Is there a reason to prune individually instead of introducing a tombstoning approach like kimi?
Most of my harnesses around agentic retrieval wind up implementing the poor man’s version of this via isolated context windows and recursion. But it seems like an entire trajectory is more likely to be erroneous than its docs, and you could just rewrite true positives in poor trajectories as the summary?
