PDF → CAD (SVG + DXF) Converter

Turn a PDF drawing into two deliverables:
- .svg — layered (Inkscape layers: Geometry / Text / Images), 1:1 real-world mm.
- .dxf — DXF R2010, mm units, 4 layers (Geometry / Hatch / Text / Images),
imports cleanly into AutoCAD, 中望, 浩辰, LibreCAD, and most AI CAD tools.

When to use

• User says "把这份 PDF 图纸转成 CAD/DXF/矢量图" or equivalent (zh/en).
• User wants editable vector geometry from a PDF (not a raster trace).
• User is prepping a PDF for AI-assisted CAD redraw / parametric editing.

Inputs

• The source PDF path — usually @-referenced in the message or pasted as text.
  If missing, ask the user for the PDF file.
• Optional output directory (default: same folder as the PDF).

Prerequisites (one-time per environment)

Managed Python venv with pymupdf + ezdxf. Set this up once:

VENV="C:/Users/muxinshu/.workbuddy/binaries/python/envs/default"
PY="$VENV/Scripts/python.exe"
"$VENV/Scripts/pip.exe" install --quiet pymupdf ezdxf

If the venv doesn't exist yet, create it first:

"C:/Users/muxinshu/.workbuddy/binaries/python/versions/3.13.12/python.exe" \
  -m venv "$VENV"

Steps

1. Locate the PDF. Read the path from the user message. Resolve to an absolute path.
2. Run the bundled converter (this skill's convert.py). It does everything:
   vector SVG export via PyMuPDF, layer reorganization, DXF build via ezdxf,
   and the critical extents patch.

bash VENV="C:/Users/muxinshu/.workbuddy/binaries/python/envs/default" "$VENV/Scripts/python.exe" \ "C:/Users/muxinshu/.workbuddy/skills/pdf-to-cad/convert.py" \ "ABSOLUTE_PDF_PATH" [--out OUTPUT_DIR]

3. Validate (quick). Confirm non-zero entity counts printed at the end
   (polylines / hatches / texts / images). The script's DXF extents patched
   line must appear — that's what prevents a blank open.
4. Report. Tell the user the output paths and the key note below.

Outputs

• base.svg — layered vector, open in Inkscape/Illustrator.
• base.dxf — CAD-ready; must sit in the same folder as base_images/
  (29-ish PNGs referenced by relative path). Move both together.
• base_images/ — extracted raster images (photos / renders) referenced by the DXF.

CRITICAL gotchas (why a naive ezdxf export fails)

1. Blank-on-open fix: ezdxf's auditor recomputes $EXTMIN/$EXTMAX on saveas
   and fails for docs with many HATCH/IMAGE, resetting them to ±1e20. CAD apps
   frame the initial view from these → blank screen. The script PATCHES the ASCII
   DXF after save with the real A4-range extents. Never remove that patch.
2. Chinese font: doc.styles.add("SimSun", font=...) — font= must be a
   keyword arg, else the style silently drops and text uses a missing font.
3. Mirrored images: PyMuPDF emits image transforms with negative determinant
   (vertical flip). add_image(size_in_units, rotation) can't express a mirror, so
   the script sets u_pixel / v_pixel vectors directly.
4. Orientation: SVG/PDF are y-DOWN; DXF is y-UP. Flip with (H_PT - y).
   Page height H_PT is read dynamically from the PDF mediabox (don't hardcode).
5. Inherited paint: fill/stroke are often set on ancestor "g" groups, not the
   path. resolve_paint() walks the ancestor chain; SVG default fill is black.
   Skip (255,255,255) fills (background) to avoid a black sheet.

Notes

• Vector curves stay exact (Bézier flattened to polylines at 0.25 mm step in DXF).
• Raster photos are embedded as images, NOT traced — they remain raster by nature.
• If the user later reports "DXF blank", re-check the patched extents line first.