Skill
Skill Router
两阶段路由：embedding粗筛之后LLM精排，977技能里挑top5。中文query自动翻成英文再匹配，外加三重拒选防线，纯信息类查询直接踢到L0工具，别浪费LLM token。
When to use Skill Router
Choose if
You're routing an agent across a large skill catalog (README cites 977 skills) and want a two-stage pipeline — BM25 + embedding coarse filter, then LLM fine ranking — to pick top-K candidates, with automatic Chinese-to-English query translation and rejection filters that kick pure information queries to L0 tools before paying LLM tokens. Picks over a single-stage LLM router when catalog size and token cost both matter.
Avoid if
Your skill library is small (README says benefit kicks in above ~50 skills), you don't run Python 3.10+ with sentence-transformers, or your workload is dominated by information-retrieval queries the rejection filters might mishandle. Also avoid if you need a turnkey solution — the L0 tool layer expects external API credentials.
Risk Flags
- MEDIUM scope README states the three rejection filters may block legitimate requests in addition to pure info queries — false-positive risk worth tuning before production.
- LOW scope L0 tool layer calls external APIs (PubMed, Semantic Scholar, TinyFish per README) and requires environment-variable credentials; not turnkey.
- LOW cost Two-stage routing only beats single-stage cost when skill libraries exceed ~50 options per README — small libraries get no benefit.
Cost
Type: Unknown
Dependencies
Minimum runtime: Python 3.10+ with sentence-transformers (all-MiniLM-L6-v2) and numpy per README; initial setup downloads ~80MB model
Distribution
- ClawHub
yan-skill-router- License
- MIT-0