feat(skills): 新增 office-text-extract + markdown-to-docx 技能(从 Hermes 迁移，含脚本，路径改相对)

2026-08-19 12:59:52 +08:00 · 2026-08-19 12:59:52 +08:00 · 972e9307c8
commit 972e9307c8
parent 31c05c8ceb
7 changed files with 29917 additions and 0 deletions
--- a/skills_library/all/markdown-to-docx/SKILL.md
+++ b/skills_library/all/markdown-to-docx/SKILL.md
@ -0,0 +1,95 @@
+---
+name: markdown-to-docx
+description: "Convert Markdown to .docx with CJK fonts, no pandoc."
+version: 1.0.0
+metadata:
+  hermes:
+    tags: [docx, word, markdown, cjk, chinese, conversion, docx-js, node]
+    category: productivity
+    related_skills: [docx, pdf, officecli]
+---
+
+# Markdown → DOCX (no pandoc / LibreOffice)
+
+## When to use
+
+Trigger on: "转成 Word / 导出 docx / Word 版本 / 白皮书 Word 版" — any request to turn a Markdown document (whitepaper, report, spec, README) into a `.docx`, especially on hosts where `pandoc`, `LibreOffice`/`soffice`, or `python-docx` are NOT installed.
+
+The bundled `docx` skill covers creating `.docx` by hand and editing existing files, but assumes pandoc/soffice for reading/verification. When those binaries are missing, this skill is the fallback: parse Markdown with `marked` and emit `.docx` with `docx-js`, fully controlling CJK fonts.
+
+## Prerequisites (no sudo)
+
+```bash
+# node + npm are usually already present; registry reachable via `npm ping`
+mkdir -p /tmp/md2docx && cd /tmp/md2docx && npm init -y >/dev/null
+npm install docx marked adm-zip
+```
+
+`docx` (docx-js) and `marked` generate/parse; `adm-zip` is for verification only.
+
+## The approach
+
+`marked.lexer(md)` returns a flat token stream. Walk it once and map each `type` to a docx-js element:
+
+| marked token | docx-js output |
+|--------------|----------------|
+| `heading` | `Paragraph({ heading: HeadingLevel.HEADING_<depth> })` |
+| `paragraph` | `Paragraph` from inline runs |
+| `code` | single `Paragraph` — one `TextRun` per line joined with `new TextRun({ break: 1 })`, gray `shading` + box `border` |
+| `table` | `Table` — `columnWidths` + per-cell `width` (both DXA), header row shaded + bold |
+| `list` | bullet (`•`) / numbered (`N.`) paragraphs with `indent` + `hanging` |
+| `blockquote` | indented paragraph with left border, italic gray |
+| `hr` | empty paragraph with bottom border (NOT a table) |
+
+A complete, tested converter is shipped at `scripts/convert_md_to_docx.js`. Run:
+
+```bash
+node scripts/convert_md_to_docx.js input.md output.docx [title]
+```
+
+Optional 4th arg sets the Word `title` property (defaults to "Converted from Markdown" — pass the real doc title so the `.docx` metadata isn't a stale leftover). The converter auto-inserts a `PageBreak` before every H1 (after the first) so each `#` chapter starts on a fresh page — right for multi-chapter whitepapers and 投标文件 (bid documents); for continuous-flow docs, delete the `firstH1Seen` block in the main loop.
+
+Inline runs are built recursively from `tok.tokens`, mapping `strong`/`em`/`codespan`/`link` to run formatting. `codespan` (`` `...` ``) renders as red Consolas — critical for documents full of `${param}$`, `sqlExe`, `load_xxx()`.
+
+## CJK font handling (the main gotcha)
+
+Word does NOT reliably auto-fallback Chinese glyphs. Set the **`eastAsia`** attribute on every run's `font`, alongside `ascii`/`hAnsi`:
+
+```js
+const FONT_BODY = { ascii: "Times New Roman", hAnsi: "Times New Roman", eastAsia: "宋体" };
+const FONT_HEAD = { ascii: "Arial", hAnsi: "Arial", eastAsia: "微软雅黑" };
+const FONT_CODE = { ascii: "Consolas", hAnsi: "Consolas", eastAsia: "宋体" };
+// ...new TextRun({ text, font: FONT_BODY })
+```
+
+Convention for Chinese technical docs: 标题 `微软雅黑` navy, 正文 `宋体`, 代码 `Consolas`. Omit `eastAsia` and the whole document renders in a Latin fallback font.
+
+## docx-js footguns (reinforced from the bundled `docx` skill)
+
+- **Tables need dual widths**: `columnWidths` on the `Table` AND `width` on every `TableCell`, both `WidthType.DXA` (PERCENTAGE breaks in Google Docs). Sum to the usable page width (A4 11906 DXA − 1440×2 margins = **9026 DXA**).
+- **Shading type is `ShadingType.CLEAR`, never `SOLID`** (SOLID renders black).
+- **No `\n` in a `TextRun`** — multi-line code blocks use `new TextRun({ break: 1 })` between line runs.
+- **`PageBreak` must live inside a `Paragraph`.**
+- Lists: either use a `numbering` config or prefix the marker manually (`• ` / `N. `); never assume a literal bullet is styled as a list marker.
+
+## Verification without soffice
+
+No LibreOffice available to render → verify by reading the OOXML directly with `adm-zip`:
+
+```js
+const AdmZip = require("adm-zip");
+const z = new AdmZip("out.docx");
+const xml = z.readAsText("word/document.xml");
+(xml.match(/<w:tbl>/g) || []).length        // table count
+(xml.match(/w:fill="F5F5F5"/g) || []).length // code blocks
+(xml.match(/w:fill="D9E2F3"/g) || []).length // header cells
+xml.includes("eastAsia")                      // CJK fonts present
+```
+
+Spot-check key strings (proper nouns, code identifiers) with `xml.includes(...)` to catch content loss. The docx-js `Packer.toBuffer` output is well-formed OOXML — adm-zip reading `word/document.xml` is sufficient structural proof of validity in the absence of a renderer.
+
+## Pitfalls
+
+- **`require('docx/package.json')` throws** `ERR_PACKAGE_PATH_NOT_EXPORTED` — the `exports` map blocks it. Verify install with `typeof require('docx').Document === 'function'` instead.
+- **Don't reach for pandoc/soffice installs** on sudo-less hosts as a first move — npm + docx-js is self-contained and faster to stand up.
+- Keep the `.md` source alongside the `.docx`; regenerate by rerunning the script after any content edit (single source of truth).
--- a/skills_library/all/markdown-to-docx/scripts/convert_md_to_docx.js
+++ b/skills_library/all/markdown-to-docx/scripts/convert_md_to_docx.js
@ -0,0 +1,244 @@
+// convert_md_to_docx.js — Markdown → .docx with CJK font support.
+// No pandoc / LibreOffice / python-docx required.
+//
+// Usage:  node convert_md_to_docx.js input.md output.docx
+// Deps:   npm install docx marked
+//
+// Verified 2026-08-17 on a sudo-less host (A4, 42 headings, 6 tables,
+// 15 code blocks, CJK fonts via eastAsia). Uses marked.lexer() token walk.
+
+const fs = require('fs');
+const {
+  Document, Packer, Paragraph, TextRun, HeadingLevel, Table, TableRow, TableCell,
+  WidthType, AlignmentType, BorderStyle, ShadingType, PageBreak,
+} = require('docx');
+const { marked } = require('marked');
+
+const SRC = process.argv[2];
+const OUT = process.argv[3];
+const TITLE = process.argv[4] || 'Converted from Markdown';
+if (!SRC || !OUT) {
+  console.error('Usage: node convert_md_to_docx.js input.md output.docx [title]');
+  process.exit(1);
+}
+
+// ---- fonts (CJK via eastAsia) ----
+const FONT_BODY = { ascii: "Times New Roman", hAnsi: "Times New Roman", eastAsia: "宋体" };
+const FONT_HEAD = { ascii: "Arial", hAnsi: "Arial", eastAsia: "微软雅黑" };
+const FONT_CODE = { ascii: "Consolas", hAnsi: "Consolas", eastAsia: "宋体" };
+const NAVY = "1F3864";
+const GRAY = "595959";
+const CODE_BG = "F5F5F5";
+const CODE_BORDER = "CCCCCC";
+
+const md = fs.readFileSync(SRC, 'utf8');
+const tokens = marked.lexer(md);
+
+// A4 usable width in DXA (page 11906 - margins 1440*2 = 9026)
+const PAGE_W = 9026;
+
+function cellWidths(n) {
+  const w = Math.floor(PAGE_W / n);
+  const arr = new Array(n).fill(w);
+  arr[n - 1] = PAGE_W - w * (n - 1); // last cell absorbs remainder
+  return arr;
+}
+
+// ---- inline run builder ----
+function inlineRuns(toks, opts = {}) {
+  const out = [];
+  const font = opts.font || FONT_BODY;
+  const size = opts.size || 21; // half-points (10.5pt body)
+  const color = opts.color;
+  (toks || []).forEach(t => {
+    switch (t.type) {
+      case 'text':
+        if (t.text) out.push(new TextRun({ text: t.text, font, size, color }));
+        break;
+      case 'strong':
+        out.push(...inlineRuns(t.tokens, { ...opts, bold: true }));
+        break;
+      case 'em':
+        out.push(...inlineRuns(t.tokens, { ...opts, italics: true }));
+        break;
+      case 'codespan':
+        out.push(new TextRun({ text: t.text, font: FONT_CODE, size: size - 1, color: "C7254E" }));
+        break;
+      case 'del':
+        out.push(...inlineRuns(t.tokens, { ...opts, strike: true }));
+        break;
+      case 'link':
+        out.push(...inlineRuns(t.tokens, { ...opts, color: "0563C1", underline: {} }));
+        break;
+      case 'html':
+      case 'br':
+        break;
+      default:
+        if (t.tokens) out.push(...inlineRuns(t.tokens, opts));
+        else if (t.text) out.push(new TextRun({ text: t.text, font, size, color }));
+    }
+  });
+  if (opts.bold || opts.italics || opts.strike) {
+    out.forEach(r => {
+      if (opts.bold) r.bold = () => true;
+      if (opts.italics) r.italics = () => true;
+      if (opts.strike) r.strike = () => true;
+    });
+  }
+  return out;
+}
+
+function plainText(toks) {
+  let s = '';
+  (toks || []).forEach(t => {
+    if (t.type === 'text' || t.type === 'codespan') s += t.text || '';
+    else if (t.tokens) s += plainText(t.tokens);
+  });
+  return s;
+}
+
+// ---- block handlers ----
+function headingPara(tok) {
+  const lvl = tok.depth;
+  const isTitle = lvl === 1;
+  const size = isTitle ? 36 : lvl === 2 ? 28 : 24;
+  const map = { 1: HeadingLevel.HEADING_1, 2: HeadingLevel.HEADING_2, 3: HeadingLevel.HEADING_3,
+                4: HeadingLevel.HEADING_4, 5: HeadingLevel.HEADING_5, 6: HeadingLevel.HEADING_6 };
+  return new Paragraph({
+    heading: map[lvl],
+    alignment: isTitle ? AlignmentType.CENTER : AlignmentType.LEFT,
+    spacing: { before: isTitle ? 0 : 240, after: isTitle ? 240 : 120 },
+    children: inlineRuns(tok.tokens, { font: FONT_HEAD, size, color: NAVY, bold: true }),
+  });
+}
+
+function paragraphPara(tok) {
+  const children = inlineRuns(tok.tokens || []);
+  if (!children.length) return new Paragraph({ children: [] });
+  return new Paragraph({ spacing: { after: 120, line: 360 }, children });
+}
+
+function codePara(tok) {
+  const lines = (tok.text || '').replace(/\n$/, '').split('\n');
+  const children = [];
+  lines.forEach((line, i) => {
+    if (i > 0) children.push(new TextRun({ break: 1 }));
+    children.push(new TextRun({ text: line, font: FONT_CODE, size: 18 }));
+  });
+  const border = { style: BorderStyle.SINGLE, size: 4, color: CODE_BORDER };
+  return new Paragraph({
+    spacing: { before: 80, after: 120 },
+    shading: { type: ShadingType.CLEAR, fill: CODE_BG },
+    border: { top: border, bottom: border, left: border, right: border },
+    children,
+  });
+}
+
+function tableBlock(tok) {
+  const headerCells = tok.header || [];
+  const ncols = headerCells.length || 1;
+  const widths = cellWidths(ncols);
+  const rows = [];
+
+  const buildRow = (cells, isHeader) => new TableRow({
+    tableHeader: isHeader,
+    children: cells.map((c, i) => new TableCell({
+      width: { size: widths[i], type: WidthType.DXA },
+      shading: isHeader ? { type: ShadingType.CLEAR, fill: "D9E2F3" } : undefined,
+      children: [new Paragraph({
+        spacing: { after: 0 },
+        children: [new TextRun({
+          text: plainText(c.tokens || []), font: FONT_BODY, size: 20,
+          bold: isHeader, color: isHeader ? NAVY : undefined,
+        })],
+      })],
+    })),
+  });
+
+  rows.push(buildRow(headerCells, true));
+  (tok.rows || []).forEach(r => rows.push(buildRow(r, false)));
+
+  const b = { style: BorderStyle.SINGLE, size: 4, color: "BFBFBF" };
+  return new Table({
+    width: { size: PAGE_W, type: WidthType.DXA },
+    columnWidths: widths,
+    borders: { top: b, bottom: b, left: b, right: b,
+               insideHorizontal: b, insideVertical: b },
+    rows,
+  });
+}
+
+function listBlock(tok, ordered) {
+  const paras = [];
+  const walk = (items, depth) => {
+    items.forEach((item, idx) => {
+      const marker = ordered ? `${idx + 1}. ` : '•  ';
+      const indentLeft = 360 + depth * 360;
+      item.tokens.forEach(bt => {
+        if (bt.type === 'list') { walk(bt.items, depth + 1); return; }
+        if (bt.type === 'text' || bt.type === 'paragraph' || bt.type === 'space') {
+          const children = inlineRuns(bt.tokens || []);
+          children.unshift(new TextRun({ text: marker, font: FONT_BODY, size: 21, bold: ordered }));
+          paras.push(new Paragraph({
+            spacing: { after: 40, line: 340 },
+            indent: { left: indentLeft, hanging: 240 },
+            children,
+          }));
+        }
+      });
+    });
+  };
+  walk(tok.items, 0);
+  return paras;
+}
+
+function blockquotePara(tok) {
+  return new Paragraph({
+    spacing: { after: 120 },
+    indent: { left: 360 },
+    border: { left: { style: BorderStyle.SINGLE, size: 12, color: NAVY } },
+    children: [new TextRun({ text: plainText(tok.tokens || []), font: FONT_BODY, size: 21, italics: true, color: GRAY })],
+  });
+}
+
+function hrPara() {
+  return new Paragraph({
+    spacing: { before: 120, after: 120 },
+    border: { bottom: { style: BorderStyle.SINGLE, size: 6, color: "BFBFBF" } },
+    children: [],
+  });
+}
+
+// ---- main loop ----
+const children = [];
+let firstH1Seen = false;
+for (const tok of tokens) {
+  switch (tok.type) {
+    case 'heading':
+      if (tok.depth === 1 && firstH1Seen) {
+        children.push(new Paragraph({ children: [new PageBreak()] }));
+      }
+      if (tok.depth === 1) firstH1Seen = true;
+      children.push(headingPara(tok));
+      break;
+    case 'paragraph': children.push(paragraphPara(tok)); break;
+    case 'code':      children.push(codePara(tok)); break;
+    case 'table':     children.push(tableBlock(tok)); break;
+    case 'list':      children.push(...listBlock(tok, !!tok.ordered)); break;
+    case 'blockquote':children.push(blockquotePara(tok)); break;
+    case 'hr':        children.push(hrPara()); break;
+    case 'space':     break;
+    default:          break;
+  }
+}
+
+const doc = new Document({
+  creator: 'Hermes Agent',
+  title: TITLE,
+  sections: [{ properties: {}, children }],
+});
+
+Packer.toBuffer(doc).then(buf => {
+  fs.writeFileSync(OUT, buf);
+  console.log('WROTE', OUT, buf.length, 'bytes');
+});
--- a/skills_library/all/office-text-extract/SKILL.md
+++ b/skills_library/all/office-text-extract/SKILL.md
@ -0,0 +1,87 @@
+---
+name: office-text-extract
+description: "Extract text from .docx/.pptx/.xlsx/.pdf without pandoc."
+version: 2.0.0
+author: Hermes Agent
+license: Apache-2.0 (docx/pptx/xlsx engine from genspark-ai/genoffice)
+metadata:
+  hermes:
+    tags: [office, docx, pptx, xlsx, pdf, text-extraction, headless, genoffice, pymupdf]
+    related_skills: [docx, xlsx, powerpoint, pdf, markdown-to-docx, ocr-and-documents]
+---
+
+# Office Text Extraction (headless)
+
+从 `.docx` / `.pptx` / `.xlsx` / `.pdf` 提取纯文本，**无需 pandoc / LibreOffice**。统一入口一个命令覆盖四种格式：
+
+- `.docx` / `.pptx` / `.xlsx` → GenOffice 引擎（github.com/genspark-ai/genoffice，Apache-2.0）的 `file-parse` + `docx-engine`，esbuild 打成单文件 Node CLI
+- `.pdf`（文本型）→ pymupdf（Python）
+- `.pdf`（扫描件/图片型，需 OCR）→ marker-pdf，见 `ocr-and-documents` 技能
+
+## 何时用
+
+- 需要**读取**已有 Office/PDF 文档内容（本机无 pandoc/LibreOffice 时尤其适用）
+- 给 AI 附加 Office/PDF 附件前先转文本
+- 快速抽取正文做检索、摘要、diff
+
+## 工具位置
+
+```
+scripts/office-extract            # 统一入口（bash，按扩展名分发）★ 主入口
+scripts/genoffice-extract.mjs     # docx/pptx/xlsx 引擎（Node，单文件，~1MB）
+scripts/extract_pdf.py            # pdf 文本提取（pymupdf）
+scripts/build.sh                  # 可复现重建 Node 引擎（从 genoffice 源重新打包）
+```
+
+要求：Node 18+（docx/pptx/xlsx）+ `pymupdf`（pdf）。
+
+## 用法
+
+```bash
+BIN=scripts/office-extract
+
+# 统一入口，四种格式都行（纯文本到 stdout）
+"$BIN" /path/to/file.docx
+"$BIN" /path/to/file.pdf
+
+# JSON 模式（{ok,kind,ext,text|error}），便于程序化处理
+"$BIN" --json /path/to/file.xlsx
+```
+
+也可以直接调底层：`node scripts/genoffice-extract.mjs <docx|pptx|xlsx>`、`python3 scripts/extract_pdf.py <pdf>`。
+
+## 输出格式
+
+- **docx**：标题转 `#`/`##`（按级别），列表项转 `-`，有序列表转 `1. 2. 3.`，表格行转 `cell1 | cell2 | ...`，正文段落原样。接近 Markdown。
+- **pptx**：每页一段 `## Slide N`，每段文字一行，按 slide 编号升序。
+- **xlsx**：每个工作表一段 `# SheetName`，行内单元格用 ` | ` 连接（空单元格占位对齐），inlineStr/sharedStrings/布尔/数字均正确处理。
+- **pdf**：每页文本用 `\n\n` 分隔，中文/多页正常。
+- 中文、ASCII 图、多列表格均无损（docx/pdf 实测通过）。
+
+## 坑
+
+1. **PDF 分两类**：
+   - 文本型 PDF（有嵌入文本层）→ `extract_pdf.py`（pymupdf），即时返回。
+   - 扫描件/图片型 PDF（无文本层）→ pymupdf 提取为空，脚本会提示改用 marker-pdf OCR（`ocr-and-documents` 技能的 `extract_marker.py`）。marker-pdf 需 ~5GB（PyTorch+模型）且首次运行下载模型。
+2. **只有提取，没有生成/编辑**：本工具只读文本。生成 docx 用 `docx` 技能（docx-js）、md→docx 用 `markdown-to-docx` 技能；编辑已有 docx 的批注/修订/图表需 `docx-engine` 的 patch 能力（未封装）。
+3. **Node 版本**：bundle target 是 node18，老环境可能不兼容。
+4. **大文件**：docx 解析会展开整个 OOXML；超大文件（>50MB）内存占用高。
+
+## 重建 Node 引擎（genoffice 上游更新时）
+
+```bash
+bash scripts/build.sh
+```
+
+`build.sh` 从 genoffice main 分支重新下载 `file-parse` + `docx-engine` 源码、装依赖、esbuild 打包，覆盖 `scripts/genoffice-extract.mjs`。需网络可达 raw.githubusercontent.com（若被墙，先建 SOCKS5 隧道并设 `PROXY=socks5h://127.0.0.1:1080`）。
+
+## 迁移到其他机器 / pipeline-app
+
+- **docx/pptx/xlsx**：`genoffice-extract.mjs` 是纯 Node 单文件工具，**只需 Node 18+**，零依赖。拷过去 `node <path> <file>` 即用。
+- **pdf**：需要目标机 `pip install pymupdf`（~25MB，一条命令）。
+- **pdf OCR**：需要 `pip install marker-pdf` + ~5GB 磁盘（首次运行下载模型）。
+
+三者迁移成本递增：docx/pptx/xlsx（零依赖）< pdf 文本（pip 装 pymupdf）< pdf OCR（5GB）。
+
+- 若 pipeline-app 服务器有 Node：拷 `genoffice-extract.mjs` 即可在服务器 shell 用。
+- 若要 pipeline-app 的 DSPY/前端调用：另写一个薄 Sage 模块，用 `subprocess`/`child_process` 调对应 CLI 取 stdout（后续集成决策，本 skill 不涉及）。
--- a/skills_library/all/office-text-extract/scripts/build.sh
+++ b/skills_library/all/office-text-extract/scripts/build.sh
@ -0,0 +1,146 @@
+#!/usr/bin/env bash
+# 从 genspark-ai/genoffice 重建 genoffice-extract.mjs
+# 用法: [PROXY=socks5h://127.0.0.1:1080] bash build.sh
+# 产物覆盖本目录下的 genoffice-extract.mjs
+set -euo pipefail
+
+BASE="https://raw.githubusercontent.com/genspark-ai/genoffice/main"
+PROXY="${PROXY:-}"
+CURL=(curl -s --fail --connect-timeout 30)
+[ -n "$PROXY" ] && CURL+=(-x "$PROXY")
+
+SKILL_DIR="$(cd "$(dirname "$0")" && pwd)"
+WORK="$(mktemp -d)"
+trap 'rm -rf "$WORK"' EXIT
+
+# 仅需 file-parse + docx-engine 两个包的源码（PDF 未打包，pdf.ts 不会被 cli.ts 引入）
+FILES=(
+  packages/docx-engine/package.json
+  packages/docx-engine/src/blank.ts
+  packages/docx-engine/src/chart.ts
+  packages/docx-engine/src/generate.ts
+  packages/docx-engine/src/index.ts
+  packages/docx-engine/src/ink.ts
+  packages/docx-engine/src/list-markers.ts
+  packages/docx-engine/src/math.ts
+  packages/docx-engine/src/metafile.ts
+  packages/docx-engine/src/notes.ts
+  packages/docx-engine/src/parse.ts
+  packages/docx-engine/src/patch.ts
+  packages/docx-engine/src/protection.ts
+  packages/docx-engine/src/scan.ts
+  packages/docx-engine/src/section.ts
+  packages/docx-engine/src/sources.ts
+  packages/docx-engine/src/symbol-fonts.ts
+  packages/docx-engine/src/text-patch.ts
+  packages/docx-engine/src/theme.ts
+  packages/docx-engine/src/tiff.ts
+  packages/docx-engine/src/types.ts
+  packages/docx-engine/src/watermark.ts
+  packages/docx-engine/src/xml-utils.ts
+  packages/docx-engine/src/zip-load.ts
+  packages/docx-engine/src/vendor/emf-converter/index.mjs
+  packages/file-parse/package.json
+  packages/file-parse/src/docx.ts
+  packages/file-parse/src/index.ts
+  packages/file-parse/src/parse.ts
+  packages/file-parse/src/pdf.ts
+  packages/file-parse/src/pptx.ts
+  packages/file-parse/src/xlsx.ts
+)
+
+echo "[1/4] 下载源码 ($((${#FILES[@]})) 个文件)..."
+for f in "${FILES[@]}"; do
+  mkdir -p "$WORK/$(dirname "$f")"
+  "${CURL[@]}" "$BASE/$f" -o "$WORK/$f"
+done
+
+echo "[2/4] 写入口 cli.ts..."
+cat > "$WORK/cli.ts" <<'CLIEOF'
+#!/usr/bin/env node
+import { readFile } from 'node:fs/promises'
+import { extname } from 'node:path'
+import { docxToText } from './packages/file-parse/src/docx.ts'
+import { pptxToText } from './packages/file-parse/src/pptx.ts'
+import { xlsxToText } from './packages/file-parse/src/xlsx.ts'
+
+const SUPPORTED = new Set(['docx', 'pptx', 'xlsx', 'pdf'])
+
+function usage() {
+  console.error(
+    [
+      'genoffice-extract — 提取 Office 文档纯文本（无 pandoc/LibreOffice 依赖）',
+      '',
+      '用法:',
+      '  genoffice-extract <file.docx|pptx|xlsx>        输出纯文本到 stdout',
+      '  genoffice-extract --json <file>               输出 {ok,kind,ext,text|error} JSON',
+      '',
+      '支持: .docx .pptx .xlsx  （.pdf 未打包，请用 pdf/pymupdf 技能）',
+    ].join('\n'),
+  )
+}
+
+async function main() {
+  const args = process.argv.slice(2)
+  const json = args.includes('--json')
+  const files = args.filter((a) => !a.startsWith('--'))
+  if (files.length === 0) {
+    usage()
+    process.exit(2)
+  }
+  const file = files[0]
+  const ext = extname(file).slice(1).toLowerCase()
+
+  if (!SUPPORTED.has(ext)) {
+    const err = `Unsupported file type: .${ext || 'unknown'} (支持 docx/pptx/xlsx)`
+    if (json) console.log(JSON.stringify({ ok: false, kind: 'unsupported', ext, error: err }))
+    else console.error(err)
+    process.exit(1)
+  }
+  if (ext === 'pdf') {
+    const err = 'PDF 提取未打包（pdfjs-dist 体积/worker 复杂）；请用 pdf 或 ocr-and-documents 技能'
+    if (json) console.log(JSON.stringify({ ok: false, kind: 'unsupported', ext, error: err }))
+    else console.error(err)
+    process.exit(1)
+  }
+
+  let bytes
+  try {
+    bytes = await readFile(file)
+  } catch (e) {
+    const err = `无法读取文件: ${e instanceof Error ? e.message : String(e)}`
+    if (json) console.log(JSON.stringify({ ok: false, kind: 'unsupported', ext, error: err }))
+    else console.error(err)
+    process.exit(1)
+  }
+
+  try {
+    let text = ''
+    if (ext === 'docx') text = await docxToText(bytes)
+    else if (ext === 'pptx') text = await pptxToText(bytes)
+    else if (ext === 'xlsx') text = await xlsxToText(bytes)
+    if (json) console.log(JSON.stringify({ ok: true, kind: 'text', ext, text }))
+    else console.log(text)
+  } catch (e) {
+    const err = e instanceof Error ? e.message : String(e)
+    if (json) console.log(JSON.stringify({ ok: false, kind: 'text', ext, error: err }))
+    else {
+      console.error('提取失败: ' + err)
+      process.exit(1)
+    }
+  }
+}
+
+main()
+CLIEOF
+
+echo "[3/4] 安装依赖并 esbuild 打包..."
+cd "$WORK"
+npm init -y >/dev/null 2>&1
+npm install jszip fast-xml-parser utif2 esbuild >/dev/null 2>&1
+npx esbuild cli.ts --bundle --platform=node --format=esm --target=node18 \
+  --outfile="$SKILL_DIR/genoffice-extract.mjs" \
+  --alias:@genoffice/docx-engine="$WORK/packages/docx-engine/src/index.ts" 2>&1 | grep -v 'tsconfig.base.json' || true
+
+chmod +x "$SKILL_DIR/genoffice-extract.mjs"
+echo "[4/4] 完成: $SKILL_DIR/genoffice-extract.mjs ($(du -h "$SKILL_DIR/genoffice-extract.mjs" | cut -f1))"
--- a/skills_library/all/office-text-extract/scripts/extract_pdf.py
+++ b/skills_library/all/office-text-extract/scripts/extract_pdf.py
@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""从 PDF 提取纯文本（pymupdf）。输出格式与 genoffice-extract.mjs 对齐。
+
+用法:
+  extract_pdf.py <file.pdf>          纯文本到 stdout
+  extract_pdf.py --json <file.pdf>   {ok,kind,ext,text|error} JSON
+
+注意: pymupdf 只提取文本型 PDF（有嵌入文本层）。扫描件/图片型 PDF 无文本层，
+提取为空 —— 此时提示改用 marker-pdf OCR。
+"""
+import sys
+import json
+import pymupdf
+
+
+def main() -> int:
+    args = sys.argv[1:]
+    as_json = '--json' in args
+    files = [a for a in args if not a.startswith('--')]
+    if not files:
+        print('usage: extract_pdf.py [--json] <file.pdf>', file=sys.stderr)
+        return 2
+
+    path = files[0]
+    ext = 'pdf'
+
+    def emit(ok: bool, kind: str, text: str = '', error: str = '') -> int:
+        if as_json:
+            payload = {'ok': ok, 'kind': kind, 'ext': ext}
+            if ok:
+                payload['text'] = text
+            else:
+                payload['error'] = error
+            print(json.dumps(payload, ensure_ascii=False))
+        else:
+            if ok:
+                sys.stdout.write(text)
+            else:
+                print(error, file=sys.stderr)
+        return 0 if ok else 1
+
+    try:
+        doc = pymupdf.open(path)
+    except Exception as e:  # noqa: BLE001
+        return emit(False, 'unsupported', error='无法打开 PDF: %s' % (e,))
+
+    try:
+        pages = [page.get_text() for page in doc]
+        text = '\n\n'.join(pages).strip()
+        doc.close()
+        if not text:
+            return emit(
+                False, 'text',
+                error=('该 PDF 无文本层（可能是扫描件/图片型）。'
+                       'pymupdf 无法 OCR，请改用 marker-pdf OCR（见 ocr-and-documents 技能）。'),
+            )
+        return emit(True, 'text', text=text)
+    except Exception as e:  # noqa: BLE001
+        try:
+            doc.close()
+        except Exception:  # noqa: BLE001
+            pass
+        return emit(False, 'text', error='提取失败: %s' % (e,))
+
+
+if __name__ == '__main__':
+    sys.exit(main())
--- a/skills_library/all/office-text-extract/scripts/genoffice-extract.mjs
+++ b/skills_library/all/office-text-extract/scripts/genoffice-extract.mjs
--- a/skills_library/all/office-text-extract/scripts/office-extract
+++ b/skills_library/all/office-text-extract/scripts/office-extract
@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# office-extract — 统一 Office/PDF 文本提取入口
+# 按扩展名分发: .docx/.pptx/.xlsx → genoffice-extract.mjs (Node)
+#              .pdf              → extract_pdf.py (pymupdf)
+# 用法: office-extract [--json] <file.docx|pptx|xlsx|pdf>
+set -euo pipefail
+
+DIR="$(cd "$(dirname "$0")" && pwd)"
+
+if [ $# -eq 0 ]; then
+  echo "office-extract — 统一 Office/PDF 文本提取" >&2
+  echo "用法: office-extract [--json] <file.docx|pptx|xlsx|pdf>" >&2
+  exit 2
+fi
+
+# 找最后一个非 -- 开头的参数作为文件路径
+file=""
+for a in "$@"; do
+  case "$a" in
+    --*) ;;
+    *) file="$a" ;;
+  esac
+done
+
+ext="${file##*.}"
+ext="$(printf '%s' "$ext" | tr '[:upper:]' '[:lower:]')"
+
+case "$ext" in
+  docx|pptx|xlsx)
+    exec node "$DIR/genoffice-extract.mjs" "$@"
+    ;;
+  pdf)
+    exec python3 "$DIR/extract_pdf.py" "$@"
+    ;;
+  *)
+    echo "不支持的格式: .${ext:-未知}（支持 docx/pptx/xlsx/pdf）" >&2
+    exit 1
+    ;;
+esac