@junma11/dsh-scholar-lab
适用于 DSH 的 Google Scholar 搜索代理：提供结构化论文元数据、JCR 影响因子、完整摘要和 CSV 导出。可作为 profile bundle（npm）安装，也可作为动态 Cordis 插件安装。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @junma11/dsh-scholar-lab@0.1.4Scholar Lab 以 @junma11/dsh-scholar-lab 发布，当前版本为 0.1.4。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
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.
| 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 |
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.
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'sshellservice (default web profile; fetches go throughcurl).
| 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.
| 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 |
# 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
# 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>).
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:
/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.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.
MIT — covers your code in lib/ and this repo's files. The bundled JCR data is
third-party (see above).
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 ofhttps://raw.githubusercontent.com/medfm-flare/dsh-scholar-lab/main/dynamic/plugin.host.jsand use it ascode.host; do the same withhttps://raw.githubusercontent.com/medfm-flare/dsh-scholar-lab/main/dynamic/plugin.client.jsforcode.client. If the network is unavailable, read both files from the installed package'sdynamic/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.