Spec driven development can help some, even for brownfield projects. You can have an LLM swallow up your entire project and spit out a spec, and then review that yourself.

For any issue, start a brand new context, point it to the spec, explain the issue and explain if it's a regression.

Also on it might seem like an obvious one, the more test coverage you have, the more your llm can tell if something has broken or if there's been a regression without needing to eat up context.

All of these things can help but there's no perfect solution.