jev-skill-suggester
Codex skill and Python CLI that recommends one installed skill per task using Jev Choice ranking and Noul fit checks, or returns none.
Overview
jev-skill-suggester recommends a suitable installed skill for the current task. It reads SKILL.md names, descriptions and bodies, ranks descriptions with a Jev Choice and a group usefulness Noul, keeps up to three candidates, then reviews body excerpts with a Choice and per-candidate fit Nouls before recommending. It may return none or uncertain, and explicitly named skills are resolved locally with no API call.
It is a Codex skill plus a standalone stdlib Python CLI for users with many similar skills. It does not execute or install skills, change agent settings or add hooks, and the host must still read the recommended skill before using it. Recommendation thresholds are fit at least 0.80 and Choice confidence at least 0.65, which the README calls exploratory rather than calibrated. A first live test reports 15 of 16 labelled tasks with the expected result and no wrong final recommendations.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate skill best matches the task? | — |
| Yes / no | Is this group of skills useful for the task? | — |
| Yes / no | Does this candidate skill fit the task? | — |
How it works
- Scan skill roots; resolve explicitly named skills locally
- Choice ranks descriptions and a Noul checks group usefulness, keeping up to three
- Review body excerpts with Choice and per-candidate fit Nouls
- Recommend if fit >= 0.80 and confidence >= 0.65, else return none/uncertain
Strengths
- Stdlib-only Python, no pip install needed
- Can abstain, with explicit status codes for none, uncertain and incomplete
- Optional 24-hour response cache and HTTP attempt budget
- README reports 0 of 16 wrong final recommendations in its test
Caveats
- Thresholds are exploratory, not calibrated accuracy
- Only the Codex skill and CLI were validated; other hosts untested
- Median about 5 seconds per task in the author's test
Getting started
git clone the repo, then `python3 -I -B scripts/suggest.py suggest --task '...' --mode jev --prompt-key --out results/report.json` (Python 3.10+, TypeSafe key for jev mode).At a glance
- Type
- Agent skill
- Status
- beta
- Language
- Python
- Runtime
- CLI / Codex skill
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Coding agents: recommends which installed skills apply to a request, keeping the recommendation bounded and letting Jev decide.
yibie/awesome-jevA Python CLI and Codex Skill that recommends a suitable installed skill for the current task using TypeSafe Jev Choice and Noul checks without executing candidate skills.
logicrw/awesome-jev-projects用 TypeSafe Jev 推荐已安装 Skill / Bounded installed-skill recommendations with TypeSafe Jev. Python CLI, Codex skill, bilingual docs and live examples.
hellogumbo/awesome-jev
Frequently asked questions
Does it run the recommended skill?
No. It only recommends; the host agent must read the skill and decide how to use it.
What does local mode do?
It lists keyword candidates only and always reports local_only with no recommendation.
Which Jev model does it use?
It pins jev-1.13.0 and needs a TypeSafe API key in jev mode.
Related projects
Decision plugin for agent harnesses where Jev recommends tools, skills or owners and scores output, for DeepSeek Harness and iPolloWork.
Agent skillbetaTypeScriptCodex plugin that consults Jev to choose tools, models and skills, diagnose failed commands and check completion claims.
PluginbetaTypeScriptPortable decision layer for agent harnesses: Jev picks from host-supplied capabilities while the host keeps permissions and execution.
ProjectbetaTypeScriptLocal Python MCP service that keeps the skill catalog out of the agent's context and lets Jev pick which SKILL.md files to load.
MCP serverbetaPythonLocal gateway, MCP server and Claude Code hooks that let coding agents delegate explicit choices, classification and scoring to Jev.
MCP serverbetaTypeScriptStandard-library MCP server giving Codex advisory Jev tools for role and resource selection, with local validation and 85 offline tests.
MCP serverbetaPython