opencode-caveman

OpenCode plugin that injects caveman compression rules into every session — automatically, without manual activation.

Full credit: This plugin is a direct adaptation of caveman by Julius Brussee (52k+ stars). All compression rules, intensity levels, and the core philosophy — "why use many token when few token do trick" — originate from that project. This repo only packages those rules for the OpenCode plugin system.

How it works

Three layers, each strengthening the last:

| Layer | File | Mechanism | |-------|------|-----------| | Always-on rules | AGENTS.md | Loaded into system prompt at session init. Survives /clear. | | Per-turn enforcement | index.ts → experimental.chat.system.transform | Rules injected into every LLM call. Model sees them every turn — no drift. | | Compaction survival | index.ts → experimental.session.compacting | Rules injected into compaction prompt context. Model's summary includes caveman state — survives /compact. | | Live mode switching | index.ts → command.execute.before + caveman_set_level tool | /caveman commands intercepted before reaching model. Mode written to ~/.config/opencode/.caveman-active — persists across sessions. | | On-demand skill | skills/caveman/SKILL.md | Loadable via skill picker for intensity switching docs. |

Install

Add to ~/.config/opencode/opencode.json:

{
  "plugin": ["@mumme-it/opencode-caveman"]
}

Modes

Switch intensity via ctrl+p → Skills → caveman-lite / caveman-full / caveman-ultra.

| Command | Effect | |---------|--------| | /caveman | Activate full mode | | /caveman lite | No filler. Articles kept. Professional but tight. | | /caveman full | Drop articles, fragments OK. Classic caveman. (default) | | /caveman ultra | Max compression. Prose abbreviations, arrows for causality. | | stop caveman | Return to normal mode. |

Mode persists across sessions via ~/.config/opencode/.caveman-active.

Credit

| | | |---|---| | Original | caveman | | Author | Julius Brussee | | Stars | 52k+ | | Philosophy | why use many token when few token do trick | | Benchmarked savings | ~65% output token reduction (range 22–87%) | | Accuracy | 100% — technical substance preserved |