Are the token and accuracy claims believable?

The digest reports roughly a fourfold context reduction per turn with the right skill in front of the model more often, while correct-answer rate holds and rule-carrying skills improve. Intervals and test details are in the results section.

What counts as a judge besides Jev?

Anything implementing Pick plus YesNo with declared limits and an ask method, from hosted models to rules. The router splits large catalogs by those limits and ScriptedJudge plus a contract check cover tests.

What happens on each user turn?

Rank and need-gate run in parallel, top candidates are verified against SKILL.md starts, confident picks load fully and uncertain ones become a short list with a search tool. Each turn is capped at two seconds by default.

How do I wire it into an existing deepagents agent?

Put skills in a skills directory with name and description front matter, install with the Jev extra, set both TypeSafe and provider keys, and swap the middleware list. It needs Python 3.11+.

What happens if the judge times out or is down?

A judge timeout or outage hands the agent the full catalog, exactly as it would be without Skill Router, so nothing is lost; bad credentials, by contrast, raise instead of hiding a broken setup.

How does it handle a catalog larger than the judge can see?

The judge declares per-call limits, and for Jev that is 32k tokens and 255 options; a bigger catalog is split into parts, keeping related skills together, and the part winners are ranked again.

How are uncertain picks handled?

A pick verified at 0.9 or higher is loaded with its instructions; otherwise the model gets a short list of up to three candidates and can search the rest of the catalog with the find_skill tool.

CommunityJev Agents & MCP

LangChain Skill Router

LangChain deepagents middleware that replaces SkillsMiddleware with per-turn judging, loading only the SKILL.md files a turn needs so large catalogs stay out of the prompt.

See highlights

Open github.com

Stars: 6
Language: Python
License: MIT
Published by: Community
Author: deyna256
Last verified: Oct 2026

deyna256/langchain-skill-router

User repository on GitHub

View on GitHub

Per-turn skill selection for LangChain and deepagents agents: a fast judge picks the few skills a turn needs, so a catalog of hundreds stays out of the prompt.

Forks: 5
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Per-turn skill routing for LangChain deepagents. A drop-in replacement for the deepagents SkillsMiddleware : on each user turn a fast judge decides which SKILL.md skills are needed, and only those are loaded, so a catalog of hundreds stays out of the prompt. Bring any judge: a hosted model, a self-hosted one, or plain rules. An adapter for Jev is included.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing Jev for Ranking Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Per-turn rank, gate and verify keeps prompts small and cache-friendly.
Pluggable judge protocol separates routing core from LangChain and providers.
Conversation-aware ranking keeps follow-ups on the thread's skill.
Short-list plus search tool covers uncertain picks without loading all skills.

Watch out

Synchronous runs fall back to the ordinary skills middleware, oversized catalogs are split within judge limits, and judge outages hand the agent the full catalog while bad credentials raise. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-judge-mcp

MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.