Caveman: Token Compression

Caveman is a token-compression plugin installed via make install-skills. It reduces output token consumption through progressively terser communication styles. make install-skills copies the skills from the repository pinned as INFINITO_SKILLS_REPOSITORY in the generated .env (override via custom.env, see docs/contributing/customize.md); the skills repository pins every skill in its skills-lock.json and also supports global installs via its own make install. This works universally across all supported agents: Claude Code, Codex, Gemini CLI, Cursor, Copilot, Windsurf, Cline, and more.

Default Mode

Agents MUST operate in /caveman lite mode by default. Lite keeps grammar and readability intact while still reducing output size.

Startup Notification

At the start of every conversation, before doing any other work, the agent MUST display the following block verbatim (substituting the active mode if it has been changed by the user):

[caveman: lite]  Switch: /caveman full · /caveman ultra · /caveman wenyan-lite · /caveman wenyan · /caveman wenyan-ultra · stop caveman

Mode Reference

Command	Style	Use when
/caveman lite	Professional brevity, grammar intact	Default: balanced readability
/caveman full	Fragment-based, drops articles	Context growing, prefer brevity
/caveman ultra	Maximally compressed, telegraphic	Context critical, every token counts
/caveman wenyan-lite	Classical Chinese, semi-formal	Alternative compression style
/caveman wenyan	Full classical Chinese terseness	Deep compression, classical style
/caveman wenyan-ultra	Extreme classical compression	Maximum reduction
stop caveman	Normal prose	User explicitly requests full output

Tier Progression: Token Budget Guidance

As the conversation grows, the agent MUST proactively suggest a higher tier. Base the estimate on observable signals: message count, response sizes, and whether context compression has been triggered by the system.

Approximate token budget consumed	Action
< 20 %	Stay on current tier. No suggestion needed.
≥ 20 %	Recommend upgrading from lite → full
≥ 50 %	Recommend upgrading from full → ultra
≥ 75 %	Strongly recommend ultra (or a wenyan variant); warn that context is running low

Suggestions MUST be short, inline, and non-blocking. Example:

[caveman: lite | ~30 % context used → consider /caveman full]

The agent MUST NOT interrupt work to ask about tier changes. Recommendations appear as a single line at the start of the next response after the threshold is crossed.