description: "Semantic vector search over a 6,097-skill local corpus (skills.sh top 100 + the a5c-ai/babysitter library), installed as a DeepSeek Harness bundle"
kind: "package-reference"

dsh-awesome-skills

A DeepSeek Harness bundle that gives agents semantic access to a curated local
skill corpus — the skills.sh top 100 (83 GitHub-hosted skills) plus the full
a5c-ai/babysitter library (~2,100 skills) — 6,097 skills total — without
ever putting that corpus into the per-turn model catalog.

Why

A skill directory that DeepSeek Harness discovers becomes a catalog entry, and
every catalog entry is injected into the model's context on every turn. A
large corpus there is a very large per-turn token bill for something almost
never needed on a given turn.

This plugin inverts that. The corpus stays out of the catalog; a single small
skill-router skill is installed instead, and it teaches the agent to search
the corpus on demand.

The corpus

The corpus is organized by source, one directory per owner/repo, with each
skill directory carrying its complete content — SKILL.md plus every
reference file it points at (examples, templates, scripts):

skills/
├── a5c-ai/babysitter/            # ~2,100 skills (specializations + methodologies incl. the domains tree)
├── mattpocock/skills/            # 17 skills (tdd, grilling, code-review, ...)
├── microsoft/azure-skills/       # 20 skills (incl. microsoft-foundry: 191 files)
├── anthropics/skills/frontend-design/
├── vercel-labs/agent-skills/     # vercel-react-best-practices (62 rules)
├── obra/superpowers/             # brainstorming, systematic-debugging, ...
└── ...                           # 157 owner/repo groups, 6,097 skills, 17,600+ files

The 17 site-only entries on the skills.sh leaderboard (the open.feishu.cn
lark suite and similar, which have no public repository) are excluded — there
is nothing to fetch from.

The index ships prebuilt and is committed. Skill bodies live in the
canonical corpus directory (~/.dsh/awesome-skills/skills), referenced — not
copied — by the package: search results return paths into it, so there is
exactly one copy of the corpus on disk.

Install

dsh plugin --profile web add github:ryasrk/dsh-awesome-skills

The plugin mounts as a cordis bundle (see cordis.patch.yml) and, on apply,
installs the bundled skill-router skill into ~/.agents/skills/skill-router.
The bundled skill-router is refreshed on every apply — the shipped copy is
the source of truth for its behaviour.

Model-facing tools

On hosts that expose the tools service, the plugin registers two tools that
run in-process with host authority — the standard-permission-mode path to the
corpus, since the agent needs no Bash or out-of-workspace Read:

• skills_search(query, k?) — the calibrated hybrid search; settings knobs
  (prio/blacklist/whitelist) apply.
• skills_read(path, file?) — one file from a hit's directory, guarded to the
  corpus root (no traversal), text extensions only, 64 KiB cap.

The skill-router skill teaches the tool-first flow and keeps the node lib/query.js CLI (JSON on stdin → JSON on stdout) as a shell-agnostic
one-paragraph fallback for hosts without the tools service — it reads stdin and
writes stdout, so no bash-specific syntax is required on any shell (bash,
PowerShell, cmd).

Ranking

Three lanes are fused, then re-ranked over a candidate pool:

score = (1 - WEIGHT) * semantic + WEIGHT * lexical + GRAM_WEIGHT * char-3-gram

• Semantic — MiniLM cosine, brute force over all rows. Exact; no
  quantization drift.
• Lexical — IDF-weighted token overlap.
• Char 3-gram — script-agnostic, so CJK/Cyrillic queries and technical
  identifiers still discriminate.

WEIGHT 0.55, GRAM_WEIGHT 0.5, pool 1200. Re-checked on the shipped
corpus with a 150-label canonical set: R@1 80%, R@3 93%.

Speed

Measured on the shipped 6,097-skill corpus (brute force still scores every
row; the per-process model load dominates):

Derived caches (per-skill char grams, query embeddings) live next to the
corpus and are keyed by a corpus fingerprint, so a corpus change invalidates
them automatically.

Service surface

Other plugins and tools can use the search service directly:

const search = ctx.get('skills-search')
const hits = await search.search('set up end-to-end browser tests', 5)
const dir = search.skillDir(hits[0].path) // e.g. .../skills/mattpocock/skills/tdd

A hit's path is the subpath under the corpus root (owner/repo/skill), and
every consumer joins corpusDir + path + /SKILL.md — the priority loader, the
skill-router template, and the settings UI all share that one convention.
Reference files live beside the SKILL.md, so a hit's directory is the whole
playbook.

Configuration

All fields optional, via a cordis.patch.yml row:

Environment: DSH_AWESOME_SKILLS_CORPUS (corpus), DSH_AWESOME_SKILLS_INDEX
(index directory for the CLI), DSH_AGENTS_HOME (agents home; resolved the same
way the harness skill provider does, so a relocated agents home still receives
the router skill).

Rebuilding the index

The index ships prebuilt in skills/skills.json and skills/vectors.f32; a
user never needs to rebuild it. Rebuilding after changing the corpus is a
maintainer step done in the separate corpus-ingestion workspace (the walker
lives there, not in this package) — it reads the corpus recursively and rewrites
skills.json and vectors.f32, which are then copied back into this repo's
skills/. If a deployed profile carries a stale index, re-sync this repo's
lib/ and skills/ directories into the profile's installed copy.

A skill directory is any directory holding a SKILL.md; directories nested
inside one (a sub-skill shipped as reference material) are not indexed
separately. Regenerate the client bundle after pulling source changes:
npx tsdown -c tsdown.client.ts, then run node scripts/preflight.mjs.

License

MIT