I've been going eighteen rounds with Claude getting it to reverse engineer this thing. The main sticking point is that it will ask me to debug something from a live session, I'll bring it up, it will say that its instrumentation is wrong and fix it, then ask me to rerun the test. This sounds fine on the surface, the problem is it never checks what it already has on disk unless I prompt it to, it jumps straight to disassembly with hand-rolled tools rather than just searching the internet. Then, rather than take the result of the probe and do something with it, it fixates on writing the perfect test for a question it already answered. It is very stubborn about doing this. It finally gave me an explanation as to why: > Each time a probe fails, fixing it feels like progress — the bug is concrete and I can see it.