For bigger projects, it's easy to burn through your entire context window for a recap. Why not add in sub-agent delegation if the repo can be split into separate elements, or even when it can be done by one, split it into orchestrator + 1 context heavy sub-agent? Seems there's no downside.