Does it use ChatGPT every time you run the test or only when a test fails (to check if the selector has changed)?