Are the benchmark numbers real?

Labels came from agents forbidden from running jev with file and line citations plus near-misses, alongside paired agent transcripts; find, ask, hook, token, and timing suites live under bench.

Will it actually make my agent faster?

No. Wall-clock is a wash since about 13 seconds per find offsets removed turns; the win is billed and context tokens from avoided re-sent history rather than speed.

What if the feature does not exist?

Jev answered no match 6 times out of 6 on absent features, while grep and BM25 still offered files, because lexical methods cannot report absence.

Can the Read hook hide the line I need?

It never touches explicit offsets, refuses files above 80 KB where cross-section scores are incomparable, and measured 50 of 50 targets kept versus 28% for a random window.

How do I install and verify it?

Needs Go 1.22 or newer and a TypeSafe API key. Run claude plugin marketplace add BorisLeMeec/jev then claude plugin install jev@jev, and export TYPE_SAFE_AI_KEY in your shell profile. To use jev from your own terminal too, make install copies it to ~/.local/bin, and jev probe verifies the API contract with raw and decoded responses.

What did the context measurements show?

Across nine real Claude Code sessions, cache reads re-sent 422,164,852 tokens, average context re-sent per turn was 183,740 tokens, and all tool results combined were 196,317 tokens, about 0.04% of the total. That is why the lever is keeping bytes out of context rather than shrinking tool output.

How do jev ask and the Read hook behave?

jev ask sweeps a directory and returns only the yes-files with the line, with strongly bimodal scores (true files 0.92 to 0.99, false files 0.02 to 0.13) at about $0.0016 and two seconds per question. The Read hook narrows files between 400 lines and 80 KB to about a fifth, adds about 1.25 s per read, refuses files above 80 KB, and never touches a read with an explicit offset.

CommunityJev Agents & MCP

Jev — BorisLeMeec

Claude Code plugin with jev find, jev ask, and a Read hook that keeps large-file context out of the agent by scoring relevance with Jev.

See highlights

Open github.com

Stars: 40
Language: Go
License: MIT
Published by: Community
Author: BorisLeMeec
Last verified: Oct 2026

BorisLeMeec/jev

User repository on GitHub

View on GitHub

A claude code plugin for jev

Forks: 6
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Jev — mayank953.

About this project

Answer questions about a codebase without reading it into the agent's context. A Claude Code plugin backed by TypeSafe's Jev — a small model that returns calibrated probabilities instead of prose, at $0.042 per million tokens. It reads your files so the agent doesn't have to. Every turn is checked. Only searches and large reads are taken over — an Update, a Bash, a file under 400 lines pass straight through. The panels in it are a diagram, not the UI: the plugin is a hook and draws nothing.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Go

Highlights

Find reaches 0.96 P@1 overall and 0.92 on vocabulary-gap queries where grep and BM25 fall to chance levels.
Ask returns only yes-files with lines at 1.00 precision and 0.96 recall, strongest on rare properties.
Read hook narrows 400-line to 80 KB files to about a fifth, keeping 50 of 50 targets at median 4 lines away.
Paired agent runs bill 30-41% fewer input tokens at equal quality; wall-clock is roughly unchanged.

Watch out

Scores above 80 KB are refused rather than narrowed, gates fail toward doing nothing, and savings compound only when sessions read large files repeatedly. 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

jevyoumean

Shell wrapper that intercepts unknown subcommands, reads the target CLI help output, and asks Jev which documented subcommand matches the intent behind the typo.