grade

Output language

Read INTERFACE_LANG from .course-meta (default en). All user-facing prose — chat output, grade-table commentary, the OCR quality escape-hatch menu — must be in that language. Keep in English regardless: file paths, slash command names (/paideia:grade, /paideia:blind, …), pattern IDs (P1, P2…), YAML keys, LaTeX, OCR engine names (claude, ollama, tesseract), and the grade table's column headers (P#, Pattern, Vars, SymPy, End form, Overall). vision_ocr.py reads INTERFACE_LANG from .course-meta on its own to set the VLM's prose-language rule and the tesseract lang= code, so the bash invocations below don't need to pass it explicitly.

Load skills/vision-ocr/SKILL.md, skills/pdf/SKILL.md, and skills/answer-processing/SKILL.md.

Arguments: $ARGUMENTS

If $ARGUMENTS contains --ocr=<engine>, that overrides the default for this call. Otherwise read OCR_ENGINE from .course-meta in CWD (one line of the form OCR_ENGINE: <engine>). If .course-meta is absent or the key is missing, default to claude.

Target answer file: the non-flag positional in $ARGUMENTS. If no positional, find the most recently modified file in answers/ (not answers/converted/).

Follow the answer-processing skill pipeline:

1. Identify. Is target a .pdf or .md?

   • .pdf → proceed to step 2
   • .md → skip step 2, go to 3

2. Convert PDF → MD. Dispatch on the selected OCR engine:

   2a. claude (default) — native Claude vision, no external model

   STEM=$(basename "answers/<stem>.pdf" .pdf)
   TMPDIR="answers/converted/.tmp-${STEM}"
   mkdir -p "$TMPDIR"
   pdftoppm -r 200 -png "answers/${STEM}.pdf" "$TMPDIR/page"
   
   # Downsize to max 1800px width to keep Read-tool image payloads small.
   # Without this, 200-DPI letter-size pages are ~1700–2200px wide and each page
   # eats ~0.5–1.0 MB of image tokens — fine for 1–2 pages, brutal for 10+.
   # Mirrors the resize step used by /paideia:ingest for lecture/homework scans.
   python3 - "$TMPDIR" <<'PY'
   import sys, pathlib
   from PIL import Image
   MAX_W = 1800
   for p in sorted(pathlib.Path(sys.argv[1]).glob("page-*.png")):
       img = Image.open(p)
       if img.width > MAX_W:
           ratio = MAX_W / img.width
           img.resize((MAX_W, int(img.height * ratio))).save(p, optimize=True)
   PY
   

   This produces $TMPDIR/page-1.png, $TMPDIR/page-2.png, ... (each ≤1800px wide). Now use the Read tool on each PNG in order and synthesize clean markdown yourself, following the transcription prompt contract from skills/vision-ocr/SKILL.md:

   • Prose stays in its original language (English, Korean, etc.) — do not translate.
   • Math as $...$ / $$...$$.
   • Preserve problem numbering (P1, (1), (a), ...).
   • Do NOT interpret or grade — pure transcription.
   • [?] for ambiguous glyphs.
   • Skip crossed-out work.
   • Markdown only.

   Write the synthesized result to answers/converted/<stem>.md with header:

   # Vision-OCR transcription
   
   <!-- SOURCE: <stem>.pdf, claude-vision (native), N pages -->
   
   ## Page 1
   
   <transcription>
   
   ## Page 2
   
   <transcription>
   

   Clean up: rm -rf "$TMPDIR".

   2b. ollama — local Qwen3-VL 8B

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/vision_ocr.py" --engine=ollama \
     "answers/<stem>.pdf" "answers/converted/<stem>.md"
   

   Uses qwen3-vl:8b via ollama. The script reads INTERFACE_LANG from .course-meta in CWD so the prose-language rule in the VLM prompt matches the course's language. Auto-falls back to tesseract on any exception (timeout / ollama down / model missing). Tier is recorded in the file header. See skills/vision-ocr/SKILL.md for details.

   2c. tesseract — explicit, skip ollama

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/vision_ocr.py" --engine=tesseract \
     "answers/<stem>.pdf" "answers/converted/<stem>.md"
   

   Pure pytesseract (eng if the course's INTERFACE_LANG=en, eng+kor if ko — also read from .course-meta). Fastest, lowest fidelity on handwriting.