@casualjim/pi-caveman

Caveman skills packaged for Pi

Packages

Package details

extensionskill

Install @casualjim/pi-caveman from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:@casualjim/pi-caveman
Package
@casualjim/pi-caveman
Version
0.1.2
Published
Jun 1, 2026
Downloads
122/mo · 37/wk
Author
casualjim
License
MIT
Types
extension, skill
Size
76.5 KB
Dependencies
0 dependencies · 0 peers
Pi manifest JSON
{
  "skills": [
    "skills"
  ],
  "extensions": [
    "extensions/caveman"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-caveman

Pi-native port of JuliusBrussee/caveman: terse agent communication skills and Pi-native session hooks that reduce output tokens while keeping technical accuracy.

Install

pi install npm:@casualjim/pi-caveman

For local development from this monorepo:

pi install ./packages/pi-caveman

Included skills

  • caveman — persistent terse communication mode with lite, full, ultra, wenyan-lite, wenyan-full, and wenyan-ultra levels.
  • caveman-commit — terse Conventional Commit message generator.
  • caveman-review — terse, actionable code review comments.
  • caveman-compress — compress natural-language memory files while preserving code, commands, paths, URLs, and structure.
  • caveman-help — quick reference for package skills and modes.
  • caveman-stats — documents stats support status for Pi.
  • cavecrew — guidance for compressed subagent delegation in Pi.

Pi registers skills as /skill:<name> commands when skill commands are enabled. Natural-language triggers in each skill also work through normal skill discovery.

Pi-native hooks

pi-caveman includes a Pi extension at extensions/caveman that mirrors upstream Claude Code hook behavior without installing Claude hooks or mutating ~/.claude:

  • session_start loads the default Caveman mode, writes safe Pi mode state, and injects filtered skills/caveman/SKILL.md rules as hidden context.
  • input tracks /skill:caveman, /skill:caveman <mode>, /skill:caveman-commit, /skill:caveman-review, /skill:caveman-compress, natural-language enable/disable, stop caveman, and normal mode.
  • before_agent_start reinforces active base Caveman mode each turn. Independent modes (commit, review, compress) do not inject base reply rules because their skills own behavior.

Default mode is full. Override with CAVEMAN_DEFAULT_MODE or config JSON at the Pi Caveman config path. Set CAVEMAN_DEFAULT_MODE=off to disable startup activation.

Known limitations

This package ports Caveman behavior to Pi. It activates only the Pi-native extension hooks described above. It does not activate upstream installers, Claude Code hooks, statusline integrations, Codex/Gemini/Cursor plugin manifests, or any non-Pi agent configuration.

caveman-stats in upstream Caveman is implemented through Claude Code hooks and Claude transcript logs. The Pi extension does not fake token savings or read Claude logs. A future Pi stats/status API could add native session-token stats.

cavecrew includes upstream prompt resources under agents/, but Pi subagent execution still depends on the subagents configured in the running Pi environment.

Development

npm test --workspace @casualjim/pi-caveman
npm run typecheck --workspace @casualjim/pi-caveman
npm run test:e2e --workspace @casualjim/pi-caveman

The e2e test installs this package into an isolated Pi home and invokes representative skills through the real pi CLI.

Attribution

This package is a Pi port of Caveman by Julius Brussee: