dsh-convert-core 0.1.0-alpha.1

package/scripts/asr.py

@@ -0,0 +1,55 @@
+#!/usr/bin/env python3
+"""media-to-md 转写 helper:音频 → 带时间戳的转写 JSON(本地 faster-whisper,不调用云端/LLM)。
+
+由 `src/models.ts` 通过 `uvx --with faster-whisper python` 拉起(转换内核的 iddoc adapter 也走这里)。
+
+用法:
+    python asr.py <audio> [model] [lang]
+
+- model: faster-whisper 模型名(small / large-v3-turbo / medium)或**本地模型目录**。
+- lang:  语言代码(zh / en …),省略则自动检测。
+
+输出(stdout,单行 JSON):
+    {"model","language","language_probability","duration","segments":[{"start","end","text"}]}
+"""
+import json
+import os
+import sys
+
+
+def main() -> int:
+    if len(sys.argv) < 2:
+        print("usage: asr.py <media> [model] [lang]", file=sys.stderr)
+        return 2
+    media = sys.argv[1]
+    model_name = sys.argv[2] if len(sys.argv) > 2 and sys.argv[2] else os.environ.get("MEDIA_TO_MD_ASR_MODEL", "small")
+    lang = sys.argv[3] if len(sys.argv) > 3 and sys.argv[3] else None
+
+    from faster_whisper import WhisperModel
+
+    device = os.environ.get("MEDIA_TO_MD_ASR_DEVICE", "cpu")
+    compute_type = os.environ.get("MEDIA_TO_MD_ASR_COMPUTE", "int8")
+    model = WhisperModel(model_name, device=device, compute_type=compute_type)
+    segments, info = model.transcribe(
+        media,
+        language=lang,
+        beam_size=int(os.environ.get("MEDIA_TO_MD_ASR_BEAM", "1")),
+        vad_filter=True,
+    )
+    payload = {
+        "model": model_name,
+        "language": getattr(info, "language", None),
+        "language_probability": getattr(info, "language_probability", None),
+        "duration": getattr(info, "duration", None),
+        "segments": [
+            {"start": round(s.start, 3), "end": round(s.end, 3), "text": (s.text or "").strip()}
+            for s in segments
+            if (s.text or "").strip()
+        ],
+    }
+    print(json.dumps(payload, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/scripts/ocr.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""media-to-md 图片 OCR helper:图片 → 画面文字 JSON(本地 RapidOCR,默认不启用)。
+
+只有在 `imageOcr=true` 时才会被 `src/convert.ts` 调用(约定:图片默认直接收录、不做 OCR)。
+
+用法:
+    python ocr.py <img1> [img2 ...]
+输出(stdout,单行 JSON 数组):
+    [{"path","lines":[...],"text":"..."}]
+"""
+import json
+import sys
+
+
+def main() -> int:
+    paths = sys.argv[1:]
+    if not paths:
+        print("usage: ocr.py <img...>", file=sys.stderr)
+        return 2
+
+    from rapidocr_onnxruntime import RapidOCR
+
+    engine = RapidOCR()
+    results = []
+    for path in paths:
+        entry = {"path": path, "lines": [], "text": "", "error": None}
+        try:
+            res, _elapse = engine(path)
+            lines = [(item[1] or "").strip() for item in (res or [])]
+            entry["lines"] = [line for line in lines if line]
+            entry["text"] = "\n".join(entry["lines"])
+        except Exception as exc:  # 单张失败不影响整体
+            entry["error"] = f"{type(exc).__name__}: {exc}"
+        results.append(entry)
+    print(json.dumps(results, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/scripts/pdf_pages.py

@@ -0,0 +1,49 @@
+#!/usr/bin/env python3
+"""media-to-md 轻量扫描档 helper:PDF → 逐页 PNG(pypdfium2,无系统依赖)。
+
+为什么不用 poppler:`pdftoppm` 是外部二进制,不是每台机器都有(本机就没有);
+pypdfium2 是 pip wheel,自带 pdfium,跨平台、可被 uvx 按需拉起。
+
+用法:
+    python pdf_pages.py <in.pdf> <out_dir> [dpi] [max_pages]
+
+输出(stdout,单行 JSON):
+    {"pages":[{"index":1,"path":"...","width":..,"height":..}],"totalPages":N,"renderedPages":M}
+"""
+import json
+import os
+import sys
+
+
+def main() -> int:
+    if len(sys.argv) < 3:
+        print("usage: pdf_pages.py <in.pdf> <out_dir> [dpi] [max_pages]", file=sys.stderr)
+        return 2
+    pdf_path = sys.argv[1]
+    out_dir = sys.argv[2]
+    dpi = int(sys.argv[3]) if len(sys.argv) > 3 and sys.argv[3] else 150
+    max_pages = int(sys.argv[4]) if len(sys.argv) > 4 and sys.argv[4] else 50
+
+    import pypdfium2 as pdfium
+
+    os.makedirs(out_dir, exist_ok=True)
+    document = pdfium.PdfDocument(pdf_path)
+    total = len(document)
+    limit = max(0, min(total, max_pages))
+    scale = max(0.1, dpi / 72.0)
+
+    pages = []
+    for index in range(limit):
+        page = document[index]
+        bitmap = page.render(scale=scale)
+        image = bitmap.to_pil()
+        path = os.path.join(out_dir, f"page-{index + 1:04d}.png")
+        image.save(path)
+        pages.append({"index": index + 1, "path": path, "width": image.width, "height": image.height})
+
+    print(json.dumps({"pages": pages, "totalPages": total, "renderedPages": len(pages)}))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/scripts/vendor.mjs

@@ -0,0 +1,76 @@
+/**
+ * Copy the browser dependencies into `web/vendor/` during `pnpm build`.
+ *
+ * The console is deliberately buildless: one HTML file + one app.js using the
+ * Vue 3 Options API against the global `Vue` object. That keeps the kernel
+ * runnable from a plain `node bin/tita-convert.mjs` with no bundler in the
+ * loop, which matters because the DSH plugin serves this page as static assets.
+ *
+ * Lucide is vendored as a plain SVG map rather than as the runtime
+ * `createIcons()` scanner: the scanner replaces DOM nodes behind Vue's back,
+ * while a map lets the template use `v-html` on a node Vue owns.
+ */
+
+import { mkdir, readFile, writeFile } from 'node:fs/promises'
+import { createRequire } from 'node:module'
+import { dirname, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+
+const require = createRequire(import.meta.url)
+const root = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const vendorDir = resolve(root, 'web', 'vendor')
+
+/** Every icon the console template references. */
+export const CONSOLE_ICONS = [
+  'layers', 'file-up', 'link', 'video', 'file-text', 'clipboard-type',
+  'refresh-cw', 'hard-drive', 'alert-triangle', 'check-circle', 'activity',
+  'inbox', 'package', 'scan-eye', 'x', 'rotate-ccw', 'loader', 'play',
+  'copy', 'download', 'circle-alert', 'circle-check',
+]
+
+await mkdir(vendorDir, { recursive: true })
+
+async function vendorScript(specifier, filename, label) {
+  try {
+    const source = require.resolve(specifier)
+    const data = await readFile(source)
+    await writeFile(resolve(vendorDir, filename), data)
+    console.log(`[convert-core] vendor: ${label.padEnd(20)} ${(data.length / 1024).toFixed(0)} KiB`)
+    return true
+  } catch (cause) {
+    console.warn(`[convert-core] vendor: ${label} 跳过 —— ${cause.message}`)
+    return false
+  }
+}
+
+await vendorScript('vue/dist/vue.global.prod.js', 'vue.global.prod.js', 'vue (Options API)')
+
+/** Lucide renamed a few icons; keep the template names stable. */
+const ALIASES = {
+  'alert-triangle': 'triangle-alert',
+  'check-circle': 'circle-check',
+}
+
+// Lucide ships raw SVGs in `lucide-static`; read them and emit one JS map.
+const icons = {}
+let missing = []
+for (const name of CONSOLE_ICONS) {
+  try {
+    const file = ALIASES[name] ?? name
+    const path = require.resolve(`lucide-static/icons/${file}.svg`)
+    const svg = (await readFile(path, 'utf8'))
+      .replace(/\s+width="24"/, '')
+      .replace(/\s+height="24"/, '')
+      .replace(/\n+/g, '')
+      .trim()
+    icons[name] = svg
+  } catch {
+    missing.push(name)
+  }
+}
+await writeFile(
+  resolve(vendorDir, 'icons.js'),
+  `// Generated by scripts/vendor.mjs from the lucide-static icon set.\nwindow.__CONVERT_ICONS__ = ${JSON.stringify(icons, null, 2)};\n`,
+)
+console.log(`[convert-core] vendor: ${'lucide icons'.padEnd(20)} ${Object.keys(icons).length}/${CONSOLE_ICONS.length} icons`)
+if (missing.length > 0) console.warn(`[convert-core] vendor: 缺少图标 ${missing.join(', ')}`)

package/src/adapters/anydoc.ts

@@ -0,0 +1,356 @@
+/**
+ * The document engine: `@firecrawl/anydoc` (MIT, Rust, prebuilt Node binding).
+ *
+ * One adapter replaces what used to be five (pdfjs text-layer reader, mammoth,
+ * the high-quality docling tier, plus the format routing between them). anydoc
+ * covers doc/docx/docm, ppt/pps/pot/pptx/pptm/ppsx/ppsm, xls/xlsx/xlsm/xlsb,
+ * odt/ods/odp, rtf, epub, csv and text-layer pdf, detects the format from the
+ * bytes, and needs no Python, no models and no external binary.
+ *
+ * The one thing it cannot do for us is the 0013 image contract: Markdown has no
+ * way to embed bytes, so an embedded image renders as its alt text while the
+ * bytes stay on `document.assets`. Retrieval only reads local relative image
+ * paths, so this adapter re-materialises every embedded image into `assets/`
+ * and puts a real `![](assets/img-…)` reference back where the document had it.
+ */
+
+import { createRequire } from 'node:module'
+import { readFile } from 'node:fs/promises'
+import { basename, extname } from 'node:path'
+import type { ConvertWarning, SourceFormat } from '../contract.js'
+import { sha1, sha256 } from '../markdown.js'
+import { AdapterError, type Adapter, type AdapterAvailability, type AdapterContext, type AdapterOutput, type PreparedAsset } from './types.js'
+
+export const ANYDOC_EXTENSIONS = [
+  // Word
+  '.doc', '.docx', '.docm',
+  // PowerPoint
+  '.ppt', '.pps', '.pot', '.pptx', '.pptm', '.ppsx', '.ppsm',
+  // Excel
+  '.xls', '.xlsx', '.xlsm', '.xlsb',
+  // OpenDocument
+  '.odt', '.ods', '.odp',
+  // Other containers anydoc reads
+  '.rtf', '.epub', '.csv', '.pdf',
+]
+
+interface AnyDocAsset {
+  id?: number
+  mediaType?: string
+  originPart?: string
+  data?: Uint8Array
+}
+
+interface AnyDocInline {
+  kind: string
+  alt?: string
+  source?: { kind?: string; assetId?: number; url?: string }
+  text?: string
+}
+
+interface AnyDocBlock {
+  kind: string
+  content?: AnyDocInline[]
+}
+
+interface AnyDocDocument {
+  blocks?: AnyDocBlock[]
+  notes?: unknown[]
+  assets?: AnyDocAsset[]
+}
+
+interface AnyDocModule {
+  toDocument(bytes: Uint8Array): Promise<AnyDocDocument>
+  toMarkdown(path: string, options?: Record<string, unknown>): Promise<string>
+  toMarkdownBytes(bytes: Uint8Array, format?: string): Promise<string>
+  formatFromBytes(bytes: Uint8Array): string | null
+  formatFromPath(path: string): string | null
+}
+
+const MEDIA_EXTENSIONS: Record<string, string> = {
+  'image/png': 'png',
+  'image/jpeg': 'jpg',
+  'image/jpg': 'jpg',
+  'image/gif': 'gif',
+  'image/webp': 'webp',
+  'image/tiff': 'tiff',
+  'image/bmp': 'bmp',
+  'image/svg+xml': 'svg',
+  'image/avif': 'avif',
+  'image/heic': 'heic',
+}
+
+let modulePromise: Promise<AnyDocModule> | undefined
+let moduleVersion: string | undefined
+let moduleError: string | undefined
+
+/** Load the binding once; the platform binary ships as an optional dependency. */
+export async function loadAnydoc(): Promise<AnyDocModule> {
+  if (!modulePromise) {
+    modulePromise = (async () => {
+      const loaded = (await import('@firecrawl/anydoc')) as unknown as AnyDocModule & { default?: AnyDocModule }
+      const candidate = typeof loaded.toMarkdown === 'function' ? loaded : loaded.default
+      if (!candidate || typeof candidate.toMarkdown !== 'function') throw new Error('找不到 toMarkdown 导出')
+      try {
+        const require = createRequire(import.meta.url)
+        const pkg = JSON.parse(await readFile(require.resolve('@firecrawl/anydoc/package.json'), 'utf8')) as { version?: string }
+        moduleVersion = pkg.version
+      } catch {
+        /* version is cosmetic */
+      }
+      return candidate
+    })().catch(error => {
+      moduleError = error instanceof Error ? error.message : String(error)
+      modulePromise = undefined
+      throw error
+    })
+  }
+  return await modulePromise
+}
+
+export async function anydocAvailability(): Promise<AdapterAvailability> {
+  try {
+    await loadAnydoc()
+    return { available: true, ...(moduleVersion ? { versions: { '@firecrawl/anydoc': moduleVersion } } : {}) }
+  } catch (cause) {
+    return {
+      available: false,
+      reason: `无法加载 @firecrawl/anydoc：${moduleError ?? (cause instanceof Error ? cause.message : String(cause))}`,
+      suggestion: '在 packages/dsh-convert-core 目录执行 pnpm install；Apple Silicon 需要 @firecrawl/anydoc-darwin-arm64（随 optionalDependencies 自动安装）。',
+    }
+  }
+}
+
+function extensionFor(asset: AnyDocAsset): string | undefined {
+  const mediaType = asset.mediaType?.split(';')[0]?.trim().toLowerCase()
+  if (mediaType && MEDIA_EXTENSIONS[mediaType]) return MEDIA_EXTENSIONS[mediaType]
+  const fromPart = asset.originPart ? /\.([a-zA-Z0-9]{2,5})$/.exec(asset.originPart)?.[1]?.toLowerCase() : undefined
+  if (fromPart && fromPart !== 'bin') return fromPart
+  return undefined
+}
+
+/** In document order, every inline image that points at an embedded asset. */
+function collectEmbeddedImages(document: AnyDocDocument): Array<{ assetId: number; alt: string; asset: AnyDocAsset }> {
+  const found: Array<{ assetId: number; alt: string; asset: AnyDocAsset }> = []
+  for (const block of document.blocks ?? []) {
+    for (const inline of block.content ?? []) {
+      if (inline.kind !== 'image') continue
+      const source = inline.source
+      if (!source || source.kind !== 'asset' || typeof source.assetId !== 'number') continue
+      const asset = document.assets?.[source.assetId]
+      if (!asset) continue
+      found.push({ assetId: source.assetId, alt: (inline.alt ?? '').trim(), asset })
+    }
+  }
+  return found
+}
+
+export interface ImagePlacement {
+  markdown: string
+  placed: number
+  appended: number
+  warnings: string[]
+}
+
+/**
+ * Put real image references back into the Markdown.
+ *
+ * anydoc renders an embedded image as its alt text, so placement is a match on
+ * that text. A standalone-line match is used only when it is unambiguous;
+ * anything that cannot be located is appended under a `## 图片` heading rather
+ * than guessed at, and the count is reported.
+ */
+export function placeImageReferences(markdown: string, images: Array<{ rel: string; alt: string }>): ImagePlacement {
+  const lines = markdown.split('\n')
+  const consumed = new Set<number>()
+  const remaining: Array<{ rel: string; alt: string }> = []
+  const warnings: string[] = []
+
+  for (const image of images) {
+    const alt = image.alt.trim()
+    let index = -1
+    if (alt) {
+      // Only accept a line that is exactly the alt text, and only once.
+      const candidates = lines
+        .map((line, position) => (line.trim() === alt && !consumed.has(position) ? position : -1))
+        .filter(position => position >= 0)
+      if (candidates.length === 1) index = candidates[0]!
+    }
+    if (index >= 0) {
+      consumed.add(index)
+      lines[index] = `![${alt}](${image.rel})`
+      continue
+    }
+    remaining.push(image)
+  }
+
+  let output = lines.join('\n')
+  if (remaining.length > 0) {
+    const section = remaining.map(image => `![${image.alt || basename(image.rel)}](${image.rel})`).join('\n\n')
+    output = `${output.trimEnd()}\n\n## 图片\n\n${section}\n`
+    warnings.push(
+      `${remaining.length} 张内嵌图片在正文中没有可定位的占位（无 alt 文本或 alt 不唯一），已追加到文末「图片」小节。`,
+    )
+  }
+  return { markdown: output, placed: images.length - remaining.length, appended: remaining.length, warnings }
+}
+
+function sourceFormatFor(anydocFormat: string | null, filename: string): SourceFormat {
+  const format = anydocFormat ?? extname(filename).replace('.', '').toLowerCase()
+  if (format === 'pdf') return 'pdf'
+  if (format === 'csv') return 'text'
+  return 'office'
+}
+
+/** anydoc's error codes are already precise; keep them instead of flattening. */
+const ANYDOC_SUGGESTIONS: Record<string, string> = {
+  needsOcr:
+    '该 PDF 是扫描件/纯图片页，anydoc 本地不做 OCR（它只覆盖文本层 PDF）。走扫描档：把 adapter 指定为 "scanned"，' +
+    '或设环境变量 SCANNED_OCR=1 让 .pdf 直接走它 —— 扫描档用 pypdfium2 渲染页图 + 本地 RapidOCR 识别文字，不出网，' +
+    '但不做版面还原；需要版面/表格/公式时见 issues/0017 的可选高质量档。' +
+    '不要使用 anydoc 的 ocr:"hosted"：它会把整份文档发往 Firecrawl Parse 云端，与 0015 的本地优先原则冲突。',
+  encrypted: '文档有密码保护，请先去掉密码再转换。',
+  unsupported: '该格式不在本版范围内（doc/docx/ppt/pptx/xls/xlsx/odt/ods/odp/rtf/epub/csv/文本层 PDF）。',
+  malformed: '文件结构损坏，无法提取有意义的内容。',
+  resourceLimit: '文件触发了 anydoc 的解压/嵌套安全上限，可能是文档炸弹或异常嵌套。',
+  missingPart: '文档缺少必需的结构部件（可能被截断或由非标准工具生成）。',
+}
+
+function toAdapterError(cause: unknown): AdapterError {
+  const error = cause as Error & { code?: string; pages?: number[] }
+  const code = typeof error?.code === 'string' ? error.code : undefined
+  if (code && code !== 'io') {
+    const pages = Array.isArray(error.pages) && error.pages.length > 0 ? `（第 ${error.pages.slice(0, 8).join(', ')} 页）` : ''
+    return new AdapterError(
+      `ANYDOC_${code.replace(/([a-z])([A-Z])/g, '$1_$2').toUpperCase()}`,
+      `${error.message || code}${pages}`,
+      ANYDOC_SUGGESTIONS[code],
+    )
+  }
+  return new AdapterError('ANYDOC_FAILED', error instanceof Error ? error.message : String(cause), ANYDOC_SUGGESTIONS.malformed)
+}
+
+async function run(context: AdapterContext): Promise<AdapterOutput> {
+  const sourcePath = context.inputPath
+  if (!sourcePath) throw new AdapterError('INPUT_REQUIRED', '文档适配器需要文件路径。')
+
+  const anydoc = await loadAnydoc().catch(cause => {
+    throw new AdapterError(
+      'DEPENDENCY_MISSING',
+      `无法加载 @firecrawl/anydoc：${cause instanceof Error ? cause.message : String(cause)}`,
+      '在 packages/dsh-convert-core 目录执行 pnpm install。',
+    )
+  })
+
+  const filename = context.filename || basename(sourcePath)
+  const bytes = await readFile(sourcePath)
+  context.report(0.15, 'detect', '识别文档格式')
+
+  let detected: string | null = null
+  try {
+    detected = anydoc.formatFromBytes(new Uint8Array(bytes))
+  } catch {
+    detected = null
+  }
+
+  context.report(0.35, 'convert', `解析 ${detected ?? extname(filename)}`)
+  let markdown: string
+  let document: AnyDocDocument = {}
+  const assetNote: string[] = []
+  try {
+    markdown = await anydoc.toMarkdown(sourcePath)
+  } catch (cause) {
+    throw toAdapterError(cause)
+  }
+  // `toDocument` covers the OOXML / OpenDocument / RTF / EPUB / CSV family. PDF
+  // renders straight to Markdown through pdf-inspector and *rejects* the
+  // document model, so a PDF simply has no embedded assets to place.
+  if (detected !== 'pdf') {
+    try {
+      document = await anydoc.toDocument(new Uint8Array(bytes))
+    } catch (cause) {
+      const error = cause as Error & { code?: string }
+      // Never lose a good Markdown because the asset model was unavailable.
+      assetNote.push(
+        `内嵌资源枚举失败（${error?.code ?? 'unknown'}: ${error?.message ?? String(cause)}），本次只交付 Markdown，未落地图片。`,
+      )
+    }
+  }
+  if (!markdown.trim()) {
+    throw new AdapterError('CONTENT_REQUIRED', '文档解析后没有内容（可能全为图片或受保护内容）。')
+  }
+
+  // ---- asset glue: bytes live on the document model, not in the Markdown ----
+  const embedded = collectEmbeddedImages(document)
+  const images: Array<{ rel: string; alt: string }> = []
+  const assets: PreparedAsset[] = []
+  const warnings: ConvertWarning[] = []
+  let skippedObjects = 0
+
+  for (const entry of embedded) {
+    const data = entry.asset.data ? Buffer.from(entry.asset.data) : undefined
+    if (!data || data.length === 0) {
+      skippedObjects += 1
+      continue
+    }
+    if (!entry.asset.mediaType?.startsWith('image/')) {
+      // 0013 only consumes images; other embedded objects are reported, not dropped.
+      skippedObjects += 1
+      continue
+    }
+    const extension = extensionFor(entry.asset)
+    if (!extension) {
+      skippedObjects += 1
+      continue
+    }
+    const digest = sha1(data)
+    const rel = `assets/img-${digest.slice(0, 12)}.${extension}`
+    images.push({ rel, alt: entry.alt })
+    assets.push({
+      rel,
+      data,
+      ...(entry.asset.mediaType ? { contentType: entry.asset.mediaType } : {}),
+    })
+  }
+
+  const placement = placeImageReferences(markdown, images)
+  for (const note of assetNote) warnings.push({ code: 'ASSET_ENUMERATION_FAILED', message: note })
+  if (placement.appended > 0) {
+    warnings.push({ code: 'IMAGES_APPENDED', message: placement.warnings[0]! })
+  }
+  if (skippedObjects > 0) {
+    warnings.push({
+      code: 'EMBEDDED_OBJECTS_SKIPPED',
+      message: `文档里有 ${skippedObjects} 个内嵌对象不是可落地的图片（非 image/* 或缺少字节），未写入 assets/。`,
+    })
+  }
+  if (embedded.length > 0) {
+    warnings.push({
+      code: 'IMAGES_LOCALIZED',
+      message: `内嵌图片 ${embedded.length} 张已落地为本地 assets/ 并改写正文引用（${placement.placed} 张原位、${placement.appended} 张追加）。`,
+    })
+  }
+
+  const heading = /^#{1,6}\s+(.+)$/m.exec(placement.markdown)?.[1]?.trim()
+  return {
+    title: (context.request.title?.trim() || heading || basename(filename, extname(filename))).slice(0, 300),
+    markdown: placement.markdown.trim(),
+    sourceFormat: sourceFormatFor(detected, filename),
+    parser: `anydoc@${moduleVersion ?? 'unknown'}/${detected ?? 'unknown'}`,
+    sourcePath,
+    warnings,
+    dependencyVersions: { '@firecrawl/anydoc': moduleVersion ?? 'unknown' },
+    externalId: `sha256:${sha256(bytes)}`,
+    assets,
+  }
+}
+
+export const anydocAdapter: Adapter = {
+  id: 'anydoc',
+  description: 'DOC/DOCX/PPT/PPTX/XLS/XLSX/ODT/ODS/ODP/RTF/EPUB/CSV/文本层 PDF → GFM Markdown（Rust 内核，无模型、无外部二进制；内嵌图片落地到 assets/）',
+  kinds: ['file'],
+  extensions: ANYDOC_EXTENSIONS,
+  availability: () => ({ available: true, ...(moduleVersion ? { versions: { '@firecrawl/anydoc': moduleVersion } } : {}) }),
+  probe: anydocAvailability,
+  run,
+}