caveman
shorter replies from the agent, so fewer output tokens.
- status
- tests only ships with colonizer, but is covered by tests only: it has not yet done real work in a colony.
- version
- v3.1.0
- pinned
- tag v3.1.0 · archive sha256 676ae2e9
- licence
- mit ruleset
- source
- juliusbrussee/caveman ↗
- docs
- token savings
what it does for you
caveman is a compressed reply style. switched on, the agent answers tersely at lite, full or ultra, while code, commands and exact errors stay intact and pull request descriptions, questions to you, memory proposals and code comments stay in plain sentences.
what it needs
- the claude code settler
how to turn it on
settings → modules on the claude code settler: the caveman switch and its level (lite, full or ultra). off by default.
how far it has been tested
covered by tests only; it has not run in a colony yet.
how colonizer runs it
- pinned by sha256 in vendor.lock (tag v3.1.0)
- only skills/caveman/SKILL.md (without its frontmatter) and LICENSE are staged, both mit; the compression engine, proxy and mcp server are BSL-1.1 and are neither staged nor used
- the runner puts the ruleset into the system prompt, followed by the one colonizer exception for descriptions, questions, proposals and comments
- the level is the caveman_level setting (lite, full or ultra; default full)
what it costs
not measured as a separate figure: only the ruleset ships, and the runner puts it in the system prompt.