Code
Find your flaky tests
Which ones fail without a code change, ranked by what they cost.
Fill it in
Where the run history is.
How far back to look.
days
With the rule for getting back out.
Your prompt
Find the flaky tests in [your ci history], over 90 days. A flake is a test whose result changes without the code changing. Compute it per test per commit: same commit, different outcomes. Anything else is measuring failure rate, and a genuinely broken test looks identical to a flaky one on that metric. Rank by COST, not by flake rate. A test that flakes 2% of the time on every run blocks more pull requests than one that flakes 40% in a job nobody waits for. Propose a quarantine list with the rule for getting back OUT. Quarantine without an exit is deletion with extra steps, and the test rots there while the thing it covered breaks quietly. Group by likely cause: timing, shared state, ordering, external dependency. Usually a handful of causes explain most of the list.
Use Find your flaky testsOpens with everything above already filled in.
Why this works
Measuring failure rate cannot tell a flaky test from a genuinely broken one. A flake is the same commit producing different outcomes, which is what this computes, then ranks by pull requests blocked rather than by flake percentage.