dsh-skill-recommender
A skill recommender for DeepSeek Harness. It browses your local session records across Codex, Claude and DSH, builds a compact user profile, ingests an open-source skill catalog, and ranks skills with a weighted score gated by a tunable match index — the higher the index, the higher-relevance the recommendations only.
What it does
- Reads sessions from three agent sources — DSH (
~/.dsh/sessions/**/session.jsonl.zstd), Codex (~/.codex/*.jsonl), Claude (~/.claude/projects/**/*.jsonl) — no API keys, purely local file reads.
- Builds a profile: topic distribution, high-frequency tools, task types, common project directories, language/model source.
- Ingests a skill catalog: local skill dirs (
~/.agents/skills, ~/.dsh/skills, your Obsidian 2️⃣ AI/Skill) + remote awesome lists (awesome-dsh-skills, awesome-dsh-plugin, awesome-deepseek-harness, Claude skills ecosystem), plus a small bundled seed so it always returns something.
- Scores & ranks with a weighted model:
score = Σ(w_i × sim_i) / Σ(w_i), where dimensions are topic / tool / task / proximity-to-installed. A global match index (0–100, default 60) is the gate: only skills with score ≥ index are returned, then top-N.
Why it's different from dsh-skill-studio
dsh-skill-studio extracts reusable skills from your own sessions. This plugin recommends third-party open-source skills by extrapolating your profile. They complement each other.
Tools (model-facing)
| Tool | Purpose |
|---|
recommender_scan | Scan sessions, build the profile, produce an initial recommendation run. |
recommender_recommend | Recommend open-source skills; optionally override index / topN / weights on the fly. |
recommender_profile | Show the current user profile (topics, tools, tasks, projects). |
recommender_config | Configure sources, catalogs, window, match index, per-dimension weights, optional LLM enrichment. |
recommender_status | Show plugin status without leaking secrets. |
Web settings panel
A Skill 推荐器 card in the Web settings page: scan button, live 匹配指数 slider (0–100), four per-dimension weight sliders (topic / tool / task / proximity), catalog toggles, and recommendation cards with scores, per-dimension breakdowns and a GitHub link. The index + weights are saved to ~/.dsh/dsh-skill-recommender/config.json (mode 0600).
Compatibility
Requires DeepSeek Harness ≥ 0.1.5-rc.1 (declared as dsh.engines.dsh in the package manifest, so the DSH plugin marketplace can report it) and is verified against 0.1.5-rc.1. This build carries the DSH 0.1.5 adaptations: the strict tool-result contract (lossless-JSON snapshot, additionalProperties: false schema validation, and output.render returning ContentBlock[]) plus executable resolution that survives a launchd-started host whose PATH is only /usr/bin:/bin.
Install (development)
dsh plugin add --profile web link:/path/to/dsh-skill-recommender
Then restart the host (tools + routes) and hard-refresh the browser (client panel). Config key: skill-recommender in the bundle patch layer.
Build
pnpm install
pnpm bundle # builds lib/index.js (ESM) + lib/client.js (browser bundle)
node tests/smoke.mjs
Notes
- The default weight model is
{topic: 50, tool: 50, task: 35, near: 40}; the index gate defaults to 60.
- Whole-web discovery: the default
github-discovery source queries the GitHub search API across the entire platform (agent/claude/codex skills, SKILL.md, awesome skills — 星标排序), so recommendations are not limited to one designated list. Results are cached (6h TTL); the unauthenticated search limit (10 req/min) is respected.
- Skills vs plugins: each catalog is tagged
skill or plugin. By default only skills are recommended (types: ['skill']); tick 含插件 in the panel to also include DSH plugins. Both markdown lists and tables are parsed (DSH skill catalogs use tables).
- Remote catalogs are fetched with a 12s timeout and cached (6h TTL); offline runs fall back to the cache + bundled seed.
- Background auto-scan: set
autoScanMinutes (default 60, 0 = off) in the panel — the host refreshes the cached result in the background, and the panel shows the last result instantly on open. The cache is persisted to ~/.dsh/dsh-skill-recommender/last-result.json (0600).
- No secrets are stored or echoed; LLM enrichment (optional) reuses OpenAI-compatible config, keys never returned.
License
MIT
安装 / Install
# from npm (published package)
dsh plugin --profile web add dsh-skill-recommender
# or local development
dsh plugin --profile web add link:/path/to/dsh-skill-recommender
# then restart dsh web to activate
