A high-level look at the loop — enough to know what you're installing, before you read a single config file.
You write the issue — what should change, and why.
Your chosen coding agent works the task inside an isolated workspace.
Every change is checked against the scope you declared, and scanned for secrets. Fails closed, not open.
Your own test suite runs against the result — no separate CI to wire up.
warden opens a draft pull request. It never approves, never merges — that's always yours.
Which files and paths a run is allowed to touch.
Which coding agent runs the task — claude-code, codex, or your own.
The commands that must pass before a draft PR opens.
A cost cap per run — warden stops before it's exceeded.
Want to try it on a real repo?
Request early access