菜单
Escalate when uncertainty is too high, risk exceeds agent capability, or human judgment is required. Use when verification fails, threats are found, or the agent is stuck.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | metacog-escalate |
| description | Escalate when uncertainty is too high, risk exceeds agent capability, or human judgment is required. Use when verification fails, threats are found, or the agent is stuck. |
| user-invocable | true |
| disable-model-invocation | false |
| allowed-tools | Read, Grep, Glob |
When the agent cannot safely proceed, escalate with structured information.
## Escalation
**Reason**: [why the agent cannot safely proceed]
**Current state**: [what has been done so far]
**What was tried**: [approaches attempted and why they failed]
**Options**:
1. [Option A] - risk: X, effort: Y
2. [Option B] - risk: X, effort: Y
3. [Do nothing] - consequence: Z
**Recommendation**: [which option and why, or "needs human judgment"]
**Information needed**: [what the human could provide to unblock]
Synchronize the Obsidian vault (Meta-YGN/) with the current codebase state. Use after releases, major changes, or when vault docs may be stale.
Run a minimal metacognitive evaluation pass over the current task or repository change. Use to assess quality, calibration, and overhead.
Pressure-test the current plan or conclusion. Use when reasoning may be too linear, confidence seems too high, or there are plausible counter-hypotheses.
Compact the session into goal, verified facts, dead ends, open risks, and next actions. Use before long sessions, handoffs, context limits, or after repeated failed attempts.
Classify the task, identify risk, choose a strategy, and set a proof plan before non-trivial work. Use for architecture, migrations, auth, billing, concurrency, large refactors, or whenever failure would be costly.
Build a proof packet before finalizing work. Use for risky edits, architecture claims, dependency changes, security-sensitive work, and any answer that could mislead a reviewer.