都在教写件，没人教验件——第8轮设计模式提炼：给SKILL.md立一道出厂闸

flowchart TD
    A["写件课三拍<br/>出厂之前教写对"] --> B["本期第四拍<br/>出厂之前教验对"]
    B --> C["维修手册 09-26<br/>装坏之后管修理"]
    B --> D["mcp-builder 94分闸<br/>隔壁MCP赛道的先例"]

name: skill-lint
description: Read-only quality gate for a SKILL.md file. Use when the user
  asks to validate, audit, review, or quality-check a skill before
  publishing or installing it, or asks why a skill never triggers.
  Runs mechanical checks via script, walks judgment gates, then issues an
  acceptance report without changing a word. Read-only by contract - it
  issues reports and never edits files. To write a new skill use
  skill-creator instead. To find new skills use find-skills instead.
  Not a replacement for skill-creator or catalog tools.
# Skill Lint — 出厂闸

给任何一件 SKILL.md 做交货前体检：脚本验算术，模型验判断，
只出验收单，不动一个字。

## 七关对照
关1名字（撞户死）→ 脚本验：小写连字符、与目录同名、注册表无同名
关2门牌（叫不应死）→ 脚本验触发短语在场；模型验动词覆盖与让路句
关3预算（烧钱死）→ 脚本验：正文 500 行以内
关4边界（越界死）→ 脚本验负面清单在场；模型验完成定义可验收
关5化石（重蹈死）→ 模型验：踩过的坑是否回写成带后果的红线
关6接缝（接缝死）→ 模型答四问：退出码、输出契约、只读、可重放
关7复活（复活无门死）→ 点火三句灭火三句过堂

## 工作流程
1. 跑 scripts/lint_skill.py <件目录> --registry ~/.claude/skills
2. 退出码 2：件目录读不了，如实报告并停
3. AUTO 挂项逐条抄进验收单，附判据原文
4. JUDGE 各关逐条过堂：过、挂、存疑三选一，存疑注明存疑点
5. 呈验收单：挂项按「先描述后正文最后脚本」给修理顺序

## 不要做
- 不要替用户改一个字——闸只出单，不动刀
- 不要打总分——分数会被应试优化，闸只认过与挂
- 不要把存疑写成过——判断关的存疑就是结论
- 不要为本机没验证过的市场撞名背书——那是上架日的事

## 完成定义
- 七关每关恰有一个结论（过/挂/存疑）
- 每个 AUTO 挂项附判据原文与修理方向
- 本次写盘数为零

#!/usr/bin/env python3
"""lint_skill.py - 只读体检一个 SKILL.md，出算术关体检单（TSV）。
契约：永不写盘；同一输入两次运行输出逐字节一致。
分工：算术关归本脚本，判断关归模型（SKILL.md 里的七关表）。
行格式：关卡 | AUTO或JUDGE | pass/fail/remind | 明细
退出码：0 全过，1 有 AUTO 挂项，2 件读不了。
"""
import argparse, os, re, sys

NAME_RE = re.compile(r"^[a-z0-9][a-z0-9-]*$")
TRIGGERS = ("use when", "when the user", "use this skill",
            "use this", "when asked")
NEGATIVES = ("不要做", "do not", "never ", "禁止")
DESC_LIMIT = 700          # 本件自定阈值，非官方标准
BUDGET_LINES = 500        # 09-27 卷口径
SEAM_ASKS = "四问：退出码有主？输出列固定？默认只读？两次运行逐字节一致？"

def parse_fm(text):
    m = re.match(r"^---\s*\n(.*?)\n---\s*(\n|$)", text, re.S)
    if not m:
        return None, "", text
    fm, body = m.group(1), text[m.end():]
    name, parts, in_desc = None, [], False
    for raw in fm.splitlines():
        line = raw.rstrip()
        if not line:
            continue
        nm = re.match(r"^name:\s*(.+?)\s*$", line)
        dm = re.match(r"^description:(.*)$", line)
        if nm and not in_desc:
            name = nm.group(1).strip().strip("'\"")
        elif dm:
            in_desc = True
            first = dm.group(1).strip()
            if first:
                parts.append(first)
        elif in_desc and line[:1] in (" ", "\t"):
            parts.append(line.strip())
        elif line[:1] not in (" ", "\t", "#"):
            in_desc = False
    return name, " ".join(parts), body

def name_of(skill_dir, cache={}):
    if skill_dir not in cache:
        p = os.path.join(skill_dir, "SKILL.md")
        try:
            with open(p, encoding="utf-8") as f:
                cache[skill_dir] = parse_fm(f.read())[0]
        except OSError:
            cache[skill_dir] = "?"
    return cache[skill_dir]

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("skill_dir")
    ap.add_argument("--registry", help="注册表根目录，供全局撞户查重")
    args = ap.parse_args()
    path = os.path.join(args.skill_dir, "SKILL.md")
    try:
        with open(path, encoding="utf-8") as f:
            text = f.read()
    except OSError as e:
        print(f"读不了 {path}: {e}", file=sys.stderr)
        sys.exit(2)
    name, desc, body = parse_fm(text)
    low = text.lower()
    rows = []
    same = name == os.path.basename(os.path.normpath(args.skill_dir))
    ok1 = bool(name) and bool(NAME_RE.match(name)) and same
    rows.append(("关1名字", "AUTO", "pass" if ok1 else "fail",
                 f"name={name} 目录同名={same}"))
    if args.registry:
        root = os.path.normpath(args.registry)
        me = os.path.basename(os.path.normpath(args.skill_dir))
        hits = sorted(d for d in os.listdir(root) if d != me
                      and name not in (None, "?")
                      and name_of(os.path.join(root, d)) == name)
        rows.append(("关1撞户", "AUTO", "fail" if hits else "pass",
                     f"同名户={','.join(hits) if hits else '无'}"))
    trig = any(t in low for t in TRIGGERS)
    rows.append(("关2门牌", "AUTO",
                 "pass" if desc and trig and len(desc) <= DESC_LIMIT else "fail",
                 f"描述{len(desc)}字符 触发短语={'在场' if trig else '缺席'}"))
    nlines = body.count("\n")
    rows.append(("关3预算", "AUTO", "pass" if nlines <= BUDGET_LINES else "fail",
                 f"正文{nlines}行 上限{BUDGET_LINES}"))
    neg = any(t in low for t in NEGATIVES)
    rows.append(("关4边界", "AUTO", "pass" if neg else "fail",
                 "负面清单在场" if neg else "未见负面清单"))
    rows.append(("关5化石", "JUDGE", "remind",
                 "踩过的坑是否已回写成带后果的红线条款"))
    if os.path.isdir(os.path.join(args.skill_dir, "scripts")):
        rows.append(("关6接缝", "JUDGE", "remind", SEAM_ASKS))
    rows.append(("关7复活", "JUDGE", "remind",
                 "点火三句灭火三句过堂；回炉每轮只拧一层"))
    for row in rows:
        print("\t".join(row))
    hard = sum(1 for r in rows if r[1] == "AUTO" and r[2] == "fail")
    soft = sum(1 for r in rows if r[2] == "remind")
    print(f"# AUTO挂{hard}项 JUDGE待判{soft}项", file=sys.stderr)
    sys.exit(1 if hard else 0)

if __name__ == "__main__":
    main()

flowchart TD
    A["写完一件SKILL.md"] --> B["点亮 skill-lint"]
    B --> C["跑 lint_skill.py"]
    C --> D{"AUTO关有挂项"}
    D -->|有| E["逐条抄单回炉<br/>每轮只拧一层"]
    D -->|无| F["判断关逐条过堂"]
    F --> G{"判断关有存疑"}
    G -->|有| E
    G -->|无| H["呈完整验收单<br/>写盘数为零"]
    E --> B

# 两件套：正文、脚本
mkdir -p ~/.claude/skills/skill-lint/scripts
# 将上文 SKILL.md 存为 ~/.claude/skills/skill-lint/SKILL.md
# 将上文脚本   存为 ~/.claude/skills/skill-lint/scripts/lint_skill.py

# 出厂自测第一式：拿它验它自己（退出码应为 0，JUDGE 三条留模型过堂）
python3 ~/.claude/skills/skill-lint/scripts/lint_skill.py \
  ~/.claude/skills/skill-lint --registry ~/.claude/skills

# 自测第二式：故意验一件没写负面清单的裸件，确认关4如挂
# 自测第三式：点火灭火各三句
#   点火：「我刚写了个 skill，帮我验一下能不能交」「为什么我这件从来不触发」
#   灭火：「帮我写一个新 skill」（归 skill-creator）「帮我找个处理 PDF 的件」（归 find-skills）

# 市场撞名复查（本通道今日未开，命令备上架日用）
npx skills find skill-lint