dsh-plugin-insights

Index / validate / evaluate / analyze the whole DeepSeek Harness (dsh) plugin ecosystem. 全量 dsh 插件索引、评估与分析管线（零依赖 Node）。

产品定位（Product）

dsh-plugin-insights = DSH 插件生态的质量评估数据层（quality data layer）：全量索引 + 客观健康分 + 生态分析 + 开放数据快照。它不是第 N 个插件目录/市场，而是让已有目录（awesome-dsh-plugin）、应用内市场（dsh-market）和 agent 的卡片上有据可依的那个分数。健康分口径与体检工具见 dsh-plugin-health。

产品文档：Vision · Roadmap · 产品规划 · 生态调研证据库

---

What it does: instead of being yet another curated list, this pipeline treats the ecosystem as data —

1. Discover (multi-source): GitHub topic:dsh-plugin / deepseek-harness / dsh-bundle / cordis-plugin searches + curated-list data dirs (awesome-dsh-plugin data/plugins, imsai-sh catalog/plugins) + npm search.
2. Validate → authoritative set: cheap noise filtering first, then real checks — not fork/archived, package.json declares dsh.bundle.patch, patch file committed. Noise is bucketed (data/invalid.jsonl).
3. Evaluate (metadata, full-set): manifest shape, exports["./client"], lib/ artifacts, npm publish state & version drift, docs / bilingual docs, repo age/activity/stars/topics, licenses. (Limited deep clone-based security scans come later, reusing dsh-plugin-health heuristics.)
4. Analyze + publish: aggregate stats (data/analysis.json, data/report.md) and a self-contained static site (site/index.html).

Run

GITHUB_TOKEN="$(gh auth token)" node bin/dsh-plugin-insights.mjs run [--limit N]
# or stage by stage:
npm run discover && npm run validate && npm run analyze && npm run site

Resumable: data/state/done.ids keeps progress; re-running continues where it stopped.

Outputs

file	what
data/candidates-all.jsonl	raw candidates from every source (topic:dsh-plugin full crawl via created-window sharding + supplementary topics/curated/npm)
data/plugins.jsonl	authoritative set (valid plugins + evaluation fields)
data/invalid.jsonl	rejected candidates bucketed by reason
data/analysis.json / data/report.md	aggregates + human report
site/index.html	self-contained static dashboard

Status

• pipeline scaffold + pilot run (discover 3125 raw → validate first 25 → report+site)
• [~] full validation of all repo candidates in progress (resumable; ~2889 repo candidates, batch 900 running; CONCURRENCY+resume ready)
• npm candidates → repo mapping
• limited deep checks (read-only-surface / sanitizers) on top/self plugins
• final snapshot commit + CI daily refresh template
• site polish (charts, gap matrix vs official)

Methodology notes: "authoritative" = passes the manifest gate above; it is a lower bound (plugins that ship only via tarball/GitHub releases without a committed manifest may be missed — flagged as no-signal/no-package.json for review). Searches cap at GitHub's 1000-result limit per query; curated dirs may be truncated by the contents API. Set GITHUB_TOKEN for the 5000/hr budget.