DiffAudit – human review checklist for AI edits

Surface the suspicious lines in AI-generated diffs so devs review only what matters

Problem

Developers using AI coding agents face a lose-lose: reading every line of a multi-file diff is exhausting, but skipping review risks silent breakage. Teams report small changes shipped past three review stages that still broke production, because reviewers left P2 comments while P0 issues slipped through.

Solution

A local CLI/IDE plugin that ingests an AI-generated diff and ranks changes by risk: removed auth checks, altered routing/config, quietly edited test assertions. It produces a short "must confirm by human" list and folds the rest, with per-repo whitelists for files AI may change freely.

Users

Solo developers and small engineering teams using Cursor, Claude Code, or similar coding agents

MVP

Parse git diff and detect high-risk change patterns; output red/yellow/green review checklist by file/line; support repo-level whitelists and ignore rules; run fully locally without uploading code; generate a Markdown review report with one command

Revenue

Free for individuals, per-seat subscription for teams; optional one-time purchase for a private rules pack

Why now

Agent-generated changes now exceed what humans can review line by line, yet general code review tools were not designed for batch agent edits or this risk model