount bundle

3.2 认知负载上限

同时挂载技能 ≤ 7 超过立即警告： ⚠ Cognitive Overload Risk

---

4️⃣ 执行闭环协议 (Mandatory Closure Protocol)

⚠ 这是强制规则。 任务结束前必须满足：

4.1 产出验证

至少生成以下之一：

• 文件
• 结构化笔记
• 决策记录 否则不允许标记完成。

4.2 自动归档

强制生成： /memory/YYYY-MM-DD-task.md 结构固定：

决策

核心数据

下一步行动

置信度

4.3 高级任务推送规则

若任务类型为：

• 财务
• 战略
• 重大决策 必须生成摘要用于外部同步（笔记或移动端）。 禁止停留在本地缓存。

---

5️⃣ 生命周期与淘汰 (Evolution Engine)

5.1 30 天未调用 → 移入 archived_skills/

5.2 60 天仍未恢复 → 标记为删除候选

5.3 删除前必须发送通知

⚠ 24 小时缓冲期 支持 restore <skill_name>

---

6️⃣ 故障熔断规则

单任务中： Skill 报错 ≥ 3 次 → 立即停止 → 标记为 Unstable → 禁止盲重试

---

7️⃣ 输出契约 (Output Contract)

每次技能调用输出必须包含： 🔍 一句话结论 📊 关键数据 ⚠ 置信度说明

---

使用说明

仅在以下情况加载本协议：

• 引入新技能
• 重构技能体系
• 大规模调研
• 多技能联动规划
• 高级战略任务

执行顺序：

1. 检查认知预算
2. 决定是否允许任务
3. 动态挂载所需技能
4. 执行脱水
5. 执行任务
6. 强制生成归档