That's worth testing for sure. This first version was mostly built against routine feature updates, typically spread out over several PR's. I found it was more likely agents missed context by ONLY sticking with a single PR vs seeking out related merges in a similar timeframe.

I designed the evals to filter out noisy stuff (we don't need to hear about the play by play of tiny bugfixes), and most of that skill should carry over to a monolithic refactor.

Going to test further on the next iteration, thanks!