WeChat Article Formatter

Convert Word documents and pasted text into markdown for the "谓无名" WeChat article editor, then render the markdown into paste-ready WeChat HTML.

Requirements: Python 3; Node.js 18+ with the bundled dependencies; LibreOffice/soffice only when converting legacy .doc files.

Deliverables — every job produces three files, all mandatory:

1. <文章名>.md — 谓无名-syntax markdown (the editable source of truth)
2. <文章名>.html — paste-ready inline-styled HTML, rendered from the .md by node scripts/render_html.js (never hand-written)
3. <文章名>.preview.html — browser preview generated by the same renderer for visual QA

Core Rules

• Preserve the uploaded document's body content, wording, punctuation, names, dates, citations, and paragraph order as faithfully as possible. Relocate source author/translator bios into the required post----[note] bio modules without rewriting them.
• Do not rewrite, polish, summarize, expand, translate, or fact-complete the article unless the user explicitly asks.
• Do not search online to fill missing book/person/publication data — exception: keyword blocks and extended reading recommendations always require web search to find accurate person, book, event, and photo information.
• Do not invent placeholders such as https://placeholder, [待确认], or [信息缺失] in the formatted article.
• If an image has no usable hosted URL, do not output standalone image syntax for it.
• Keyword block images are downloaded to a local images/ folder, then uploaded to the user's own image host via python scripts/upload_image.py <图片...>. The default backend is the user's PicGo Server; a user who has configured Cloudflare R2 uploads there instead. Either way the destination and the credentials belong to the user, and this skill stores no image-host credentials of its own. Fill the returned public URL into the syntax block (e.g. [book:https://cdn.example.com/book.jpg|书名|作者|出版社|年份]). If PicGo is unavailable or upload fails, leave the image field empty and report it in 校对提醒.
• For book/reading-book blocks that are already present in the source but lack a cover URL, search + download + upload as above; only if no image can be found, leave the image field empty (e.g. [book:|书名|作者|出版社|年份]).
• If typos, grammar issues, punctuation issues, or political/compliance risks are found, keep the formatted article faithful to the source and report the issues after the markdown under 校对提醒.

编者按（Editor's Note）

• Every article must begin with a 编者按. It is the first non-empty content in the markdown file. Do not place the article title, author line, or any other content above it; use the source title only for the output filename and reporting.
• Syntax format: >>> (note the trailing space) must be on the same line as the content start, and <<< must be on a separate line at the end. The space after >>> is mandatory for the editor to recognize the block.
• Correct:
  >>> 编者按内容... <<<
• Incorrect:
  >>>编者按内容...
• Also incorrect:
  >>> 编者按内容...
• The 编者按 should be approximately 500 characters (Chinese) and must:
  1. Summarize the main content and key arguments of the article.
  2. Provide a brief editorial commentary — highlighting significance, raising questions, or connecting to broader context.
• Write the 编者按 in a tone consistent with the "谓无名" editorial style: thoughtful, measured, and intellectually engaging.
• In the final report, state the 编者按 character count so the user can confirm it is ~500.

Article Structure Order

Every article must follow this order:

1. 编者按 — >>> / <<<, ~500 chars; first non-empty content, with no article title above it
2. 正文 — body text with keyword blocks (books, persons, etc.) inline
3. 目录 (if any) — ---[toc] ... ---[/toc]
4. 注释 (if any) — ---[notes] ... ---[/notes]
5. 来源说明 — ---[note] ... ---[/note] block with:
   - 本次推送内容为《书名》一书的"章节"。
   - 感谢XXX授权转载。
   - 图片源于作者/互联网。
6. 作者简介 — ---[bio-title:作者简介] ... ---[/bio]
7. 译者简介 (if any) — ---[bio-title:译者简介] ... ---[/bio]
8. 延伸阅读 — ---[reading-title:延伸阅读] ... ---[/reading]
9. staff — exactly two [staff:...] entries
10. 固定关注区 — ---[follow] ... ---[/follow]

This order is strict. Do not leave author or translator bios inside the body; move their source wording to the matching bio blocks after ---[note].

No Separator Before Special Blocks

The following blocks use their own opening syntax (e.g. ---[note], ---[bio-title:], etc.) as visual markers. Do NOT add an extra --- horizontal rule line before them:

• ---[toc] (目录)
• ---[notes] (注释)
• ---[note] (来源说明)
• ---[bio-title:作者简介]
• ---[bio-title:译者简介]
• ---[reading-title:延伸阅读]

These blocks should appear directly after the preceding content, without any intervening --- line.

## Heading Rule

• ## section headings must always be followed by ---[dot]. The ## and ---[dot] are bound together as a single unit.
• Every ## heading line must be immediately followed by a separate ---[dot] line.
• ### subheadings do NOT have ---[dot] after them. Only ## headings use ---[dot].

Keyword Blocks and Extended Reading

• Search for key information mentioned in the article: identify important persons, books, events, movies, and TV shows referenced in the text.
• Use the available web search/fetch/browser tools for online research. Search strategy:
• Use web search to discover sources and keyword candidates.
• Open/fetch known URLs to verify exact metadata.
• Use browser automation only when a public page requires JavaScript or interactive navigation.
• Always prioritize first-hand sources (official websites, original documents) over secondary reporting

Selection & Listing Rules（列举规则）

• 人物 [universal:] blocks: pick only substantive persons — those the article discusses, quotes, or builds an argument on. Skip passing name-drops, and skip a person who is already represented by their own book's [book:] block in the same paragraph. Target 3–6 person blocks per article (fewer is fine for short articles); insert each at the paragraph of first mention only, never twice for the same person. The first text field must be 姓名（生年~ 卒年） — one space after the tilde, before the death year; for living persons use 姓名（生年~） with nothing after the tilde. The second field contains identity and contribution only, without repeating the dates.
• 图片语法选择: 人物、电影/电视剧海报、广告使用 [universal:]（统一卡片尺寸）；史料照片、艺术作品、横幅或其他必须保留原始宽高比例的图片使用 [origin:]。不要用 Markdown ** 包裹第一项文字，两种语法都会自动将第一项加粗。
• 书籍 blocks: every book the article substantively discusses gets one block at first mention — [book:] for Chinese editions, [enbook:] for English, [jpbook:] for Japanese. Never duplicate a book.
• 延伸阅读: recommend 2–5 books, thematically tied to the article's subject. Must NOT repeat any book already in the article body. Prefer in-print Chinese editions (豆瓣 has an entry); order from most to least directly related.
• In the final report, list what was chosen and why (one line each), so the user can veto or swap entries.

信息获取渠道（persons / books / images, in priority order)

• 人物信息（identity, dates, 身份简介）:
  1. Wikipedia — match language to the person: 中文人物→zh.wikipedia, 日本人物→ja.wikipedia, 西方人物→en.wikipedia. Cross-read a second language edition when dates/titles matter.
  2. Kotobank (kotobank.jp) — authoritative for Japanese scholars/writers not well covered by Wikipedia.
  3. 机构页面 — university faculty pages, publisher author pages (最权威的在职头衔来源).
  4. 百度百科 / 豆瓣作者页 — fallback for Chinese figures; treat as secondary, verify against one of the above.
• The name field format is 姓名（生年~ 卒年）; use 姓名（生年~） for a living person. The year separator is the tilde ~, never a hyphen or dash (-, –, —), and there is exactly one space after the tilde whenever a death year follows — 舒群（1913~ 89）, not 舒群（1913~89）. A living person's name ends right after the tilde, with no trailing space: 姓名（生年~）. When the death year shares its first two digits with the birth year, drop those two digits: 舒群（1913~ 89）, not 舒群（1913~ 1989）. Write both years in full when the leading digits differ (钱谦益（1582~ 1664）) or when either year has fewer than four digits (李白（701~ 762）). The 简介 field format is 身份头衔，代表作或主要贡献 — factual, ≤ 40 chars, no honorifics, and no repeated dates. Birth/death years and titles must be confirmed by at least one authoritative source; if sources conflict, pick the more authoritative one and note the conflict in 校对提醒. Never invent a missing year.
• 人物照片:
  1. Wikimedia Commons (https://upload.wikimedia.org/wikipedia/commons/...) — check the person's Wikipedia infobox image first.
  2. Kotobank / 机构页 / publisher author page portrait.
  3. Amazon (author page) / 豆瓣作者页 photo.
• Prefer a clear, front-facing, watermark-free portrait; skip group photos.
• 书籍信息与封面:
  1. 豆瓣 (Douban) — for Chinese-edition books, movies, TV, music. After obtaining item_id, construct cover URL:
  • Books: https://dou.img.lithub.cc/book/<id>.jpg
  • Movies: https://dou.img.lithub.cc/movie/<id>.jpg
  • TV shows: https://dou.img.lithub.cc/tv/<id>.jpg
  • Music: https://dou.img.lithub.cc/music/<id>.jpg
    2. 日本亚马逊 (Amazon.co.jp) — for Japanese books. Cover URL patterns: https://images-fe.ssl-images-amazon.com/images/I/<image_id>.jpg or https://m.media-amazon.com/images/I/<image_id>.jpg. Also useful for person photos of Japanese authors.
    3. 出版社官网 / Google Books — for English books and to verify edition metadata (出版社、年份、译者).
• Book metadata (作者/译者/出版社/年份) must match the actual edition whose cover is used; verify on the source page, not from memory.

Image Download + Image-Host Upload

• Download each image to the images/ directory under the same project folder using Bash (curl -L -o or PowerShell Invoke-WebRequest).
• Naming convention (use Chinese name, remove punctuation 《》·： etc., keep alphanumeric and CJK):
• Book covers: book_<书名>.jpg — e.g. book_浮世通鉴日本大众文化史.jpg
• Person photos: person_<姓名>.jpg — e.g. person_鲁迅.jpg
• Movie/TV posters: movie_<片名>.jpg — e.g. movie_千与千寻.jpg
• Extended reading covers: reading_<书名>.jpg — e.g. reading_菊与刀.jpg
• If filename conflicts occur, append _2, _3 etc.
• Upload to the user's image host after all downloads finish, in one batch:
  bash python <SKILL_ROOT>/scripts/upload_image.py images/*.jpg # 每行输出「本地路径<TAB>公网URL」；--json 输出 JSON 映射
• Two backends are available and both upload to the user's own host: picgo (default) and r2 (optional, direct to the user's Cloudflare R2 bucket). The backend is resolved in this order: --backend picgo|r2 → WEIWUMING_IMAGE_BACKEND → the "backend" field in the config file → automatic (r2 when all five R2 keys are present, otherwise picgo). Do not switch backends on the user's behalf; just use whatever resolves.
• For picgo: users must configure and select their own image host in PicGo, then keep PicGo Server running. The default API is http://127.0.0.1:36677/upload; override it with PICGO_API_URL or ~/.weiwuming/image-host.json (see references/image-host.md).
• For r2: users supply account_id, access_key_id, secret_access_key, bucket and domain through ~/.weiwuming/image-host.json or the R2_* environment variables. These are the user's own credentials — never print, echo, or copy them anywhere.
• Run python <SKILL_ROOT>/scripts/upload_image.py --check before the first upload; it checks whichever backend is active. If PicGo Server authentication is enabled, configure PICGO_SERVER_SECRET or picgo.server_secret in the local config file.
• If the active backend is picgo and PicGo is not configured, run python <SKILL_ROOT>/scripts/setup_picgo.py --status. (setup_picgo.py only concerns the PicGo backend; it does nothing for R2.) When PicGo Core CLI 2.0+ is installed and the user chooses PicGo Cloud, offer --login for the official browser OAuth flow. Do not open the browser without the user's confirmation. OAuth login configures PicGo Cloud only; GitHub/S3/OSS/COS and other third-party uploaders still require their own credentials through --configure-uploader or PicGo GUI.
• setup_picgo.py --sync-config is optional and explicit because the first cloud sync may upload the local PicGo configuration and conflicts require interactive resolution. Never run it automatically.
• Remote http(s) URLs passed in are returned as-is. Local-image results are cached in ~/.weiwuming/upload-cache.json, isolated by backend instance and content hash (the older picgo-upload-cache.json is still read as a fallback).
• Cache write failure is only a warning: keep and report the successful URLs instead of retrying and uploading duplicates.
• Exit code 2 = invalid configuration or the image host is unavailable → leave image fields empty and tell the user which backend failed, pointing at references/image-host.md.
• Image field in syntax: fill the returned public URL into the image/cover field:
• [book:<公网URL>|书名|作者|出版社|年份]
• [universal:<公网URL>|姓名|身份/简介]
• [origin:<公网URL>|图说第一行|图说第二行|...]
• [reading-book:书名|作者|译者|<公网URL>]
• If upload failed or no image was found, leave the field empty.
• Search timeout handling: if a search takes too long, times out, or returns no results after reasonable attempts, leave the image URL field empty (no download needed) and report in 校对提醒.

Block Syntax Details

• Insert the appropriate syntax block immediately after the paragraph where the keyword is first mentioned:
• Persons → [universal:<URL>|姓名（生年~ 卒年）|身份/简介]；波浪号后空一格再写卒年（舒群（1913~ 89））。在世人物使用 姓名（生年~），波浪号后不留空格，简介字段不再重复年份。卒年与生年前两位相同时省略前两位（舒群（1913~ 89）），不同世纪或非四位数年份写全（钱谦益（1582~ 1664）、李白（701~ 762））。
• Books → [book:<URL>|书名|作者|出版社|年份] / [enbook:<URL>|Title|Author|Publisher|Year] / [jpbook:<URL>|書名|著者|出版社|年]
  • Field separator rule: each field MUST be separated by |, do NOT use 、 or other delimiters within fields.
  • Book name: do NOT add 《》 or similar punctuation around the book name.
  • Example: [book:http://.../weiwuming/a1b2c3d4.jpg|浮世通鉴：日本大众文化史|日文研项目组 编著，党蓓蓓 译|北京大学出版社|2025]
  • Incorrect: [book:xxx.jpg|《书名》|作者，出版社，年份] (missing separators and extra punctuation)
• Movies/TV posters and advertisements → [universal:<URL>|《片名》|导演：姓名，编剧：姓名|上映：年份，片长：分钟] — search Douban first for movie/TV posters.
  • 标题行: write only the work title (normally 《片名》); do not append 海报.
  • 主创行: director and screenwriter MUST share one field and render on the same line: 导演：...，编剧：.... Use full-width colons and a Chinese comma exactly as shown. Join multiple names within either role with /; retain 等 when the source or verified credit is intentionally non-exhaustive. Do not substitute cast/voice actors for the screenwriter. If the director also wrote the screenplay, repeat the name in both roles.
  • 电影时间行: 上映 and 片长 MUST share one field so they render on the same line: 上映：YYYY，片长：NN 分钟.
  • 电视剧时间行: use 首播：YYYY，单集片长：NN 分钟，共 NN 集 (omit the episode count only when it cannot be verified).
  • 默认不写简介: unless the user explicitly requests a synopsis or contextual note, end the card after the time line. When explicitly requested, add it as a following |-separated field (normally 说明：...); never append it to the 上映/首播 + 片长 field.
  • Example: [universal:https://.../poster.jpg|《新世纪福音战士》|导演：庵野秀明/鹤卷和哉等，编剧：庵野秀明/榎户洋司等|首播：1995.10.4，单集片长：24 分钟，共 26 集]
• Events → [universal:|事件名称|简要说明]
• Archival photos and artworks that must preserve their intrinsic proportions → [origin:<URL>|图说第一行|图说第二行|...]
• In both [universal:] and [origin:], the first text field is automatically bold. Write plain text only; never add ** around it.
• Extended reading format: [reading-book:书名|作者信息|译者信息|封面URL] — fields MUST be separated by |; last field = uploaded cover URL (empty if unavailable).
• Each reading entry MUST be separated by a blank line. No two [reading-book:...] lines should be adjacent without an empty line between them.
• 作者字段 — 只写姓名，不写"著"。著 是默认情形，一律省略。多位作者用半角 / 分隔，不用 、 或 ，：[美]彼得·海斯勒/何伟。
• 编著加"编"：编者、编著、主编的书，在姓名后空一格补 编，例如 [日]日文研项目组 编、钱理群 编。这是唯一保留的著作方式标记。
• 国籍前缀只出现在译本上，标注跨语言关系，方括号与姓名之间不留空格。引用原版（未经翻译）时作者不加国籍。译本情形下作者与译者各按本人国籍标注，中国人一律省略：
  • 中译本 → 作者 [日]实藤惠秀；译者是中国人，省略国籍，写 谭汝谦、林启彦 译
  • 外译外（非中译本）→ 作者与译者都不是中国人，两边都标：作者 [英]Barak Kushner，译者 [日]井形彬 译
  • 原版 → 四方田犬彦，不加国籍
• 译者字段：译者名 译，姓名与 译 之间空一格。多位译者用 、 分隔：谭汝谦、林启彦 译。非中国籍译者按上条加 [国籍] 前缀。引用原版时此字段留空。
• 人名写法：中译本用中译本署名的中文译名（[美]约翰·W·道尔、[日]实藤惠秀）；其余情形（外译外、原版）一律用本人原名，不用所引版本的音译或转写 —— 上例日译本封面署名是片假名「バラク・クシュナー」，条目仍写 [英]Barak Kushner。
• 完整示例：
  • 中文译本 → [reading-book:中国人留学日本史|[日]实藤惠秀|谭汝谦、林启彦 译|http://.../weiwuming/xxxx.jpg]
  • 中文原著 → [reading-book:韩国现代政治史|咸在凤||]
  • 中文编著 → [reading-book:浮世通鉴：日本大众文化史|[日]日文研项目组 编|党蓓蓓 译|http://.../weiwuming/xxxx.jpg]
  • 外译外（英文书的日译本）→ [reading-book:思想戦：大日本帝国のプロパガンダ|[英]Barak Kushner|[日]井形彬 译|http://.../weiwuming/xxxx.jpg]
  • 外文原版（无中译本）→ [reading-book:日本映画は信頼できるか|四方田犬彦||]
• 延伸阅读不写出版社：一律使用 [reading-book:书名|作者|译者|封面URL]。[reading-jpbook:] / [reading-enbook:] 的第三字段是出版社，与本栏体例不符，不要在延伸阅读中使用（它们保留给正文内联书目）。
• 未有中译本的外文书：书名保留原文（日文写日文、英文写英文），译者字段留空。引用原版时作者不加 [国籍]，例：[reading-book:日本映画は信頼できるか|四方田犬彦||]；引用它在第三国的译本时按上条为作者与译者两侧都标注国籍。
• All searched data must be factual. If uncertain, report in 校对提醒.

Staff and Fixed Footer

• The last section of every article must contain two [staff:...] entries:
  1. [staff:作者姓名|编辑] — filled with the actual author name(s) from the source.
  2. [staff:春生、|审校] — the 审校 (reviewer) always defaults to 春生、.
• After the staff entries, every article must include the fixed follow section:
  ```
  ---[follow]

东亚视角 全球视野

寻找东亚论述的"虫洞"与"黑洞"

点击下图关注"谓无名"

---[/follow]
```
- This footer block is mandatory for all articles, regardless of whether the source contains it.

Workflow

1. Detect input type.
   - For .docx, run python scripts/extract_docx.py <file.docx>.
   - For legacy .doc, use the same script. It detects the OLE container and tries LibreOffice conversion to .docx before extraction.
   - For pasted text, analyze the text directly.

2. Analyze structure.
   - Preserve paragraph order from the extraction JSON content array.
   - Use style hints, heading levels, alignment, bold runs, tables, footnotes, endnotes, and image relationships only as formatting clues.
   - Treat tables as source content. Convert simple tables to readable text blocks; if a table is structurally important, preserve rows in markdown table form.
   - MinerU PDF sources — recover footnotes before declaring them lost: full.md drops footnote contents, but the sibling *_content_list.json keeps them as page_footnote blocks (in page order; ref_text blocks hold the bibliography). Collect them, separate the author-affiliation ∗ note, and restore lost entry numbers from page order. See references/syntax_rules.md → Notes.

3. Apply editor syntax. Read references/syntax_rules.md when exact syntax is needed.
   - Do not emit a main article title or author line above the article. Use the source title for <文章名>.md / .html filenames only.
   - 编者按: write ~500-char editorial summary/commentary with >>> (trailing space mandatory) on the same line as content start. It must be the first non-empty markdown content.
   - Section heading: ## 标题 followed by ---[dot] (always bound together).
   - Blockquote: > 原文.
   - Notes: ---[notes] ... ---[/notes]; note entries are single lines in ^[N 内容] format (Arabic N; the editor auto-converts to circled ①–㊿). Inline references are ^N — never raw HTML <sup>N</sup>. Details and pitfalls in references/syntax_rules.md → Notes.
   - Source note: ---[note] ... ---[/note]. No extra --- line before this block.
   - Bio: after ---[note], place ---[bio-title:作者简介], [bio:原文], ---[/bio], followed by translator bio when present. No extra --- line before these blocks.
   - Keyword blocks: apply the Selection & Listing Rules above; search with the available web tools, download to images/, upload to the user's image host via scripts/upload_image.py, and insert [book:<URL>|...], [universal:<URL>|...], or [origin:<URL>|...] blocks at first mention. The first text field of both image syntaxes is plain text and auto-bold.
   - Extended reading: 2–5 non-duplicate related books, ---[reading-title:延伸阅读] ... ---[/reading] above staff. No extra --- line before this block.
   - Staff credit: [staff:作者姓名|编辑] + [staff:春生、|审校].
   - Follow section: the fixed ---[follow] ... ---[/follow] footer is mandatory.

4. Handle images.
   - Extracted Word image relationships are local package targets, not hosted URLs — never emit them as article image URLs.
   - If the source text already contains a hosted image URL, download it to images/ (it will pass through upload unchanged if already public).
   - Follow Image Download + Upload above; fill returned URLs into syntax blocks.
   - At the end, report all downloaded images with their filenames, uploaded URLs, and corresponding keywords.

5. Save the markdown deliverable as <文章名>.md in the project folder.

6. Render the HTML deliverable (mandatory):
   bash node <SKILL_ROOT>/scripts/render_html.js <文章名>.md --preview
   - Produces <文章名>.html (WeChat paste-ready, inline-styled — identical to the editor's 复制 output) and <文章名>.preview.html (browser QA page with a 复制排版结果 button).
   - The renderer loads the web editor and runs its own preprocessMarkdown → marked.parse → postprocessHtml → applyThemeToPreview → generateInlineStyledHtml pipeline. Never hand-write or post-edit the .html; fix the markdown or editor and re-render.
   - Editor lookup order: --editor arg → env WEIWUMING_EDITOR_HTML → ~/.weiwuming/render.json (editor_html) → live F:\py\tools.weiwuming.cn\index.html → bundled assets/editor-index.html snapshot → legacy F:\py\编辑器最终版\index.html. The bundled root index.html plus vendor/marked.min.js is a directly openable copy of the same web editor.
   - For WeChat, open <文章名>.preview.html and click 复制排版结果; do not copy raw HTML source or rely on selecting part of the preview page. The button writes the exact generated text/html payload, with a selection-copy fallback for restricted file:// clipboard environments.
   - Exit code must be 0. The renderer validates: the editor note is first, article modules are in strict order, all <img> are public http(s) URLs, every syntax block rendered, no unparsed syntax residue (e.g. ---[, <<<), and >>> has its space. If it exits 1, fix the .md and re-render — do not deliver a failing pair.
   - Re-render after ANY edit to the .md, so the two deliverables never diverge. Rendering is idempotent: same .md → byte-identical .html.

7. Validate before delivery.
   - render_html.js exited 0 (this covers image URLs, block pairing, and syntax residue mechanically).
   - No placeholder URLs or invented missing-data markers are present.
   - Every filled image field is a real public URL returned by the user's image host through upload_image.py (never a local path, never an invented URL); fields without a successful upload are empty.
   - All downloaded images exist in images/ with correct naming.
   - Syntax markers are paired: >>>/<<<, ---[notes]/---[/notes], ---[note]/---[/note], ---[reading-title:]/---[/reading], ---[bio-title:]/---[/bio].
   - [universal:] and [origin:] contain at least URL + first line + second line; their first text field contains no Markdown ** because the editor auto-bolds it. Use [origin:] only when the intrinsic image proportions must be preserved.
   - Every person [universal:] card puts verified birth/death years immediately after the name; the description does not repeat them.
   - Every movie/TV [universal:] card uses the bare work title (no 海报 suffix), lists both director and screenwriter on the second rendered line, and has a dedicated time field containing both release/premiere time and runtime. It omits synopsis/context by default; explicitly requested notes use a later | field.
   - No extra --- horizontal rule before ---[note], ---[bio-title:], ---[reading-title:], ---[toc], or ---[notes] blocks.
   - The article follows the exact structure order: 编者按 → 正文 → 目录(如有) → 注释(如有) → ---[note] → 作者简介 → 译者简介(如有) → 延伸阅读 → 两条 staff → 固定关注区.
   - Every ## heading is followed by ---[dot].
   - Notes/footnote format checks: (1) inline ^N reference count == ^[N …] entry count in ---[notes]; (2) in the rendered HTML, footnote-item divs and footnote-num spans both equal that count, with no gaps in the circled-number sequence; (3) zero <sup> tags and zero bare ^数字 left in the HTML. If the source had footnotes, every marker must have its recovered entry — no "footnotes lost" report while page_footnote blocks still exist in the MinerU JSON.
   - Punctuation normalization checks (OCR sources): no half-width , ; : followed by a space inside Chinese sentences (should be ，；：); no smart quotes U+201C/U+201D in Chinese context (unify to the article's 「」 style, keep English quotes in English citations); no unclosed 「 without 」; no empty parentheses （ ） from OCR-dropped years/terms (restore only verifiable values, otherwise report in 校对提醒). When doing exact-string replacements, diagnose actual char codes first (charCodeAt) — curly vs straight quotes look identical in output.
   - Person blocks are substantive and unique (target 3–6; fewer is fine for short articles); reading list has 2–5 items and no duplicates with body books.
   - 延伸阅读作者/译者字段格式：作者字段无 著；[国籍] 与姓名之间无空格；引用原版与中国籍人士无国籍前缀，外译外的作者与译者两侧都有；多作者用 / 而非 、；编著保留 编；译者字段形如 姓名 译（多人用 、 分隔），引用原版时留空。
   - Open <文章名>.preview.html or capture a browser screenshot. Check the opening, universal fixed-size cards, origin intrinsic-proportion images, card covers, long poetry/table layouts, and footer; fix visible blank/broken cards or spacing errors and re-render. Use its copy button for the final WeChat paste test.
   - Suspected typos, grammar, punctuation, and political/compliance issues are listed separately, not silently corrected.

Reporting Format

Return:

交付文件：
- 📄 `<文章名>.md` — 谓无名语法 markdown
- 📄 `<文章名>.html` — 微信粘贴版（render_html.js 校验通过 ✓）
- 👀 `<文章名>.preview.html` — 浏览器预览

编者按：约 NNN 字。

人物/书目选择（可替换，告诉我即可）：
- [universal:] 鲁迅 — 正文第二节重点讨论其杂文观 → 照片源：Wikimedia Commons
- [book:] 浮世通鉴 — 文章主推书目 → 封面源：豆瓣
- 延伸阅读：拥抱战败（同主题战后史）、菊与刀（经典参照）…

图片处理结果：
- ✅ `images/book_书名.jpg` → 来源：豆瓣 → 已上传 `http://.../weiwuming/<hash>.jpg` → [book:<URL>|书名|...]
- ✅ `images/person_姓名.jpg` → 来源：Wikimedia Commons → 已上传 → [universal:<URL>|姓名|...]
- ⚠️ `images/book_书名.jpg` → 上传失败，图片字段留空
- ❌ <关键词> → 下载失败，未找到可用图片源

校对提醒：
- 未发现明显问题。

If issues exist, list them with source snippets and a concise reason:

校对提醒：
- 可能错别字："..."，建议核对是否应为 "..."。
- 标点疑点："..."，中英文标点混用，建议人工确认。
- 政治/合规风险："..."，建议人工复核表述是否符合发布要求。
- 人物信息存疑：<姓名> 生卒年在 Wikipedia 与 Kotobank 记载不一致，已采用 Kotobank。
- 图片下载失败：<关键词>，未找到可用图片源。
- 图床上传失败：`images/xxx.jpg`，语法块图片字段已留空；请确认当前图床后端可用后重跑 `scripts/upload_image.py`。
- 图床不可用：`upload_image.py` 退出码 2，请运行 `--check` 并检查 `~/.weiwuming/image-host.json`；PicGo 后端如尚未配置，可先运行 `setup_picgo.py --status`，PicGo Cloud 用户可在确认后运行 `--login`。

Word Extraction Notes

The bundled extractor adopts the minimax-docx approach:

• Detects file signature instead of trusting extensions.
• Treats .docx as an OOXML ZIP package.
• Converts .doc to .docx through LibreOffice when available.
• Uses Python standard-library XML parsing instead of python-docx.
• Extracts paragraphs, style names, alignment, run bold/italic/size hints, tables, footnotes/endnotes, and image relationships.

If .doc conversion fails or LibreOffice is unavailable, ask the user for a clean .docx export.