jev-skill-gate
Scores installed Claude Code skills against the current project with Jev and hides irrelevant ones via skillOverrides.
Overview
jev-skill-gate tackles the context cost of Claude Code loading every skill description at session start, which the README puts at roughly 10,000 tokens on a full skills library. It scores each installed skill against the project you are in using Jev in one parallel pass, then writes skillOverrides so only relevant skills reach the model, with some kept as name only. The list summary cites a 217-skill install going from 12,750 to 3,185 tokens for $0.0009 a session.
The repository includes a committed eval of 20 labelled cases over 217 skills, using only the prompt as state. Jev was run on 4 cases and a free local scorer on 18, so the columns are not head-to-head. On its 4 cases Jev reached a mean pairwise AUC of 1.000, a median expected-skill rank of 3, no irrelevant skills in any top 10, and kept all 19 expected skills (14 with full descriptions, 5 name only). The local scorer reached 0.961 AUC with 92.9% of expected skills surviving. Scripts can rerun the local arm free or rederive numbers from committed scores.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How relevant is this skill to the current project? | — |
How it works
- Collect installed skills and project signals
- Jev scores every skill in one parallel pass
- Write skillOverrides to hide or shorten low scorers
- Claude Code starts with a smaller skill manifest
Strengths
- Large cut in skill manifest tokens
- Committed, reproducible eval
- Free local scorer fallback
Caveats
- Jev arm evaluated on only 4 cases
- A wrong hide could remove a needed skill
- Needs TypeSafe credits for the Jev arm
Getting started
Install from the releases, set your TypeSafe key, and run node eval/run-eval.mjs --local --fresh to try the free scorerAt a glance
- Type
- Agent skill
- Status
- beta
- Language
- JavaScript
- Runtime
- Node CLI (Claude Code hook)
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Ranks Claude Code skills for the current project to reduce the descriptions loaded by default.
logicrw/awesome-jev-projectsCut Claude Code's skill manifest by ~75% with TypeSafe Jev. Scores every installed skill for relevance and hides the rest via skillOverrides — 12,750 → 3,185 tokens on a 217-skill install, for $0.0009 a session.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
How relevant each installed skill is to the project you are working in.
Did it hide any needed skills?
Across the 4 live Jev cases, all 19 expected skills survived the gate.
Can I try it without a key?
Yes. The local scorer arm runs free in about two seconds.
Related projects
Agent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
ProjectbetaPythonDeepSeek Harness plugins that filter tool output, prune stale context and disclose skills progressively, with Jev or Laya judging what goes.
PluginbetaTypeScript- winnow★ 110
Calibrated context sieve for Claude Code: Jev judges each block of large tool output and hides confident-no blocks behind recallable stubs.
Projectbeta Local proxy for Claude Code that uses Jev to remove stale tool results from the context once it reaches about 120K tokens, failing open.
ProjectbetaJavaScriptClaude Code /fc command where Jev picks which old tool outputs to keep whole; the rest are trimmed and saved to files.
ProjectbetaTypeScriptLocal compaction proxy for Claude Code and Codex that compacts more often with local rules and keeps full history searchable.
ProxybetaGo