For developers
Every claim it makes, you can go and check.
You have watched an agent announce that it fixed something, and found no diff. Chovy is built around the assumption that you will look.
The rule
A build cannot report success on its own say-so.
This is enforced outside the agent, not asked of it in a prompt. Three facts are collected after the work and compared against what was claimed.
git rev-parse HEAD
A version was saved
Every build and every change is a commit in Git, named in the words you used to ask for it. Any of them can be made live again.
git diff --stat
Something actually changed
The list of files that moved, every time. A build that produced nothing is reported as producing nothing, never as a success with an apology.
curl -sI https://…
The address answers
Your app's real URL is asked from the open internet whether it is up. If it is dark, the build is not done and nobody is told otherwise.
Then each acceptance statement from the plan is answered on its own. A statement nothing exercised is reported as unproved rather than folded into a pass, which is the failure mode that makes most agent test reports worthless.
What you get on day one
Ordinary infrastructure, not a walled garden.
The owner's original brief, in their own words
Every note and decision since
An isolated workspace and repository for this project
Previous work with its git checkpoints and previews
A preview URL you can share back
The shape of it
A short line, not one agent in a loop.
One agent asked to plan, build, test and report on its own work is the same person marking their own homework. Chovy splits it: one step turns the plan into an instruction and writes down what will count as done, a second builds against that instruction, and a third checks the result against it.
The acceptance statements are written before the build, by a step that does not do the building. That ordering is the whole point.
1 · plan
Reads the blueprint. Produces an instruction and a list of statements that must end up true.
2 · build
Works in the repo, commits, and brings the app up on its reserved address.
3 · check
Runs the suite, opens every route, answers each statement, and sends the work back once if something failed.