dsh-paper-search

dsh-paper-search

Literature search for DeepSeek Harness: twenty-two sources, all native HTTP — no external CLI, no Python package. Twenty of them need no credential at all.

Registered as the paper-search skill and five model tools (paper_search, paper_status, paper_citations, paper_oa, paper_fulltext).

Sources

International (14) — official JSON APIs, except IACR which is parsed HTML:

id	Source	Notes
crossref	Crossref	DOI metadata across publishers, broadest coverage
acm	ACM Digital Library	ACM's own output, keyless — Crossref restricted to DOI prefix 10.1145 (upstream PR #123)
openalex	OpenAlex	metadata + abstracts; an email raises the rate limit
doaj	DOAJ	open-access journals — includes many Indonesian journals
europepmc	Europe PMC	biomedicine, full text for OA records
pubmed	PubMed	biomedicine; two calls per search (esearch + esummary)
pmc	PubMed Central	open-access full text; two calls per search
arxiv	arXiv	preprints; the plugin self-paces to 1 request / 3 s per arXiv's TOU
semantic	Semantic Scholar	a free key raises the rate limit
core	CORE	open-access repository aggregator — CORE_API_KEY required
zenodo	Zenodo	general repository: data, software, preprints
hal	HAL	French open-access repository
iacr	IACR ePrint	cryptology preprints
unpaywall	Unpaywall	DOI input, not keyword — resolves a DOI to its open-access copy; email required

Indonesian (3) — parsed from public pages:

id	Source	Notes
garuda	Garuda (Kemdiktisaintek)	Indonesian journal articles
sinta	Sinta (Kemdiktisaintek)	accredited journals, not articles
ios	IOS OneSearch (Perpusnas)	Indonesian repository aggregator

sources: "all" and the DOI-input source

unpaywall answers only when the query is a DOI. So when you search with keywords, it is skipped rather than run and reported as an empty result — and the output says so:

CATATAN: unpaywall dilewati karena query ini bukan DOI — sumber itu memang
mencari dengan DOI, bukan kata kunci.

Pass a DOI and it runs. Pass sources: "unpaywall" with keywords and it runs too, returning nothing — because that is genuinely how it behaves, and the output says that rather than pretending the source is broken.

Deliberately excluded, with the reason

Every exclusion was measured on 2026-09-28. An always-empty source is worse than an absent one, so these are dropped rather than registered to look larger:

Source	Reason
moraref	landing page only, no parseable result markup (3,964 characters)
dblp	API is behind bot protection — returns "Making sure you're not a bot!", not JSON
openaire	test request timed out at 30 s and again at 40 s
biorxiv, medrxiv	their API is date-range based, not keyword search
base	OAI-PMH requires institutional IP registration (Access denied for IP address …)
citeseerx	returns HTTP 404, then hangs until the request is killed
ssrn	rejects ordinary requests with HTTP 403
google_scholar	bot detection active; a proxy is required

Install

# from npm
dsh plugin --profile <profile> add dsh-paper-search

# or straight from the repository
dsh plugin --profile <profile> add github:rasyidmmz/dsh-paper-search

The package ships a dsh.bundle.patch, so DSH installs its loader row automatically. Do not also write an id: paper-search row by hand — the loader throws duplicate loader entry id rather than warning.

Restart DSH after installing.

Plugin options

Option	Default	Meaning
tools	"five"	"five" registers the five separate tools. "one" registers a single tool named paper with a mode parameter (search · status · citations · oa · fulltext).
anydocPath / markitdownPath	auto-detected	Explicit path to the PDF→Markdown converter used by paper_fulltext.
folderTeksLengkap	~/.dsh/paper-search	Where downloaded PDFs and converted Markdown are kept.

Why tools: "one" exists. Every tool definition is sent to the model on every request. Measured 29 Sep 2026 (test/uji-satu-pintu.mjs):

	characters of JSON	≈ tokens
"five" (default)	4,959	~1,240
"one"	2,159	~540
saved	56%	~700 per request

What it costs, stated plainly. Tool names stop describing themselves (paper_fulltext no longer exists as a name; it becomes paper({ mode: "fulltext" })), and one JSON schema cannot require different arguments per mode — so per-mode required-argument checking moves from the harness into the plugin. paper({ mode: "oa" }) without an identifier is rejected by the plugin, with a message that mirrors a schema rejection, instead of by DSH before the call. That is why the default stays "five".

An unrecognised value falls back to "five" and says so in the plugin's startup note — it is never applied silently. The option is read from the plugin's config in your DSH profile; the shipped skill adds a short note to its own body when "one" is active, so the documentation never points at a tool name that does not exist in that profile.

Use

paper_search({ query: "machine learning" })
paper_search({ query: "pendidikan karakter", sources: "indonesia" })
paper_search({ query: "CRISPR", sources: "pubmed,europepmc,doaj", max_results: 10 })
paper_search({ query: "digital literacy", open_access_only: true, year_from: 2020 })
paper_search({ query: "eye tracking", sources: "acm" })           # ACM Digital Library only
paper_status()                    # which sources answer right now
paper_status({ probe: false })    # configuration only, no network
paper_citations({ identifier: "10.1145/3290605.3300233" })                  # who cites it
paper_citations({ identifier: "10.1145/3290605.3300233", arah: "rujukan" }) # what it cites
paper_oa({ identifier: "10.1371/journal.pone.0266789" })                    # free copy?
paper_fulltext({ identifier: "1706.03762" })                                # read the whole paper

For Indonesian topics, use Indonesian keywords: "pendidikan karakter" returns real Indonesian journals, while the English equivalent returns international results that do not match the intent.

Citation graph and open-access lookup

Two tools use OpenAlex — free, no key, no email (upstream PR #96):

• paper_citations walks the citation graph in either direction. arah: "pengutip" (default) lists papers that cite the work — forward snowballing; arah: "rujukan" lists the works it cites — backward snowballing. The input may be a DOI, an OpenAlex id (W…), or a title.
• paper_oa reports the open-access status (gold/green/hybrid/bronze/closed) and every OA copy with its PDF/landing link, licence, and version. It does not need the Unpaywall email that the unpaywall source requires.

Three limits are reported rather than hidden:

1. Title input can resolve to the wrong record. Measured 2026-09-29: the title "Attention is all you need" resolves to W2626778328, a duplicate record with an unrelated DOI. Every output therefore names the work it actually used — check that before drawing conclusions.
2. Backward snowballing reads at most 100 references per call, so "most cited first" applies only among those examined.
3. A source that fails says so. paper_citations/paper_oa translate OpenAlex's 429/503/504 into an actionable sentence instead of a bare status code.

This plugin does not download PDFs for paper_citations/paper_oa — those two only report what exists. Downloading is paper_fulltext's job, below.

Full text: open-access PDF → Markdown

paper_fulltext fetches the full text of a paper so an agent can read it, not just its title and abstract. It accepts a DOI, an OpenAlex id (W…), an arXiv id (1706.03762), or a title.

paper_fulltext({ identifier: "1706.03762" })                      # arXiv id
paper_fulltext({ identifier: "10.1371/journal.pone.0266789" })    # OA DOI
paper_fulltext({ identifier: "Attention Is All You Need" })       # title
paper_fulltext({ identifier: "…", paksa_ulang: true })            # force re-download

How it finds the PDF — arXiv first, which is upstream PR #124:

1. the work's own arXiv id (from 10.48550/arXiv.<id> or an arXiv location), or else the arXiv version looked up by title;
2. every open-access PDF location OpenAlex reports;
3. Unpaywall, last, using the configured email.

Then the PDF is saved next to its Markdown in ~/.dsh/paper-search/md/ (override with the folderTeksLengkap plugin option), converted locally by anydoc — fallback markitdown — and the tool returns the file paths, the source it used, and either the whole text (when ≤ 20,000 characters) or a preview plus the read({ file_path, offset, limit }) call to continue with.

Measured 2026-09-29: Attention Is All You Need (2.1 MB PDF) → 39.8 KB of Markdown in 424 ms; a PLOS ONE paper (2.9 MB) → 70.8 KB in 363 ms, with reference links intact. End to end through the tool, including download and the arXiv lookup: 2.4–4.4 s, and 0.3 s when the file already exists.

Install a converter (neither is bundled):

npm i -g @firecrawl/anydoc     # recommended: Rust binary, no Python
uv tool install markitdown     # fallback

paper_status reports which converter was found and where — before you hit a failure.

Limits that are stated, not hidden

1. Open-access copies only. If no legitimate free copy exists, the tool fails and says so, listing the candidates it tried. It never bypasses a paywall and never uses Sci-Hub or Anna's Archive.
2. Conversion is always local. anydoc has an --ocr hosted mode that uploads the document to Firecrawl; this plugin never uses it.
3. Title lookups can pick the wrong work. Measured: the title "Attention is all you need" resolves in OpenAlex to W2626778328, a duplicate record. Every output names the work it actually used — check it before trusting the result. A DOI or an arXiv id is safer.
4. Files are never deleted for you. The library grows ~2–3 MB per paper (PDF + Markdown).

Verification status

Stated precisely, because "tested" on its own is not a useful claim:

Verified — the connectors, against live APIs and pages. node test/uji-sumber.mjs runs all twenty-two against the real services: 79/79 assertions, 1 skip. On the last run 21 of 22 answered; the one that did not was ios (IOS OneSearch) with a network-level fetch failed. That was verified not to come from this release: the previously installed build fails on the same source in the same way, so the site was simply unreachable — and the test now reports that as a skip rather than a failure. Source acm answered with real ACM records in 670 ms. The sources added in 0.2.0 and 0.3.0 answered with real records too — openreview in 1.8 s (with DOIs recovered from BibTeX, so its records merge with Crossref on deduplication), ssrn in 2.3 s, ieee in 1.6 s, biorxiv in 1.0 s, and medrxiv in 0.9 s.

Verified — the wiring, against a mock Cordis context. node test/uji-plugin.mjs, 57/57 assertions: all five tools register, the new tools are actually called and return data, paper_fulltext downloads and converts a real paper with the resulting .md checked on disk, the skill provider reports rank 450, the frontmatter parses within the 500-character catalogue limit, and every tool declares the output: { schema, render } that DSH requires.

Verified — loading the bundle inside a live harness. node test/uji-di-dsh.mjs copies the headless profile to a throwaway test profile, inserts this plugin plus a witness plugin, boots a real DSH (0.2.0-rc.1), and checks from inside that all five tools are registered, the skill provider is registered, the skill appears in the catalogue, and paper_search is actually called and returns real papers. 27 assertions, 0 failures. The witness records schemas=Array(15): paper_search, paper_status, paper_citations, paper_oa, paper_fulltext, ….

That last test earned its place: it caught a fatal bug the mock never could. DSH requires every tool to declare output: { schema, render }, and 0.1.0 did not — so registration threw, and because the throw happened inside apply(), the whole plugin failed to load. 0.1.0 is unusable; use 0.1.1.

What makes the reporting honest

Every result separates sources that failed from sources that answered:

  v Crossref            3 results  (2316 ms)
  x Semantic Scholar    FAILED — HTTP 429 — {"message": "Too Many Requests...
  v Garuda              3 results  (1428 ms)

A source that fails is never allowed to masquerade as "no results". This is not a detail: a widely-used tool in this space returns an empty list for every non-200 response with no error recorded at all, so a rate-limited request reads as "this literature does not exist". Here, the failure and its HTTP status are printed, and paper_status exists to prove liveness on demand.

Credentials (all optional)

No key is bundled. Read order:

1. plugin config in your DSH profile,
2. environment variables PAPER_SEARCH_MCP_<NAME>, then <NAME>,
3. the file ~/.config/paper-search-mcp/.env (belonging to the upstream CLI).

Name	Purpose
UNPAYWALL_EMAIL	any email — used as the Crossref/OpenAlex "polite pool" contact
OPENALEX_EMAIL	same, OpenAlex only
SEMANTIC_SCHOLAR_API_KEY	raises Semantic Scholar rate limits
DOAJ_API_KEY	raises DOAJ's hourly limit
CORE_API_KEY	required for the core source — without it CORE answers HTTP 429 with an empty body

Twenty of the twenty-two sources work with no credential at all; keys raise rate limits and stability. Two are exceptions, and each is reported as a clear, actionable failure rather than a confusing 429:

• core needs CORE_API_KEY.
• unpaywall needs an email (any real address) — it also needs a DOI as input.

What this package does not do

• It does not bypass paywalls or access controls, and never uses Sci-Hub or Anna's Archive. paper_fulltext downloads only copies that are legitimately open access; downloaded files stay on your machine and nothing is redistributed.
• It does not upload your documents anywhere. PDF→Markdown conversion runs locally through anydoc/markitdown; anydoc's hosted-OCR mode is never used.
• It does not call any service other than the sources listed above.
• It does not require or install paper-search-mcp. If that CLI happens to be installed, you can use it for its additional sources (CORE, Zenodo, HAL, SSRN, Unpaywall, OpenAIRE, CiteSeerX, BASE) — this plugin never depends on it.

Licence and attribution

MIT — see LICENSE. Attribution and source terms are in NOTICE.md.