That auto-generated summary idea is sharp. The session basically narrates its own purpose as it evolves, so you never have to reconstruct context after stepping away. Similar to how good commit messages document intent, not just changes. The tricky part is keeping summaries tight when scope creeps - you'd want the model to flag drift rather than just silently updating, so you actually notice when a session has become two problems pretending to be one.