Chasing a Flaky Test Down With Cursor
A test that passes sometimes is gone when you look. Use the model for hypotheses, not patches.
A flaky test is the worst kind of bug because it's gone when you look. I paste the failure and the test into Cursor and ask 'what could make this pass sometimes,' not 'fix it.'
The model is good at listing the usual suspects: timing, order dependence, shared state. I read the list and pick the likely one, then ask Cursor to show me where that state leaks between tests.
Once I know the cause, I let it write the fix and the test that proves it, then I deliberately break the condition to confirm the test catches it. If it doesn't, the fix was a guess.
The lesson is to use Cursor as a thinking partner on the hypothesis, not the author of the patch. A flaky test fixed by vibes comes back, and it comes back at 2 a.m.