lau4tin1/dsh-skill-router

A Skill Router for DeepSeek Harness

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.

• Design: see DESIGN.md.
• Status: V1 complete and running — type-checks against the real DSH 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.

Install

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. If dsh plugin stops with ERR_PNPM_IGNORED_BUILDS, open $DSH_HOME/profiles/<name>/pnpm-workspace.yaml, set the allowBuilds entries pnpm wrote there to true, then re-run the same dsh plugin ... add command.

The first real task downloads the embedding model once (~543MB, q8) into $DSH_HOME/skill-router/models, then runs fully offline.

Layout

• 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.

Build

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)

Demo

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

Config (defaults)

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.

Languages

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).

Embedding model

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.

Dev note

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).