The Pi extensions and skills I use to keep long agent sessions useful without building a fake operating system around them.

Everything is published as a separate npm package. Install a bundle for the full setup, or pick only what you need. Revolutionary stuff. A table.

Pi packages run with your local permissions. You can obviously trust me, a stranger on the internet with a folder called pi-stuff, but read the source before installing it anyway.

pi install npm:@howaboua/pi-stuff
# or
pi install npm:@howaboua/pi-extensions
pi install npm:@howaboua/pi-skills

pi-codex-conversion is separate because it changes Pi's tool surface for GPT/Codex models. pi-codex-web-run and pi-codex-imagegen are separate because they use Codex endpoints and login. omarchy-help is separate because it targets Arch desktops configured with Omarchy.

Install pi-browser separately for logged-in browser control. The standalone pi-subdir-agents loader is also opt-in.

Pi discovers installed skills automatically and loads them when a task matches. Use /skill:<name> when you want to invoke one explicitly.

Map an unfamiliar repo, set /marker once the useful context is in place, implement one coherent change, and run /review. After triage and QA, /end carries the accepted result forward. Broad changes get a codebase-hygiene pass.

For UI work, I give the agent references first—apps, screenshots, and interface details I like—then iterate through browser inspection and screenshots. One-shotting a good frontend is mostly a party trick.

See CHANGELOG.md. Package-level changelogs remain beside packages that have them.

Individual packages include their own license files. They are MIT-licensed unless noted in the package directory.