AI column

Coding@donghyeok99Staff

Jevgrep: find code by asking what it does

On a 2,761-file repo, one team's 20 questions all put the right file first

A jg question sweeps the repository's folders, only three files light up, and their source comes back with line numbers before End context.
A jg question sweeps the repository's folders, only three files light up, and their source comes back with line numbers before End context.
Who
Made by David Zhang (dzhng), for coding agents, and the people running them, who need the right files in an unfamiliar repository.
When
Public since 26 September 2026. The latest GitHub release is 0.8.0 (Windows support) on 1 October; tags reach v0.8.5.
Where
A skill for Claude Code, Codex, OpenCode and others. macOS, Linux or Windows, Node.js 22+. Jev runs through the provider you pick.

What was hard

You need to know where authentication is checked before a request reaches a handler, but you don't know what the function is called, so grep has nothing to grab. A coding agent in an unfamiliar repository spends part of every task just finding the right files, and pays for it.

What it does about it

Ask jg a question about behaviour, and Jev judges whether each folder, file and declaration is relevant to it. Only the branches that pass are explored, and the chosen source comes back untouched, with line numbers. It hands over evidence rather than an answer, so editing and testing stay with your agent.

README's comparison: 8 of 10 solved either way, about 30% lower agent cost
README's comparison: 8 of 10 solved either way, about 30% lower agent cost

What people made with it

GitHub@lslogis2026-09-29

After it worked on their 165-file project, they ran 20 Korean questions on Wan2GP (2,761 files): the right file came first every time, at a median 65 s and about 5M tokens each.

Accuracy on the large repo was excellent: the correct file was ranked first for all 20 queries.
comments 2 (2026-10-29 as of)GitHub ↗
GitHub@oconnorjoseph2026-09-29

A team that wired jg into their coding-agent harness: well-scoped questions find the right file first, but asking a folder with no answer marked 41% of files relevant and printed 5.4 MB.

We wired it into our coding-agent harness and it finds the right file first on well-scoped questions.
comments 1 (2026-10-29 as of)GitHub ↗

How it runs inside

  1. 1

    Claude Code · Codex → jevgrep skill

    A how / where questionThe skill's description tells the agent to start behaviour questions with jg rather than grep.skills/jevgrep/SKILL.md
  2. 2

    jevgrep skill → jg CLI

    jg "question" rootThe CLI reads the question and search root, loads the provider key saved by jg auth and starts a search.apps/cli/src/index.ts
  3. 3

    jg CLI → Repository files

    Reads eligible files onlyIgnored, hidden, dependency, binary, .env and credential-looking files are left out.packages/core/src/filesystem.ts
  4. 4

    Repository files → Jev (AI Gateway etc.)

    Batches of folder and file previewsRetrieval walks the tree in bounded batches and opens only the branches judged relevant.packages/core/src/retrieve.ts
  5. 5

    Jev (AI Gateway etc.) → jg CLI

    Relevance judgments, cached locallyJudgments go through the AI SDK with retries and a rate budget, and answers are cached on disk.packages/core/src/evaluator.ts
  6. 6

    jg CLI → stdout context

    Summary, file list, verbatim sourceStatus and file list first, then source blocks with line numbers and declaration locations, ending in End context.apps/cli/src/render.ts

What goes in, what comes out

jg "How are telemetry events recorded and sent?"Source of src/telemetry.ts, src/backend/events.ts and tests/telemetry.test.ts (README's output example)
20 Korean questions on a 2,761-file repo (Wan2GP)Right file ranked first in all 20, median 65 s and about $0.22 a question (measured in issue #38)

Is it for you

Good for

  • Finding where a behaviour lives in an unfamiliar repository when you don't know its name
  • Letting Claude Code, Codex or OpenCode look before they edit
  • Following a flow that spans files, like a request on its way to a handler
  • Finding which tests cover a given behaviour

Not for

  • Needs a key for Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen or a compatible endpoint, saved with jg auth.
  • Source under the search root is sent to Jev. .env and credential-looking files are left out, which the README calls best-effort, not a guarantee.
  • The output is evidence for the agent, not a promise that every relevant file was found.
  • When you know the exact symbol or path, rg or a direct read is enough.

Reviewed by Devalley’s editor before publishing. How the column is made

Sign in to post, comment and like.Sign in

Report

Reason

Read the community rules