dsh-skill-router
适用于 DeepSeek Harness 的 RAG 风格技能路由器：对技能进行嵌入，通过相似度匹配用户任务，并自动注入相关技能正文。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:lau4tin1/dsh-skill-router#cb3d6c2678f8eda79132e0bb5e8796bd7ac6335dSkill Router 以 dsh-skill-router 发布，当前版本为 0.2.0。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。
RAG-style skill routing plugin for DeepSeek Harness: embeds skills (their
name + description + whenToUse routing surface), embeds each user task, keeps
only the clearly-relevant skills with a gap-based selection rule, and
auto-injects their full bodies into the prompt — no model round-trip to load a
skill.
0.1.5-rc.2 types, builds, mounts in a live DSH profile (verified in the
web and headless profiles), and passes 20/24 bilingual routing prompts on
real skills (node demo/e2e.ts). Compatible with DSH 0.1.5-rc.2 through
0.2.0-rc.2 (peer range >=0.1.5-rc.2 <0.3.0-0); all routing APIs it uses —
ctx.skills.snapshot/get, skills/change, renderSkillContent,
createUserMessage, and the agent/pre-step middleware — are unchanged in
0.2.0-rc.2.dsh plugin --profile web add dsh-skill-router
Installs into the web profile; swap web for another profile name to target
it. Requires dsh and pnpm on
PATH.
pnpm 12 build approval — the first install pulls native dependencies (
onnxruntime-node,sharp,protobufjs) whose build scripts pnpm blocks by default. Ifdsh pluginstops withERR_PNPM_IGNORED_BUILDS, open$DSH_HOME/profiles/<name>/pnpm-workspace.yaml, set theallowBuildsentries pnpm wrote there totrue, then re-run the samedsh plugin ... addcommand.The first real task downloads the embedding model once (~543MB, q8) into
$DSH_HOME/skill-router/models, then runs fully offline.
src/index.ts — the Cordis plugin entry (name / inject / Config / apply).src/selection.ts — the selection rule (largest-gap / ratio-to-max, weak floor).src/similarity.ts — cosine similarity + vector helpers (pure).src/embedding.ts — one real embedding model via transformers.js (ONNX).src/skillIndex.ts — the index: build, embed, digest-diff sync on change.src/config.ts — schemastery config schema + mappers.src/render.ts — renders the injected <system-reminder> block (reuses DSH's renderSkillContent).demo/e2e.ts — end-to-end routing test over downloaded skills.npm install # toolchain (typescript, esbuild); DSH types come from a local checkout symlink
npm run typecheck # tsc --noEmit against the real DSH types
npm run build # esbuild -> lib/index.js (@deepseek-ai/* kept external)
Skills live in .agents/skills/ — a real DSH scanned root (rank 200, see
DESIGN.md §14). The demo parses them the same way the registry would surface
them:
node demo/e2e.ts # plugin defaults
node demo/e2e.ts 0.2 # experiment with the minScore floor
enabled: true
embedding:
model: Xenova/bge-m3 # optional; any HF ONNX embedding model
dtype: q8 # or fp32
cacheDir: ~/.dsh/skill-router # optional; models + index live here
rule: largest-gap # or ratio-to-max
minScore: 0.2 # confidence floor (precision > recall)
ratioThreshold: 0.75
maxSkills: 4
The plugin default above is classic gap/ratio mode. The bundled
cordis.patch.yml ships with the scale-invariant override instead:
minScore: 0.05 and zThreshold: 2.5.
Chinese and English both work out of the box — the default model is
multilingual, and a Chinese prompt matches an English skill description (and
vice versa) by meaning, cross-lingual. One DSH registry rule to respect:
skill names must stay kebab-case ASCII (pdf-tools), but description,
whenToUse, and the body can be Chinese. The injected instruction block is
bilingual (中文/English).
One engine: a real local model via transformers.js (ONNX) — dense semantic
vectors; synonyms and paraphrases match, word sharing is not required.
Default Xenova/bge-m3 (q8, ~543MB download, 1024-dim, [CLS] pooling) —
flagship multilingual quality for Chinese + English. Alternatives via
embedding.model: multilingual-e5-small (fast, needs query/passage
prefixes — handled automatically), bge-large-en-v1.5, bge-large-zh-v1.5.
Scoring de-biases the vectors by subtracting the corpus mean ("centering"), which removes the shared direction that would otherwise make unrelated texts score ~0.8. Inference is pinned to a single thread so results are deterministic.
All heavy data lives OUTSIDE the working directory: the model downloads once
from huggingface.co into <cacheDir>/models, and the skill-embedding index
persists at <cacheDir>/skill-index.json (atomic writes, tagged with the
model id), so unchanged skills are never re-embedded after a restart.
node_modules/@deepseek-ai and node_modules/js-yaml are local symlinks into
a DSH checkout for type-checking and demos; recreate them after a fresh
install. At runtime a real DSH profile provides the @deepseek-ai/* packages
(they are declared as peerDependencies and kept external by the build).