An agent patch rarely needs your whole suite. It needs a scope you can defend line by line, and a safety net over everything the scope excludes. Most review pipelines get this backwards: they run everything, watch it go red at random, and then learn nothing about which failure the patch actually caused.
This is a selection strategy, not a speed trick. The gate runs a narrow set of tests, emits a receipt explaining why each test was included, and escalates to the full suite whenever a widening trigger fires. Flaky tests are frozen separately β and a freeze is only granted once a property check covers the same behavior.
The cost is a product, not a number
Write the cost down before optimizing it. For a patch p :
cost(p) = |selected(p)| * t_avg + |frozen| * t_skip + widen(p) * |full| * t_avg
Enter fullscreen mode
Exit fullscreen mode
widen(p) is binary here on purpose. Either a trigger fired or it did not.β¦