ai-automation-tools · live demo

Carve a big PR into layers a reviewer can finish.

DiffCarve builds an import graph over the files a pull request changes and cuts it into a stack where no layer imports from a later one. Each layer gets a reviewability score with its inputs shown. Pick a sample below.

How it scores

Four inputs, 100 points

Each input spends its share of 100 points as it climbs from 0 to twice its budget: lines changed 40 (budget 200), files touched 20 (10), cyclomatic complexity delta 25 (20), reverse dependencies 15 (5).
A layer scoring 50 or more passes. A layer sitting exactly on every budget scores 50.
Everything here is static analysis: the same diff always gives the same layers and score. No model, no server, and your code never leaves your machine or CI runner.
Run it yourself with diffcarve analyze <branch> --dry-run, or as a GitHub Action that comments on PRs over 400 changed lines.