jev-assist
Pipe grep or file lists into Jev to get every candidate scored, sorted and cut at a calibrated threshold, so agents read only what matters.
Overview
jev-assist is a CLI and agent skill that classifies search results rather than performing searches. You pipe in raw candidates (grep -rn output, git ls-files, test runs) and each line comes back with a calibrated probability, sorted, with an exact total and a cut line; the reply is about 30 rows whether the pool was 40 or 4,000. It uses Jev's three judgments: noul (is this hit real?), choice (what is it?) and score (how bad is it?). Finding things stays with grep, rg or Glob by design.
The README example filters getUser hits to the two lines that call the user API in production code rather than tests, mocks or comments, at a 0.70 cut. Per the yibie list it can also rank every tracked file by relevance to a one-line task description in parallel batches, and a validate command grades rankings against past commits. It needs Node 18+ and a TypeSafe or OpenRouter key, caches results, writes only the skill directory, key file and cache in current versions, has CI and an npm package, and is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this line call the user API in production code, not a test, mock or comment? | yes · no |
| Choice | What kind of hit is this? | — |
| Score | How severe is this hit? | — |
How it works
- Run grep, rg or git ls-files
- Pipe the output into jev noul/choice/score with a question
- Jev scores every candidate in batches
- Output is sorted with a calibrated cut line
- Agent reads only rows above the cut
Strengths
- Bounded output size regardless of input size
- Reports totals for the full pool; nothing capped
- Caching avoids paying twice
- Validate command checks rankings against history
Caveats
- Sends code lines to the API
- Older versions installed hooks that need manual removal
Getting started
With Node 18+ and a key, e.g. `grep -rn "getUser" src/ | npx jev-assist noul '...'`, or install the skill via skills.sh.At a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- Node CLI, agent skill
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Coding agents: ranks every tracked file by relevance to a one-line task description — Jev asks each file the same typed question in parallel batches, so an agent in a 600-file repo starts from the handful it actually needs — with a validate command that grades the ranking against past commits.
yibie/awesome-jevDon't burn your expensive main model on grep-and-guess grunt work — let jev rank the whole repo, and save the main model for reading the right files and writing the right code.
logicrw/awesome-jev-projectsDon't burn your expensive main model on grep-and-guess grunt work — let jev rank the whole repo, and save the main model for reading the right files and writing the right code.
hellogumbo/awesome-jev
Frequently asked questions
Does jev-assist search my code?
No. It classifies candidates you pipe in from grep, rg or file lists.
How big is the output?
About 30 rows regardless of how many candidates you pipe in, with totals for the full pool.
Which keys work?
A TypeSafe or OpenRouter API key.
Related projects
Retrieval layer for Codex that finds evidence locally, optionally ranks sanitized excerpts with Jev, and returns bounded, cited evidence.
ProjectbetaCodex search plugin that runs ripgrep and keeps only passages Jev judges relevant to the question.
PluginprototypeTypeScriptDeterministic code-graph context for coding agents on Postgres, AGE and pgvector, with optional Jev or local re-selection of symbols.
MCP serverbetaPythonClaude Code plugin that adds Jev as an independent judge for review findings, debugging hypotheses, design options and search hits.
PluginbetaDeepSeek Harness jev_search tool that runs ripgrep, scores every hit line with Jev, and returns only the top rows above a cut.
PluginbetaJavaScriptGo-based Claude Code plugin where Jev reads code so the agent doesn't: semantic find, bounded questions and large-read takeovers.
PluginbetaGo