dsh-market

Two Editions

The DSH ecosystem is growing fast, and plugins & skills are scattered across GitHub — it's hard to know which ones are good, how to install them, and how to combine them into a working environment. DSH Market gathers them all in one place and offers two ways to consume them:

The two editions share only the same plugins.json data (refreshed daily at 06:00 with stars & descriptions) — nothing else. The Web site is a standalone browsing site; the plugin is a standalone cordis plugin. They are independent and optional: you don't need the plugin to use the Web edition, and installing the plugin doesn't affect the Web site.

Demo

Web edition	DSH plugin edition

Requirements

• DSH ≥ 0.1.5 (verified on the latest release 0.1.5-rc.1)
• Node.js ≥ 20

The panel entry adapts automatically: 0.1.5+ uses the standard entry (icon on the left nav rail); older builds fall back to a sidebar-foot button + overlay — still working, but in maintenance mode and removed in 0.5.0.

Quick Start

Web edition

No install needed, just visit:

https://dsh.market/

Install the plugin edition

npx @deepseek-ai/dsh plugin --profile web add @dsh-market/plugin

Restart the harness after installing — the "Plugin Market" entry appears at the bottom of the sidebar.

DSH Plugin Edition

The plugin market lives in your DSH sidebar: up and running in 3 minutes, and it learns what you like over time.

Beginner-friendly · zero learning curve

• A cold-start quiz on first open (pick your common scenarios & plugin types) → instantly get curated picks + for-you recommendations
• No commands to memorize — one-click install on every card (skill/cordis auto-routing, retry & rollback on failure)
• Zero-token passive: no panel open, no resources consumed, never in daily conversations

Personalized recommendations · the more you use, the smarter it gets

• Your profile comes from quiz answers / favorites / GitHub stars / installed plugins — all stored locally
• "For You" updates dynamically as you use it (EMA decay); every recommendation carries a "why" line

Scene recommendations · reads your session

• Manually triggered, reads the current session title & messages (zero token) → suggests what fits "right now"

Chinese intent search · zero token

• Searching 「记事本」 (notepad) also finds plugins described as "笔记" (notes): a built-in 100+ Chinese intent dictionary expands colloquial queries into bilingual recall terms, and results are labeled with the matched intent; the web edition adds a "Chinese categories" chip row for one-click browsing (notepad / todo / screenshot / MCP…)
• Experimental toggle: AI semantic re-ranking (local recall of 60 candidates → LLM re-ranks the top 20 with reasons; off by default to save tokens)

Backup & restore

• Export the installed list / favorites / preferences as JSON in one click; import on another machine merges — installs only what's missing, never touches what came later; backups contain no credentials

AI-assisted install

• Not sure whether to install? Hand it to a DSH subagent: read the README → verify → install; asks you first when configuration is needed

Features

• Continuous collection — scans dsh-plugin / dsh GitHub topics and curated community lists every day, collecting everything (currently 16914 plugins + packs channel)
• Practical 5-dimension scoring — maintenance activity / usefulness / ecosystem heat / convenience / signal quality, fused with a weighted geometric mean; every plugin comes with a "why recommended" explanation
• Chinese experience — auto-generated Chinese summaries and feature tags; Chinese intent search (colloquial queries hit, zero token) & Chinese-category filters
• One-click install — deterministic routing in the plugin edition: git clone for skill plugins, dsh plugin add for cordis plugins; retry & rollback on failure
• AI install — hand it to a DSH subagent that reads the README, verifies, then installs; asks you first when configuration is needed
• Recommendation system — cold-start quiz / beginner-friendly / for-you (profile-based) / scene recommendations (reads the current session context; see DSH Plugin Edition)
• Pack ecosystem — Web "Packs" section + plugin packs tab (browse / entry resolvability checks / install entry); backed by the dsh.pack.json pack protocol (see dsh-bundler)
• Zero-token resident — the plugin runs purely passively; no panel open, no resources consumed

Usage

Web edition

Plugin edition (5-tab panel)

Scoring System

Practical five dimensions (weighted geometric mean, inspired by the StarRadar fusion mechanism; oriented toward "practical & convenient"):

Every plugin carries an explanation (one sentence on why it was scored that way). See the scoring guide.

Listing Mechanism

Positioning: DSH Market is the plugin listing & discovery platform for the DeepSeek Harness ecosystem — it lists only DSH-specific plugins, not generic projects.

The daily 06:00 pipeline scans these sources:

GitHub Actions (daily 06:00 collection + deploy)
  └─ collector (Node, concurrency 10, 24h cache)
       ├─ Sources:
       │    ├─ Topic scan: dsh-plugin / dsh / deepseek-harness-plugin / dsh-bundle / dsh-skill
       │    │    └─ Multi-sort union (stars + updated + created, 1000 each, deduped)
       │    │    └─ GitHub Search API caps a single query at 1000 → union covers long tail (~72%+)
       │    │    └─ Rate limit: Search API 30 req/min → 2.3s pause per page to avoid 403
       │    ├─ Awesome lists ×2 (community-curated)
       │    ├─ dsh-external org
       │    └─ Submission issues in this repo (label: submission / title "[提交插件]")
       ├─ Listing criteria (all must hold):
       │    1. Not fork / archived / official (deepseek-ai/deepseek-harness)
       │    2. Feature detection passes: root SKILL.md (skill type)
       │       or skills/ dir with SKILL.md (skill collection)
       │       or cordis markers (dsh.profile / cordis.patch.yml)
       │       or package.json deps contain cordis keywords (double-checked)
       │    3. No stars / score threshold — passes detection, gets listed
       ├─ Metadata + README: GitHub API (stars / descriptions / install command parsing)
       ├─ Practical 5-dimension scoring + explanation layer
       └─ DeepSeek incremental i18n (only new plugins, saves API cost)
            → data/plugins.json
                 ├─ synced to web/public/plugins.json (shared by Web site & plugin)
                 └─ commit → build → deploy GitHub Pages

Note: adding the dsh-plugin topic only gets you into the candidate pool — the repo must actually be a DSH plugin (SKILL.md or cordis markers) to be listed. Many of the 2000+ topic repos are not DSH plugins (casual tagging / forks); feature detection filters them out automatically.

Directory Structure

├─ collector/   # Data pipeline (Node + tsx): scan → detect → score → i18n
├─ web/         # Web site (Vite + React + TS + Fuse.js)
├─ plugin/
│  ├─ core/     # Plugin core layer (pure Node, zero DSH deps, independently testable)
│  └─ ui/       # Plugin UI layer (cordis Host RPC + browser Client panel)
├─ schema/      # Shared types (DshPlugin / MarketData / PracticalScore)
└─ scripts/     # Tooling (screenshots / data injection / visual review)

Local Development

# Clone & install
git clone https://github.com/2BingLing/dsh-market.git
cd dsh-market
npm install
cp .env.example .env        # GITHUB_TOKEN (required), DEEPSEEK_API_KEY (optional)

# Data pipeline (scan → detect → score → i18n → data/plugins.json)
npm run collect

# Web site
npm run dev -w web          # http://localhost:5173
npm run build -w web        # production build

# Plugin
npm run build -w @dsh-market/core    # core layer
npm run build -w @dsh-market/plugin  # plugin package (lib/index.js + lib/client.js)

Contributing

Submit a plugin (either way works)

1. Add the dsh-plugin topic to your repo — fastest; the daily scan picks it up directly.
2. Open an issue via the submit template — fill in the repo URL / type / one-line description.

What happens after you submit:

Submit issue → next day 06:00 the pipeline extracts the repo → plugin detection
  ├─ Passed (it is a DSH plugin) → listed → the bot auto-replies:
  │     ✅ confirmation + "live next update" note + badge guide → issue auto-closed
  └─ Rejected (not a plugin) → not listed (reopen the issue to ask why)

• Fix data: wrong scores / descriptions / install commands — open an issue or PR

• Author intro: include an "author self-introduction" when submitting — it will be shown in the "Author's words" section on the market detail page (with an author badge), letting you describe your plugin in your own words. Already-listed plugins can submit a [data fix] issue to add or update their author intro / Chinese description / install command (applied the next day; write 作者自述：无 to clear an existing intro).

• Add the listing badge: listed plugin authors can display the DSH Market badge at the top of their README:

  

  Usage: PLUGIN-BADGE.md (Top Rated requires a practical score ≥ 80)

License

MIT