How do you keep your big repos up to date?

Like every new session does read agents.md and a little documentation.md where my repo structure is explained (by gpt himself and he keeps updating documentation after every change in the repo) But the issue is that agents often miss to analyze the whole repo turthfully, meaning often there are old files or references that just bloat and missinform. documentation is not a 1:1 state es it would be thousands of lines long, its just a overview. So yeah, i feel like the best way to let a agent run is to always completly analyze a repo and then work on that - but that is not efficient and does not work at all because lf context limits. I cant always suck in 200k kontext to have the actual states of my repos. So yeah, how do you guys solve it? Or dont you care at all?