██████╗ ██████╗ ██╗ ██╗
██╔══██╗██╔═══██╗██║ ██║
██████╔╝██║ ██║██║ ██║
██╔══██╗██║ ██║██║ ██║
██║ ██║╚██████╔╝███████╗███████╗
╚═╝ ╚═╝ ╚═════╝ ╚══════╝╚══════╝中文版 README
Roll puts the AI coding tools already on your laptop to work the way a small engineering team works — and shows you the receipts.
You already have an AI that writes code when you sit there and prompt it. Roll is the part around it: it turns your requirement into small tasks, hands one task at a time to an agent, makes that agent test its own work, opens a pull request, and keeps evidence you can check afterwards.
What Roll actually does
Give it a requirement. For each task, Roll then:
- Splits the work — one requirement becomes small numbered tasks in a to-do
list at .roll/backlog.md, each with its own acceptance criteria.
- Picks one task and hands it to an agent — Claude, Codex, Kimi, Pi,
Antigravity, Reasonix, Cursor: whichever you have installed and logged in.
- Keeps the agent in its own sandbox — the work happens in a separate git
worktree, not in the directory you are sitting in.
- Makes it prove the work — tests must pass, then a second agent reviews and
scores the result.
- Opens a pull request — CI has to be green. A task is only Done once its PR
is merged into main. Roll never marks its own homework.
- Files the evidence — every task gets its own acceptance page with test
output and screenshots, so "it works" is something you can open and look at.