Caveman separates shorter answers from compressed tool inputs
The maintainer’s skill and proxy change different token paths, so savings must be compared against the right baseline.
Caveman provides a concise-output skill and a proxy that reduces what an agent reads from some tool outputs. These are different mechanisms: asking for a shorter generated answer can reduce output tokens, while compressing a tool result before the next model call can reduce input tokens.
The current maintainer README reports results for specific tasks and baselines, including conditions where proxy overhead worsened a session. Its advertised 65% reduction is not a universal bill reduction. Repository popularity and an example of fewer words do not establish task quality or savings for a particular agent.
For a useful trial, compare billed input/output usage and task correctness against a concise-answer baseline, including the proxy’s added instructions. Inspect whether omitted tool content contains an error, negation or identifier needed by the task. Shortening text after it has already been generated cannot recover already billed output tokens.
Editorial correction (2026-10-13 UTC): Separated input/output mechanisms and removed universal savings and undated popularity claims.