fetch-arxiv-recent

Fetch arXiv papers submitted within a date window, optionally restricted to specific arXiv categories, and keyword-filter the results. Returns records in the same canonical shape as the other three literature fetchers.

Workflow

- [ ] Step 1: Validate inputs (from, to, keywords, optional categories)
- [ ] Step 2: Build the search_query string with date range + categories + keywords
- [ ] Step 3: Page through the API until results exhausted or window endpoint passed
- [ ] Step 4: Parse Atom XML; normalize each entry
- [ ] Step 5: Dedupe by arXiv ID (keep latest version)
- [ ] Step 6: Return matched records + summary

Step 1 — Validate inputs

Required:

• from: YYYY-MM-DD, inclusive
• to: YYYY-MM-DD, inclusive, to >= from
• keywords: list of strings (case-insensitive substring match against title + abstract)

Optional:

• categories: list of arXiv category codes — e.g. ["cs.LG", "cs.CL", "stat.ML"]. If omitted, search all of arXiv. Common groupings the caller may want to expose:

- CS / ML: cs.LG, cs.CL, cs.CV, cs.AI, cs.NE, stat.ML - Math / stats: math.ST, math.PR, math.OC, stat.ME, stat.AP - Quantitative biology: q-bio.QM, q-bio.GN, q-bio.MN, q-bio.NC - Physics-adjacent: physics.data-an, cond-mat.stat-mech

The skill itself does not hard-code these groupings; it accepts whatever categories the caller passes. The orchestrator's source-registry.md can document the groupings the operator cares about.

Reject if window > 31 days (same rule as fetch-preprint-recent — wider windows almost always indicate a misuse).

Step 2 — Build the query string

The arXiv search API accepts a query language with fielded search. Compose:

search_query = (date_clause) AND (category_clause)? AND (keyword_clause)

• Date clause (always include — without it you'll get the whole arXiv history):

`` submittedDate:[YYYYMMDDHHMM TO YYYYMMDDHHMM] ` Use from + 0000 and to + 2359` so the window is fully inclusive of both days.

• Category clause (optional): join categories with OR.

`` (cat:cs.LG OR cat:cs.CL OR cat:stat.ML) ``

• Keyword clause: join with OR. Use all: for full-record search (covers title + abstract + authors, which is what we want for a wide-net keyword scan):

`` (all:"protein language model" OR all:"diffusion model" OR all:gnn) `` Multi-word phrases must be double-quoted. Single words don't need quotes.

Final assembled example:

submittedDate:[202605040000 TO 202605102359]
  AND (cat:cs.LG OR cat:cs.CL OR cat:stat.ML)
  AND (all:"protein language model" OR all:"diffusion model" OR all:transformer)

URL-encode and pass as search_query to:

http://export.arxiv.org/api/query?search_query={URL_ENCODED}&sortBy=submittedDate&sortOrder=descending&start={offset}&max_results=200

sortBy=submittedDate&sortOrder=descending is important — it lets you stop paginating as soon as the result dates fall before from, instead of having to walk the whole result set.

Step 3 — Pagination

arXiv returns 200 records per page (max). After each page:

• If response has 0 entries, stop.
• If the last entry's published date is older than from, stop (the rest are out of window).
• Otherwise, increment start by 200, request again.
• Cap at 10 pages (2,000 records). If hit, surface "window may be over-broad" and return what you have.

Rate limit: arXiv asks for 1 request every 3 seconds. Sleep between paginated requests. Don't hammer.

Step 4 — Parse Atom XML

The API returns an Atom feed. Each <entry> has:

• <id>: full URL like http://arxiv.org/abs/2605.12345v1 — the canonical paper ID is the trailing 2605.12345 (post-2007 format) or arxiv.org/abs/cs.LG/0301001 (legacy format)
• <title>: paper title (may have newlines + indentation; collapse whitespace)
• <summary>: abstract (same whitespace caveat)
• <author><name>: one per author; preserve order
• <published>: ISO timestamp of v1 submission (use this as date)
• <updated>: ISO timestamp of latest version (different if revised)
• <arxiv:primary_category term="cs.LG">: primary category
• <category term="..."/>: list of all categories
• <link rel="alternate" type="text/html" href="..."/>: abstract page URL
• <link title="pdf" rel="related" type="application/pdf" href="..."/>: PDF URL
• <arxiv:doi>: optional, present once paper is published in a journal

If the Atom parser fails, retry the request once with a 5-second backoff. On second failure, log the error to fetch_errors and skip the page.

Step 5 — Dedupe by arXiv ID

The paginated results may include the same paper twice if a v2 was submitted within the window. Keep the highest version per ID.

The arXiv ID alone (e.g. 2605.12345) is the canonical key — strip the v1/v2 suffix from the URL and use that.

Step 6 — Normalize and return

Same canonical record shape as the other fetchers:

{
  "id": "arxiv:2605.12345",                                   // arxiv-prefixed for source clarity
  "title": "...",
  "authors": ["Smith J", "Doe A", ...],
  "abstract": "...",
  "date": "2026-05-07",                                       // YYYY-MM-DD parsed from <published>
  "server": "arxiv",
  "primary_category": "cs.LG",
  "categories": ["cs.LG", "stat.ML"],
  "version": 2,
  "doi": "10.1145/...",                                       // if present, otherwise null
  "url": "https://arxiv.org/abs/2605.12345",                  // abstract page (preferred for digest links)
  "pdf_url": "https://arxiv.org/pdf/2605.12345.pdf",
  "matched_keywords": ["protein language model"]
}

Apply the keyword filter client-side (the API's all: is full-record OR but doesn't preserve the match-keyword info). For each record, check title + abstract against the keyword list, populate matched_keywords, and drop records that match none (the API's recall is broader than the operator's intent — the skill must filter).

Return summary:

{
  "server": "arxiv",
  "window": "2026-05-04/2026-05-10",
  "categories": ["cs.LG", "cs.CL", "stat.ML"],
  "query": "(...full search_query...)",
  "fetched_total": 1240,
  "matched_total": 18,
  "pages_fetched": 7,
  "fetch_errors": [],
  "records": [ ... ]
}

Cache the raw Atom XML responses (one per page) to .cache/{YYYY-WW}-arxiv-{page}.xml.

Common Patterns

Pattern A — CS-only weekly scan: categories=["cs.LG", "cs.CL", "cs.CV", "cs.AI", "stat.ML"]. The default for an ML/CS-leaning watchlist.

Pattern B — Cross-disciplinary scan (CS + quant-bio): categories=["cs.LG", "stat.ML", "q-bio.QM", "q-bio.GN"]. When the operator wants computational-biology preprints from arXiv that bioRxiv may miss.

Pattern C — All of arXiv: categories=None. Maximum recall, maximum noise. Only useful when the keyword filter is very tight.

Pattern D — Track a specific arXiv account: outside this skill's default scope. Add (au:"Smith, J" OR au:"Doe, A") to the search_query as an additional AND clause.

Guardrails

1. Always include the date clause. Without submittedDate:[... TO ...] arXiv returns the entire history of the matching query — millions of records. This is the single most common arXiv-API mistake.
2. Sort by submittedDate descending. Otherwise you cannot bail out of pagination early when the dates fall out of window — you'd have to walk the whole result set.
3. Respect the 3-second rate limit. arXiv enforces it loosely but consistent abuse triggers IP-level throttling. Sleep 3s between paginated requests.
4. Don't trust the <title> and <summary> whitespace. Atom feeds often contain literal \n plus 6 spaces of indentation. Collapse all internal whitespace runs to single spaces before storing.
5. Don't conflate the id URL with the canonical ID. <id>http://arxiv.org/abs/2605.12345v2</id>; — the canonical ID is 2605.12345. The v2 is the version suffix; strip it for dedupe.
6. Don't drop the legacy ID format silently. Pre-2007 arXiv IDs look like cs.LG/0301001. The skill should accept both formats, never assert the modern one. Match ^[a-z-]+\.[A-Z]+/\d+$|^\d{4}\.\d{4,5}$ to validate.
7. Don't dedupe across sources here. A paper on arXiv may also be on bioRxiv (for cross-listing) or have a PubMed publication — that's the calling agent's job, not this skill's.
8. Don't expand keywords automatically. arXiv's full-record all: search is already broad; further synonym expansion silently widens the net. The watchlist is the gate.

Quick Reference