Need some help here....
I am working on a side agent project. A couple of iterations around 25k lines of code were spit out.
After each iteration, I had asked Claude to run the tests and assert them. Needless to say, I checked the test output, looked good, looked real, test cases passed, failed cases corrected.
Ran the agent on a real use case(browse data). The pipeline broke completely.
I looked inside the test file, and all cases were mock-ups, every single one of them. Async process simulated by sleep, sampling simulated by returning fixed numbers, and so on.
When working with a code base that is growing at 1k lines per 10 mins, what's the best testing methodology where you can't sustainably check each line of code?
I have tried many prompting techniques, but fake test cases always pop up.
AI always sneaks up. Any advice is welcome.