Design-partner pilot · one task · free
An agent PR you don’t have to take its word for.
Our agent works on one task in your repo inside a disposable sandbox. Before anything is published, a separate gate re-runs your tests and checks the diff against a written contract. If it fails, you never see a PR.
Why a gate
Agents report success they didn’t achieve.
One of our own agent runs reported that it had added two tests. The test count hadn’t moved, and without a check that PR would have gone to review looking finished. So publication no longer depends on what the agent says. The executor collects the facts itself: the staged diff, the file contents, and a fresh test run it starts after the agent has stopped. Replayed against that run, the gate fails it.
- Scope
- Changed files must be inside the contracted list; line additions and deletions are capped.
- Content
- Required and forbidden strings are checked in the changed files.
- Tests
- Named vitest runs are re-executed after the agent finishes, with the pass count read from vitest’s JSON report at a path chosen at check time.
- Record
- The diff, the test report, and every check result are kept whether the task passes or fails.
How the pilot runs
Contract in. Checked PR out.
- 01
Scope
You name one bounded task in a public GitHub repo. We turn it into a written acceptance contract: which files may change, and which tests must pass.
- 02
Run
An agent works in a disposable Cloudflare sandbox on a fresh clone of your repo. It can read, edit, and run your tests. It cannot publish.
- 03
Verify
The executor re-runs the contracted tests itself and reads the counts from the test runner’s own report. The agent’s summary is never used as evidence.
- 04
Deliver
Only a change that passes opens a pull request, from a Stackbilt fork, with the acceptance report in the description. A failing change opens nothing.
- 05
Decide
You review it like any contributor’s PR. Nothing merges unless you merge it.
Good fit
What the pilot can handle today.
These limits are real. We’d rather turn a task down than hand you a PR we couldn’t verify.
- ✓A public GitHub repository
- ✓TypeScript or JavaScript, with a committed npm or pnpm lockfile at the root
- ✓Tests that run under vitest
- ✓One task you could describe to a new contributor in a paragraph
Design-partner application
Bring one task you’d give a new contributor.
A bug with a reproducible failure, a missing test, a small, well-defined feature. We’ll reply within one business day, either with the acceptance contract we propose or with why it isn’t a fit.