egma-ai/jev-code-reviewer: Overview conduct, not simply diffs. Jev prioritizes human consideration; OpenAI explains the adjustments. Native CLI + agent ability + GitHub extension. · GitHub

Most PRs which might be generated by brokers at the moment get YOLO merged as a result of its laborious for human thoughts to understand when your agent simply abruptly reveals up with 230 file adjustments. This is an try to cut back the psychological burden by classifying every change in a overview to P0, P1, P2. Only P0 are proven by default. The priorities are configurable. The diffs are additionally present utilizing a pure language. The authentic code is one toggle away.

It runs in your pc, for reviewing your individual coding agent’s PRs, and posts nothing to GitHub.

Watch the recording · Open the demo PR. The bundled recording makes use of actual Jev classifications and labeled, ready clarification copy (provenance).

Needs Node.js 22+.

git clone https://github.com/egma-ai/jev-code-reviewer.git
cd jev-code-reviewer
npm set up
npm run demo

Open http://127.0.0.1:4731/demo for the replay, which makes no supplier calls, or load the extension (under) and open the demo PR’s Files changed page.

Needs the GitHub CLI signed in (gh auth login) and an area clone whose origin is the PR’s repository.

npm hyperlink                 # provides the jev-reviewer command
jev-reviewer setup       # shops your TypeSafe and OpenAI keys; run it in your individual terminal
jev-reviewer physician      # checks instruments, key sources, the native server, and supplier entry
jev-reviewer serve       # go away working
jev-reviewer analyze --pr https://github.com/OWNER/REPO/pull/123 --repo /path/to/clone

Then in Chrome:

1. Open chrome://extensions, activate Developer mode, click on Load unpacked, and decide this repo’s extension/ folder.
2. Run jev-reviewer token | pbcopy. In the extension popup, open Connection, paste, and click on Save. It ought to say Paired.
3. Open the PR’s Files modified web page (/pull//recordsdata) and activate Show logic instead of code.

Optional: uv software set up graphifyy provides an area code graph for higher context. To have your coding agent run analyze after it opens a PR, set up the agent skill.

• Your code leaves your machine. analyze sends modified code and close by context to TypeSafe and OpenAI. The extension solely talks to the native server on 127.0.0.1:4731.
• Keys are saved in ~/.config/jev-reviewer/credentials.json (owner-only, not encrypted). Prefer this over exported variables, which a coding agent’s shell usually can not see. Replace an OpenAI key with node scripts/setup-keys.mjs --replace-openai.
• Coverage is capped. The first 12 change models (diff hunks) in path order are analyzed; --max-units permits as much as 100. Files with an unanalyzed change preserve GitHub’s code, and the CLI and popup say what number of.
• Classic Files modified web page solely. GitHub’s new /adjustments web page shouldn’t be supported but; change again beneath your profile image → Feature preview.
• Old experiences are by no means proven. If the PR has newer commits, the extension retains GitHub’s code: rerun analyze, then click on Refresh report.
• Priorities imply consideration, not correctness. P0 recordsdata open by default; P1 and P2 begin collapsed. Tune them by copying config/coverage.json to .jev-reviewer.json within the reviewed repo, or go --policy.

npm take a look at
node scripts/test-extension.mjs   # wants Chromium and port 4731 free

See architecture, demo guide, and extension details.

MIT.