We've all done it. You fill in a long sign-up form, a job application, or a checkout, press Back or Refresh, and everything you typed is gone. These "lost my work" bugs are everywhere, and unit tests rarely catch them because they only show up in a real browser.
For the KaneCLI Online Hackathon by TestMu AI, I built StillHere: a loop that catches these bugs in a real browser and keeps fixing them until the test passes.
What it does
StillHere runs a closed loop:
- Find it. Kane CLI drives a real Chrome session through a journey described in plain English: fill in a checkout, open another page, press Back, press Refresh, and check the cart, details, and current step are still there.
- Fix it. If anything is wiped, Kane's failure report becomes a repair prompt for the Codex coding agent, which fixes the app without weakening the test.
- Prove it. The exact same Kane objective re-runs, unchanged, until it passes.
It works for any form or multi-step flow via --objective, not just checkout.
How I built it
Node.js with no dependencies. StillHere calls Kane CLI with --agent, parses its NDJSON result, writes a failure report with evidence and screenshots, and pipes a repair prompt to codex exec. A sample shop with a real Back/Refresh bug is included, and npm run demo:loop resets the bug so you always see fail, fix, pass. Built solo with Codex.
What Kane caught
In the recorded run, pressing "Back to checkout" wiped the order: the cart, the customer's details, and the review step were all gone. Codex fixed it by saving state in sessionStorage plus browser history and restoring it on load and Back. Kane then passed through both Back and Refresh. The test was never changed to make it pass.
Challenges
Chrome crashed when launched from a sandbox, so Kane runs from a normal terminal. Long Kane runs timed out, so the default timeout became 300 seconds. I also made sure a run with no final Kane result counts as incomplete, not as a failure for the agent to "fix."
What I learned
Verifying in a real browser catches bugs a mock never would, and keeping the test fixed is what makes an AI fix trustworthy.
What's next
A GitHub Action that runs the loop on every pull request, plus ready-made objectives for sign-up, login, and multi-step forms.
Try it
- Code and setup: https://github.com/Mzee04/stillhere
- Demo: https://youtu.be/4R3yvR5oKvg
What "lost my work" bugs have you run into?

