Fetch Paper

Ingest an arXiv paper as both a lean LaTeX-source context (for you and other AI agents to read) and a PDF (for humans). One command, two artifacts.

When to use

• The user gives an arXiv URL, arXiv ID, or a paper title and wants it added.
• Phrases like "download this paper", "add to refs", "grab the latex source", "ingest this arxiv link".

What it does

For input <url | id | title>:

1. Determines the arXiv ID:
   • URL or ID → taken directly, no network lookup (/abs/2605.06548, 2605.06548, arxiv:2605.06548v2 all work).
   • Title → the only case that queries the arXiv API (text→id), with 429/503 backoff.
2. Source → refs/<Title>/ — pulls arxiv.org/e-print/<id>, extracts the LaTeX bundle, then strips heavy elements: videos/animations (.mp4 .mov .gif .webm …) and font blobs (.ttf .otf .tfm .woff …). Image & vector figures are kept (.png .jpg .pdf .eps .svg) because graphs and architecture diagrams are valuable context. Text stays (.tex .bib .bbl .sty …).
3. Names intelligently by reading \title{} out of the downloaded source (handles \includegraphics logos, \textcolor, \raisebox, \thanks, \texorpdfstring, etc.). Falls back to the API title, then the ID.
4. PDF → papers/<Title> - <arxiv_id>.pdf (title first, then ID), from arxiv.org/pdf/<id>.

Usage

Run from the directory that owns refs/ and papers/ (e.g. the project root):

python3 ~/.claude/skills/fetch-paper/fetch_paper.py "<url | id | title>"

Examples:

# By URL
python3 ~/.claude/skills/fetch-paper/fetch_paper.py "https://arxiv.org/abs/2407.12282"

# By bare ID (version optional)
python3 ~/.claude/skills/fetch-paper/fetch_paper.py 2510.17206

# By title
python3 ~/.claude/skills/fetch-paper/fetch_paper.py "Chip Placement with Diffusion Models"

Flags

• --refs DIR refs output dir (default ./refs)
• --papers DIR papers output dir (default ./papers)
• --source-only download/strip LaTeX source only, skip PDF
• --pdf-only download PDF only, skip source
• --yes suppress the candidate-disambiguation printout for title searches

Agent guidance

• The script is stdlib-only (no pip installs) and idempotent — re-running overwrites the same paths.
• If dangerouslyDisableSandbox is needed for network access in this harness, use it; arXiv requires HTTPS (http://export.arxiv.org returns empty).
• A URL/ID never calls the arXiv API — only a bare-title input does. So prefer passing the URL or ID to avoid rate limits entirely.
• For a title search, the script prints the top match and any other candidates to stderr. If the resolved title looks wrong, show the candidates to the user and re-run with the correct URL/ID.
• To batch-ingest several papers, loop over inputs; arXiv is politely rate- limited, so add a short sleep between calls.

Conventions

• refs folder name = sanitized official title (path separators removed).
• PDF name = <Title> - <arxiv_id>.pdf (no version suffix on the ID).