Loop until dry
Run every check, fix what it finds and run again. Stop only when a whole run finds nothing new.
Use it When I want to know that something is done, not that it was fixed once.
The problem
A fix is a change, and a change can break something. One round of “check, then fix” ends with code that nobody has checked. It is easy to stop there, because the list of findings is empty and that feels like done.
How it works
- Run all the checks on the work, not only the one that failed last time.
- Ask one question about the report: did this run find anything new?
- If yes, fix each finding and go back to step 1.
- If no, stop. The loop is dry.
The rule that matters is that the last thing to happen is a clean run, never a fix.
What it costs
At least one run more than feels needed, since the last run finds nothing by design. When the checks are slow or cost model calls, that adds up, so the loop needs a limit on rounds as well.
Where it fails
It is only as good as the checks. A run that finds nothing means the checks found nothing. Some bugs only show in the real place: one check of the workbench passed on my machine and failed on the deployed preview, where two requests arrived at the same moment.
A check that sometimes fails by chance keeps the loop from ever going dry. That has to be fixed in the check, not in the work.
Where I used it
In every package of the AI lab: the unit tests, the script against the deployed preview and the browser pass ran again after each fix, until a run came back clean.
Used inAI lab