What problem does it solve? AI agent configurations accumulate token waste: bloated CLAUDE.md files, never-invoked skills, and context re-pasted every session. The caveman learn analyzer measures where tokens go, but acting on its findings requires careful, reversible edits that provably reduce cost without degrading the agent's recall. ## Core Features & Use Cases - Consent-gated fix loop: Reviews ranked token sinks from caveman learn report --json and proposes one fix at a time, applying each edit only after explicit user approval. - Net-token-negative gate: Re-measures every applied trim and reverts any edit that fails to lower tokens per turn. - Cavemem offload: Moves recurring context blocks into cavemem with a pointer left behind, verifying recall works before trimming the source so the agent never loses access to the content. - Skill distillation holdouts: For repeated tool-step procedures, writes a skill and starts an on/off experiment instead of claiming unproven savings. - Use Case: After running a Caveman learn scan showing a heavy CLAUDE.md as the top sink, use this Skill to dry-run a candidate trim, approve the diff, confirm the token reduction, and record the outcome for longitudinal tracking. ## Quick Start Ask the agent to run a caveman learn report and walk you through fixing the top token sinks one approved edit at a time.
