Jev Agent Skill Router

What Jev Agent Skill Router does

Jev Agent Skill Router is a small Python CLI and library that decides which skill, if any, an agent should use for a request. Given a request and a catalogue of skill names with public descriptions, it returns route, no_skill or review, naming a skill only for route, along with confidence values, each call's candidates and evidence, and measured latency and token usage. It never loads skills or runs their instructions; the README calls it an optional routing layer, not a replacement for Hermes.

Large catalogues are sorted, split into deterministic batches and judged in parallel with TypeSafe Jev. Two candidates survive each batch, and the reduction repeats until a final choice fits, with explicit no-skill and review options. The author motivates the project with a Hermes profile that exposes 443 unique skill names. On a synthetic benchmark of 24 skills and 72 requests it scored 68 of 72 correct against 51 of 72 for a lexical baseline, but the README calls these exploratory results that do not establish production accuracy.

Key features

• Returns route, no_skill or review, naming a skill only when it routes
• Confidence and winner probability for the final choice
• Batched parallel evaluation that scales to large catalogues, tested offline with 1,000 skills
• Separate need, ambiguity and candidate-fit judgements gate the result
• Reports end-to-end and per-call latency plus API-returned token usage
• Fully synthetic public benchmark with a results report and evaluation protocol

When to use it

• Selecting one skill from a Hermes profile with hundreds of skills before loading it
• Having an agent abstain for review when no skill clearly fits
• Studying skill selection separately from skill execution

Who it is for: Developers and researchers working with large agent skill libraries who want routing decisions with visible uncertainty.

How it fits with Hermes Agent

Motivated by Hermes Agent profiles with large skill catalogues; it is an optional routing layer that does not load or execute Hermes skills itself.

Note: The README describes the benchmark as exploratory, on synthetic data, and not proof of production accuracy.

FAQ