pdf-extractlisted

# pdf-extract Text extraction handles most PDFs on its own. This skill uses a fast local engine for that, then spends vision calls **only** on what it provably missed — figures, scanned pages, and tables it failed to structure. ## When NOT to use this - Merging, splitting, form filling, encryption → `anthropics/skills@pdf`. - A PDF you need one number from and already know the page of → just read it. Requires `uv`. Nothing else; dependencies resolve on first run. ## 1. Convert ```bash uv run <skill-dir>/convert.py FILE.pdf [MORE.pdf ...] ``` One JSON object per document, on stdout. Everything deterministic is now done: text extracted, images written, artifact cached. Exit code is non-zero if any document failed. ```json {"status":"ok","artifact":"/home/…/<sha>-<tag>","cached":false, "doc_md":"…/doc.md","pages_dir":"…/pages","manifest":"…/manifest.json", "pending":[{"id":"p007-render","page":7,"kind":"page_render", "reason":"dense_grid","path":"…/images/p007-render.png"}], "dropped":6,"over_scale_guard":false,"scale_guard":15} ``` Check `status` first. `encrypted` and `unreadable` are terminal — report and move on; the batch continues. Re-running the same PDF returns `cached: true` instantly and costs nothing. **If `over_scale_guard` is true**, tell the user how many calls it wants and get agreement before continuing. An 84-page scan is 84 calls. ## 2. Look at each pending item `pending` is empty → you are done, go to step 3. Otherwise, for each en