Caveman

why use many token when few do trick — caveman for OpenCode.

Caveman is an ultra-compressed output mode: drop articles, filler, pleasantries and hedging; keep every technical fact, code block, and error string exact. This package is an unofficial OpenCode port of the Caveman rules by Julius Brussee (MIT). Not affiliated with or endorsed by the upstream project.

Install

opencode plugin @miguelmartens/opencode-caveman

That installs the package and adds it to your config. Or add it by hand (global at ~/.config/opencode/opencode.json, or per project):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["@miguelmartens/opencode-caveman"]
}

Or run it from a checkout with no publish step — the repo ships a root opencode.json that does exactly this:

{ "plugin": ["./.opencode/plugins/caveman.mjs"] }

Restart OpenCode after changing config. The mode is active from the next session; no extra setup.

Commands

Command	What it does
/caveman	Activate caveman at full
/caveman lite|full|ultra|wenyan-lite|wenyan-full|wenyan-ultra|off	Set the level (wenyan = wenyan-full)
/caveman-commit	Terse Conventional Commit message for the staged changes
/caveman-review	One line per finding, severity-tagged, no praise
/caveman-help	Quick-reference card
say stop caveman or normal mode	Turn caveman off

Levels

Level	Style
lite	No filler or hedging. Full sentences, professional but tight
full (default)	Drop articles, fragments OK. Classic caveman
ultra	Strictest compression; code symbols and error strings untouched
wenyan-lite / wenyan-full / wenyan-ultra	Classical Chinese registers

How it works

Hook	Effect
config	Registers the commands, the bundled skills, and the cavecrew subagents
experimental.chat.system.transform	Injects the level-filtered ruleset into every turn's system prompt — no drift, survives /clear and compaction
command.execute.before	Persists /caveman <level> to ~/.config/opencode/.caveman-active
chat.message	Turns caveman off on a standalone stop caveman / normal mode message

The injected ruleset is the vendored caveman skill filtered to the active level: only that level's intensity-table row and worked examples survive, so level-independent rules (negation handling, Auto-Clarity, Boundaries) always apply.

cavecrew subagents

Three compressed subagents, registered for OpenCode's task tool:

Agent	Job	Edits
cavecrew-investigator	Locate code, return path:line citations only	denied
cavecrew-builder	Surgical 1–2 file edit, returns a diff receipt	allowed
cavecrew-reviewer	Diff review, one line per finding	denied

They inherit your configured model (upstream pins Haiku; OpenCode users run mixed providers, so the pin is dropped here).

Configuration

Default level resolution, highest priority first:

1. CAVEMAN_DEFAULT_MODE environment variable (lite, full, ultra, wenyan-*, or off)
2. ~/.config/caveman/config.json:

{ "defaultMode": "ultra" }

1. full

Set "off" as the default to disable auto-activation; /caveman still works. /caveman <level> persists until changed, and off persists too. The state file is $OPENCODE_CONFIG_DIR/.caveman-active when that variable is set, else $XDG_CONFIG_HOME/opencode/.caveman-active, else ~/.config/opencode/.caveman-active.

What is not included

• Reading-side compression (logs, tool output, JSON, diffs): that is the upstream caveman proxy, not a skill. Use caveman opencode from @caveman-ai/cli instead; this plugin only owns the output side.
• /caveman-compress: upstream's version needs Python scripts plus an API call and validates its own output mechanically. The plugin does not ship a validator-less copy. Install upstream's skill if you want it.
• /caveman-stats: OpenCode reports session token usage and cost natively.
• Work-pattern skills (investigate-first, lean-build, …): agent-agnostic, available via upstream npx skills add JuliusBrussee/caveman.

If you already installed the upstream skills globally (~/.agents/skills/), OpenCode will list both copies. They are independent; this package's copies are the ones updated with the plugin.

Development

npm test          # node --test, no dependencies

skills/ and agents/ are vendored verbatim from the upstream commit pinned in NOTICE.md. Don't edit them here — re-vendor and update the SHA instead.

License

MIT for this packaging. Vendored skill and agent text is MIT, © Julius Brussee; see NOTICE.md. "Caveman" and the rock logo are trademarks of Julius Brussee.