I come to conclusion that software source code will be kind of ephemeral output, like binaries are today. If you are not happy about slop accumulation, you take still valid functional requirements, docs, API contracts, test cases, some selection of still valid prompt history, and let better agent rewrite everything from scratch. Even if the session costs $200K one time - that is very cheap even for SMEs.

I really worry about "regressions" in that model. If you regenerate everything all the time. Do you keep issues fixed or constantly end up introducing new issues. Well it might work with absolute total test coverage. Which seems to be very high bar to reach.