My workflow these days is basically ultracode + PRs. I do control the features that go but barely need to check the code due to the amount of checks: lint, TS, Local AI visual Model for visual check for screen frame comparison (Mac M5 48GB mem), Pixel/Screenshot regression test for basic UI, code coverage/c8 to avoid trash code without tests, etc. etc.

Essentially the tests are just for the AI to remember what they did (as in the description + break if they change smth unintentionally).