Fetch Paper

Resolves one DOI or arXiv ID to a legal open-access copy (PDF URL, arXiv HTML rendering, landing page) and optionally downloads it for transient reading. This is the fetch layer that find-papers, literature-review, study-exemplars, and draft-related-work build on.

When to use

• The user has a DOI, doi.org URL, arXiv ID, or arxiv.org URL and wants the paper.
• Another skill needs full text for a paper it found (lit review, exemplar study).
• The user asks "is there an open-access version of this paper?"
• NOT for topic/title/venue search — run find-papers first to get the identifier.

Inputs

• Exactly one identifier per fetch: a DOI (10.1145/3589132.3625571, https://doi.org/...) or an arXiv ID (2403.12345, arXiv:2403.12345v2, cs/0309136, or any arxiv.org URL). arXiv-issued DOIs (10.48550/arXiv.*) are auto-converted.
• UNPAYWALL_EMAIL (or CONTACT_EMAIL) env var — a real address; Unpaywall rejects placeholders with HTTP 422. The script prompts if unset.

Process

1. Resolve the identifier. Run:

   python3 scripts/resolve_oa.py <DOI-or-arXiv-ID> --json
   

   The script handles everything deterministic: identifier normalization, Unpaywall best_oa_location lookup, arXiv PDF/HTML URLs, the post-2026 ACM DL open-access fallback for 10.1145/* DOIs, rate limiting, 429 backoff, and response caching under .cache/fetch-paper/. Never hand-roll these API calls.

2. Interpret the result (exit codes: 0 = OA found, 3 = no legal OA copy, 2 = bad input/config, 1 = network failure):

   • Prefer html_url (arXiv HTML) when present — easiest to read directly.
   • Otherwise use pdf_url. Read the meaning of oa_status, version, and license in references/oa-sources.md.
   • version: submittedVersion means a preprint — warn the user that the published version may differ before they quote or cite page numbers.

3. Fetch the full text transiently. Either read the URL directly, or:

   python3 scripts/resolve_oa.py <ID> --download
   

   which saves the PDF to a fresh temp directory (never the repo). Read it, extract what the task needs, then delete it. dl.acm.org blocks scripted downloads — give the user the URL to open in a browser instead.

4. If exit code is 3 (no legal OA copy): relay the script's suggestions — author homepages / institutional repositories (search via find-papers), arXiv by title, the user's library access, or emailing the authors. Never suggest Sci-Hub, LibGen, or any shadow library, and never try to bypass a paywall. Full rules: references/copyright-and-politeness.md.

5. Process, don't persist. Quote at most short excerpts in your analysis. Never write paper text, abstracts, or the PDF into the repository or any committed file. Metadata (DOI, title, BibTeX fields) is always fine.

Output

• A resolution report (human-readable, or JSON with --json): is_oa, oa_status, license, version, pdf_url, html_url, landing_url, source, plus notes.
• With --download: a transient PDF in a temp directory, path printed to stderr.
• Nothing is ever written into the repository.

Guardrails

• One identifier per invocation — never loop this script over a list of DOIs for bulk harvesting (ACM ToU and arXiv policy both prohibit it). For a handful of papers in a session, run it one at a time and let the built-in rate limits breathe.
• Legal sources only; transient processing only; no storage, no redistribution.
• Cite the published DOI, not the preprint, when both exist — and verify every citation with verify-citations before it lands in a bibliography.
• Never submit anything to any system on the user's behalf.

References

• references/oa-sources.md — Unpaywall fields and errors, arXiv URL patterns, ACM DL open access, troubleshooting.
• references/copyright-and-politeness.md — what is safe to keep vs. transient-only, politeness contract.