Jev Code Reviewer ranks agent-written PR changes by priority to cut review overload

An open-source project from egma-ai called Jev, hosted as jev-code-reviewer on GitHub and surfaced on the Hacker News front page, takes aim at a specific pain point of agent-assisted development: pull requests written by coding agents routinely arrive with enormous file counts, and when a human faces that wall of diffs, the review often collapses into a rubber stamp. The project's README opens with the example of an agent showing up with 230 changed files. Jev's answer is to rank every change and show only the urgent ones first.

What it does

According to the project's README, Jev classifies each change in a PR into three tiers, P0 through P2, and expands only the P0 items by default, leaving P1 and P2 collapsed. The thresholds are configurable: reviewers can copy a policy file into the repository being reviewed or override it with a command-line flag. The README frames the priorities as a signal for where human attention should go, not a verdict on whether a change is correct.

The diffs are also rewritten in natural language, with OpenAI generating the explanations, and the original code remains one toggle away. The tool ships in three parts: a local CLI for reviewing your own agent's PRs, a Chrome extension that modifies GitHub's Files changed page, and an agent skill that lets a coding agent run the analysis itself immediately after opening a PR.

Setup and workflow

The requirements are Node.js 22 or newer, a signed-in GitHub CLI, and a local clone whose origin points at the PR's repository. After linking the package, a setup command stores TypeSafe and OpenAI API keys, a doctor command checks the environment, a serve command runs a local server, and an analyze command is pointed at a PR URL and a repo path. The Chrome extension is loaded unpacked from the repository, paired with a token generated by the CLI, and switched on with a "Show logic in place of code" toggle on the PR's Files changed page.

A demo mode replays a pre-recorded analysis in the browser and makes no calls to AI providers; the project notes that the recording combines real Jev classifications with prepared explanation text. An optional code-graph tool, installable via uv, can be added to give the analysis better context.

Limits and privacy trade-offs

The README is candid about what the tool does not do. The analyze step sends changed code and nearby context to TypeSafe and OpenAI, so source code does leave the machine during analysis, even though the browser extension itself only talks to a local server on 127.0.0.1:4731 and nothing is posted back to GitHub. Credentials are stored in a config file with owner-only permissions, but the file is not encrypted.

Coverage is also capped. Only the first twelve diff hunks in path order are analyzed by default, with a flag raising the ceiling to one hundred; files beyond the cap keep GitHub's normal code view, and both the CLI and the extension popup report how many changes were left unanalyzed. The extension works only on the classic Files changed page, not GitHub's newer changes view, and reports go stale whenever the PR receives new commits, forcing a rerun before the extension will display them again.

Why it matters

As coding agents produce a growing share of merged code, review is becoming the bottleneck, and approving a 230-file PR unread is a real failure mode rather than a joke. Jev's core idea, treating human attention as the scarce resource and spending it where an automated triage pass says it counts most, is a pattern worth watching in other review tooling. The caveats are just as instructive: classification depends on external LLM providers, coverage is bounded, and the tiers point at attention rather than guaranteeing correctness, which makes this a triage layer on top of review rather than a substitute for it. With an MIT license and a local-first design, it is also straightforward to audit and adapt.