Cavecrew

Three subagents that return compressed output to save your main context budget across long sessions. Investigator locates code, builder handles surgical edits on 1-2 files, reviewer audits diffs. The hook is that their tool results come back caveman-style at roughly 60% of vanilla token cost, so you can delegate 20 times instead of exhausting context at 12. Trade prose for structure. If you need architectural commentary or deep rationale, stick with the standard agents. If you just need facts and your session is already running long, this keeps you in the game.