Environment

• OpenClaw version: 2026.6.9
• clawpdf version: 0.3.0

Problem
When document-extract processes a scanned/image-based PDF with multiple pages, text extraction returns empty, so it falls back to:

pdf.extract({ mode: "images", image: { maxPixels: 4000000, ... } })

clawpdf appears to enforce a hard 4,000,000 pixel render budget across the entire extraction. On multi-page 300dpi scans, later pages exhaust the budget and are rendered as 1x1 pixel images. This causes downstream vision models (e.g. dashscope/qwen-vl-ocr) to fail with:

InternalError.Algo.InvalidParameter: The image length and width do not meet the model restrictions. [height:1...]

Reproduction

1. Send a multi-page scanned PDF (e.g. 13 pages, 300dpi) to OpenClaw.
2. document-extract extracts 0 characters of text.
3. Fallback mode: "images" returns 1x1 images for later pages.
4. Vision model call fails.

Workaround
Replace the batch pdf.extract({ mode: "images" }) fallback with per-page rendering:

for (const pageNum of pageIndices) {
  const page = pdf.page(pageNum);
  const bytes = await page.png({ dpi: 150 });
  // ...
}

This avoids the shared budget and produces valid PNGs for all pages.

Suggested fix
Either:

• Fix clawpdf's mode: "images" to reset/reuse the render budget per page, or
• Change document-extract to use per-page page.png() rendering for the image fallback.