Method execute

src/tools/filesystem/pdf-read.ts:245–320

(args: z.infer<typeof PdfReadArgs>, _ctx: ToolContext)

  argsSchema = PdfReadArgs;

  async execute(args: z.infer<typeof PdfReadArgs>, _ctx: ToolContext): Promise<ToolResult> {
    try {
      const data = await fs.readFile(args.path);
      // Sanity check: starts with %PDF-
      if (data.subarray(0, 5).toString('latin1') !== '%PDF-') {
        return { content: `[PDF_ERROR] Not a PDF: ${args.path}`, isError: true };
      }
      // Encryption: presence of /Encrypt in trailer dict
      const tail = data.subarray(Math.max(0, data.length - 4096)).toString('latin1');
      if (tail.includes('/Encrypt')) {
        return {
          content: `[PDF_ENCRYPTED] ${args.path} is encrypted. Decrypt it first with: qpdf --decrypt input.pdf output.pdf`,
          isError: true,
        };
      }

      const objects = parseObjects(data);
      logger.debug(`pdf_read parsed ${objects.length} objects from ${args.path}`);

      // Find pages — objects whose dict contains /Type /Page (not /Pages)
      const pageObjects = objects.filter(o => /\/Type\s*\/Page\b(?!s)/.test(o.dict));
      const totalPages = pageObjects.length;
      const wanted = args.pages ? parsePageRange(args.pages, totalPages) : null;

      // For each page, follow /Contents reference(s) and decode the stream
      const pageText: string[] = [];
      let imagesOnly = 0;
      for (let i = 0; i < pageObjects.length; i++) {
        const pageNum = i + 1;
        if (wanted && !wanted.has(pageNum)) continue;
        const pageObj = pageObjects[i]!;
        // /Contents N 0 R OR /Contents [ N 0 R M 0 R ... ]
        const contentsMatch = pageObj.dict.match(/\/Contents\s+(\d+)\s+\d+\s+R|\/Contents\s*\[([^\]]+)\]/);
        if (!contentsMatch) { pageText.push(`\n[Page ${pageNum}: no contents]\n`); continue; }
        const refs: number[] = [];
        if (contentsMatch[1]) refs.push(parseInt(contentsMatch[1], 10));
        else if (contentsMatch[2]) {
          const refRe = /(\d+)\s+\d+\s+R/g;
          let r: RegExpExecArray | null;
          while ((r = refRe.exec(contentsMatch[2]))) refs.push(parseInt(r[1]!, 10));
        }
        const pageTextParts: string[] = [];
        for (const ref of refs) {
          const streamObj = objects.find(o => o.num === ref);
          if (!streamObj || streamObj.streamStart === undefined || streamObj.streamEnd === undefined) continue;
          const raw = data.subarray(streamObj.streamStart, streamObj.streamEnd);
          const decoded = await decodeStream(raw, streamObj.dict);
          if (!decoded) continue;
          const streamStr = decoded.toString('latin1');
          const text = extractTextFromStream(streamStr);
          pageTextParts.push(text);
        }
        const joined = pageTextParts.join('\n').replace(/\n{3,}/g, '\n\n').trim();
        if (joined.length < 20) imagesOnly++; // probably a scanned page
        pageText.push(`\n--- Page ${pageNum} ---\n${joined || '[no extractable text — likely scanned image]'}\n`);
      }

      if (imagesOnly === pageObjects.length && pageObjects.length > 0) {