Overview & result
The Korean-language repository packages Jev as a CLI plus Claude Code and Codex integrations. Agents can ask bounded questions about file relevance, failing tests or categories and receive typed answers with probabilities instead of generating prose.
From the original post
Claude Code와 Codex가 판단하느라 태우는 토큰을 줄인다.
How Jev fits in the loop
- An agent or shell pipeline supplies state and one or more bounded questions to the CLI.
- Jev returns typed Noul, Choice or Score answers and probabilities in one request.
- The coding agent uses those answers to select files or triage work; this CLI does not switch the agent's model.
How to reproduce
- Clone the repository, run `npm test`, and set `TYPESAFE_API_KEY`.
- Use the `jev` CLI for a Noul, Choice or Score question, or install its Claude Code skill / Codex prompt.
- Try the README's batch file-relevance example and inspect the agent's subsequent file reads.
Why this build matters
A reproducible multilingual example of Jev as a small decision tool inside coding agents, distinct from model routing.
Limitations
- This repository is a decision CLI and skill, not the author's separate model-routing project.
- The README's price and latency examples are author-provided, not independent measurements.