double-check

name	double-check
description	Agent 主动输出 confidence rating 0-5 + 列出可疑点。把 user 频繁说的 "double check 一下" 固化。Use when user says "double check", "再确认", "你确定吗", "confidence", "自查", "reliability".
argument-hint	[claim-or-output]
allowed-tools	Read

/double-check — Self-Confidence Audit

借鉴 user 11K 对话 "double check" 频次 + claude-research 自检 pattern。

Overview

任何 agent 报告完，自查输出：

**Confidence: 4/5**

Suspicious points to verify:
1. Last GPU util reading may be stale (cache 30s)
2. Run v6-meta wandb URL was guessed from naming pattern, not fetched
3. Smoke gate passed but only with bs=2 — real bs=8192 might OOM

Workflow

非可执行 skill——是 output style enforcement。可作为 PostToolUse hook 自动 append。

Agent 自我评：

self.confidence = self.assess_confidence(
    sources_verified=True,
    fresh_data=True,
    edge_cases_covered=False,  # → -1
    has_speculation=True,       # → -1
    user_can_re_verify=True
)
# emit confidence X/5 + list each -1 reason

Composition

• 配合 /conclusion-first 用：5-段报告之后 append confidence
• /pivot 推荐时强制 emit confidence
• /cross-review 判收敛时看每方 confidence