dsh-scholar-lab
Google Scholar search agent for DeepSeek Harness (DSH) — publishable as an npm
profile bundle (one-command install into any DSH profile) or usable as a
dynamic Cordis plugin (no install, per-session).
Search Google Scholar for academic papers and get structured metadata — title,
full journal name (when resolvable), JCR impact factor + quartile, authors,
year, cited-by count, abstract-excerpt snippet, PDF link, versions — plus full
abstracts from publisher pages and CSV export of any result set.
Feature comparison
| Capability | This plugin (Scholar Lab for DSH) | Google Scholar Labs |
|---|
| Access | No login needed, works from any session/host | Login required (no anonymous API; also no CORS from third-party pages) |
| Search source | Same underlying Scholar index (classic endpoint) | Same index, but an AI "research mode" on top |
| Structured metadata (title, authors, journal, year, citations) | ✅ Full structured JSON per paper | ✅ (the "familiar Scholar features") but as web UI, not structured data |
| Full journal names | ✅ Resolved from bundled JCR database | Shows venue as Scholar renders it |
| JCR impact factor + quartile | ✅ JCR 2025 (June 2026 release), 22.6k journals | ❌ Scholar doesn't show IFs |
| Full abstract + DOI, volume/issue/pages, publisher | ✅ viascholar_abstract (publisher meta tags) | Labs gives summaries, not structured bibliographic records |
| CSV export (all loaded papers, 22 columns) | ✅ One click → downloads to your laptop | ❌ No bulk/structured export |
| Pagination / "load more" | ✅ Up to 100 papers per search | Page-based browsing, no export |
| Use from chat / automation | ✅ Model toolsscholar_search / scholar_abstract — results feed the agent | ❌ Browser UI only |
In short
- Scholar Labs wins on: understanding research questions — AI-generated overviews, per-paper helpfulness notes, and follow-up Q&A. It's a reading/exploration assistant.
- This plugin wins on: structured, machine-usable data — clean metadata (title/authors/journal/year/citations), full journal names + JCR 2025 impact factors & quartiles (which Scholar never shows), full abstracts + DOIs, bulk CSV export to your laptop, pagination up to 100 papers, and deep integration with the chat agent (search results feed directly into what the assistant can reason about). No login, works everywhere.
They're complementary: use Labs when you want an AI conversation about a research area; use this plugin when you want a reproducible list of papers with bibliometrics you can export and cite-check. The panel even links to your logged-in Labs page, so both are one click apart.
Install (profile bundle — the community path)
dsh plugin --profile web add @junma11/dsh-scholar-lab
dsh web
The dsh plugin command forwards to pnpm inside the profile; any installed
package declaring dsh.bundle automatically joins the profile's bundle layer
stack (dsh.profile.bundles) and its cordis.patch.yml composes into the app
config. Remove with:
dsh plugin --profile web remove @junma11/dsh-scholar-lab
Requires @deepseek-ai/dsh ≥ 0.1.0 and the base profile's shell service
(default web profile; fetches go through curl).
What you get
| Surface | Description |
|---|
scholar_search tool | Chat-side search:query, max_results (1–20), start, year ("2020" or "2018-2023"), lang, export_csv (writes scholar-lab/scholar-export-*.csv into the workspace via the fs service) |
scholar_abstract tool | Full abstract + HighWirecitation_* metadata from a result/publisher URL |
| Search panel (UI) | Run-card panel with search box, pagination ("Load more papers", deduped, ≤100), andDownload CSV (browser-side, nothing written to the host) |
| Result cards | scholar_search / scholar_abstract result views with JCR IF badges |
Per-paper fields: title, authors[], authors_raw, venue, journal_full, impact_factor, jif_quartile, publisher, year, cited_by, snippet, link, pdf, versions, cluster_id, cites_id, result_id.
Package layout
| File | Contents |
|---|
lib/host.js | Packaged host half:scholar_search / scholar_abstract tools (ctx.tools.register + defineTool from @deepseek-ai/dsh-tools), the /scholar-lab/search + /scholar-lab/export webServer routes, CSV writer |
lib/client.js | Packaged client half: browser bundle in the web-shell module-table format (window.__ModuleLoader__.load), search panel + tool result cards (talks to the host over the HTTP routes) |
data/journal-impact-factors.json | JCR 2025 database (full names, IFs, quartiles) — 22,643 journals, 1.5 MB, read from the package viaimport.meta.url |
cordis.patch.yml | Bundle patch layer: inserts thescholar-lab host row (the browser row comes from the package's dsh.client manifest) |
dynamic/plugin.host.js | Dynamic Cordis plugin form (host half) — the zero-install variant; uses harness.* and host.call, no npm needed (see below) |
dynamic/plugin.client.js | Dynamic Cordis plugin form (client half) — host.call('scholar-search' / 'scholar-export') bridge |
Develop & verify locally (before publishing)
# 1. Syntax / structure checks (dynamic halves are `return { … }` bodies —
# wrap them in a function for node --check)
node --check lib/host.js
node --check lib/client.js
node -e "new Function('return async () => {' + require('fs').readFileSync('dynamic/plugin.host.js', 'utf8') + '}')"
node -e "new Function('return async () => {' + require('fs').readFileSync('dynamic/plugin.client.js', 'utf8') + '}')"
# 2. Pack and install into a throwaway profile (or your web profile)
pnpm pack
dsh plugin --profile web add ./junma11-dsh-scholar-lab-0.1.0.tgz
# 3. Confirm the layers compose without booting
dsh --profile web --dump-config | grep -A3 scholar-lab
# 4. Boot and test
dsh web
Publish to npm
# fill in package.json: name (real scope), repository, homepage, license holder
npm login
npm publish # scoped packages publish private by default — `publishConfig.access: "public"` is already set
Version semver; tag releases; attach the packed tarball to GitHub releases for
air-gapped installs (dsh plugin --profile web add <tarball-url>).
Notes
The packaged API was mapped from the in-box bundles (@deepseek-ai/dsh-base,
@deepseek-ai/dsh-web-app, @deepseek-ai/dsh-tool-web,
@deepseek-ai/dsh-client-ui-cordis). Two points should be confirmed against the
running app before publishing a public release:
- Browser → host bridge. The client bundle talks to the host over the
plain same-origin HTTP routes the host half registers (
/scholar-lab/search
and /scholar-lab/export, same-origin enforced on POST) — the same pattern
dshmarket uses. ctx.remote.* typert namespaces are first-party only and
must not be used by community client bundles. If you rename the routes,
change lib/host.js and the callHost calls in lib/client.js together.
- Slot keys. The client injects into
tool.view.cordis (panel) and
tool.call.toolview (result cards, keyed by tool name). If the packaged slot
names differ from the dynamic runner's, adjust the slots.inject(...) calls
in lib/client.js.
If a seam turns out different, fix the one file and bump a patch version —
nothing else in the package changes.
Data & legal notes
- The journal database is from hitfyd/ShowJCR v2026-1.2.
- Scholar's ToS / rate limits: results and abstracts are cached in memory
(search 10 min, abstract 30 min); a CAPTCHA/"unusual traffic" response is
returned as a clear error. Document this behavior for your users.
License
MIT — covers your code in lib/ and this repo's files. The bundled JCR data is
third-party (see above).
Zero-install alternative: dynamic Cordis plugin
No npm and no local files needed — the dynamic form is versioned in this
repository (dynamic/plugin.host.js / dynamic/plugin.client.js, kept in
sync with the packaged lib/). Create it in any session by pasting this into
chat:
Create a Cordis plugin with kind: "new", idPrefix: "schlr", name "Google
Scholar Lab", purpose "Search Google Scholar for papers with full journal
names, JCR impact factors, and CSV export of paper metadata". Use curl to
fetch the content of
https://raw.githubusercontent.com/medfm-flare/dsh-scholar-lab/main/dynamic/plugin.host.js
and use it as code.host; do the same with
https://raw.githubusercontent.com/medfm-flare/dsh-scholar-lab/main/dynamic/plugin.client.js
for code.client. If the network is unavailable, read both files from the
installed package's dynamic/ folder instead.
Optional — JCR impact factors (the plugin degrades gracefully without them):
fetch
https://raw.githubusercontent.com/medfm-flare/dsh-scholar-lab/main/data/journal-impact-factors.json
into the working directory as journal-impact-factors.json.
Dynamic plugins are per-session and in-memory — they disappear when the
process restarts.
