Audit
Audit paper data quality in the knowledge base and take over repairs. Rule-based checks surface missing fields, filename issues, DOI duplicates, and title mismatches with per-issue fix hints; agent/subagents then verify against the original text and fix meta.json directly. Use when the user wants to check data quality, find problems, or fix metadata issues.From its SKILL.md
npx -y skills add wszqkzqk/scrinium --skill auditAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 10 stars10 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its file declares
Copied from the file, not written here
The file declares its own license as GPL-3.0-or-later. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
4.1 KB, as published. Nobody here has run it
论文审计
检查已入库论文的数据质量,并接管修复——audit 不止于诊断:规则检查输出的每条问题都带修复建议(hint),agent 派 subagent 读原文核对后直接修改 meta.json 闭环。
阶段一:规则化检查
scrinium audit [--severity error|warning|info]
问题按严重程度分类,每条附 hint: 修复建议:
- 错误:缺少标题、缺少 MD 文件、JSON 解析失败、DOI 重复
- 警告:缺少 DOI/摘要/年份/作者/期刊、MD 过短、标题不一致、文件名年份不匹配
- 提示:文件名不符合规范格式
阶段二:subagent 核对(agent 深度诊断)
对需要理解才能判定的问题(title_mismatch、可疑元数据、DOI 重复),派 subagent 逐篇核对(多篇并行、互不重叠)。每个 subagent:
- 用 Read 工具读取
data/papers/<paper>/meta.json和paper.md(前 80 行起,必要时读更多) - 判断 MD 正文的实际主题/标题/作者/年份是否与 JSON 元数据一致
- 区分无害差异(MinerU H1 识别问题)vs 真正的内容错配
- 带回「判定 + 正确元数据」的 T1 结论
阶段三:修复
按问题性质分两条路径:
agent 直改 meta.json(推荐,大多数问题):
- 用 Edit 工具直接修正
data/papers/<paper>/meta.json的title/authors/year/journal/abstract等字段 - 适用:标题/作者/年份/期刊错配、摘要缺失或截断(读原文直写
abstract)、 harmless H1 误判后的标题归一
repair 命令(需要 API 补全或目录重命名时):
# 先 dry-run 预览
scrinium repair "<paper-id>" --title "正确标题" [--author "一作"] [--year YYYY] [--doi "10.xxx/..."] --dry-run
# 确认后执行(会查 API 补全元数据、重写 meta.json 并规范化目录名)
scrinium repair "<paper-id>" --title "正确标题" [--author "一作"] [--year YYYY] [--doi "10.xxx/..."] [--no-api]
修复后统一重建索引:
scrinium index
目录名不规范的批量问题转交 /rename skill;DOI 重复参照 pending 工作流由 subagent 对比两篇后决定去留(见 /ingest skill)。
检查规则
| 规则 | 级别 | 说明 |
|---|---|---|
missing_title | error | 缺少标题 |
missing_md | error | JSON 无对应 MD 文件 |
duplicate_doi | error | DOI 重复 |
missing_doi | warning | 缺少 DOI |
missing_abstract | warning | 缺少摘要 |
title_mismatch | warning | JSON 标题与 MD H1 不一致 |
nonstandard_filename | info | 文件名不符合规范格式 |
untagged | info | 未打策展标签(可用 /curate 流程或 scrinium tag 补充) |
示例
用户说:"帮我检查一下论文库有没有问题" → 执行阶段一规则化检查,按 hint 汇总问题清单
用户说:"深度检查并修掉"
→ 完整闭环:阶段一 → 派 subagent 逐篇核对 title_mismatch 等可疑项 → 直改 meta.json / repair → scrinium index → 重跑 audit 确认不再复现
用户说:"修复那些错配的论文" → 执行阶段二 + 阶段三
完成前检查
- 笔记写了吗:subagent 核对确认的真实内容错配(非无害 H1 问题),已通过 CLI 写入对应论文的笔记:
scrinium show "<paper-id>" --append-notes "## YYYY-MM-DD | <任务来源> | audit - 关键发现" - 修复闭环了吗:确认的错配都已直改 meta.json 或
repair,并执行scrinium index,重跑audit不再复现相同问题 - 输出在 workspace 吗:批量诊断结果和修复清单已保存到
workspace/下