This is my Pi toolbox: the extensions and skills I use to keep agent sessions moving without building a whole fake operating system around them.

Everything here is still published as separate npm packages. Install the full bundle if you want my setup, or pick the one package you actually 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 maybe read the package before installing it anyway.

GitHub does not render proper copy-button codeblocks inside Markdown tables. So you get a long list instead. At least the commands copy cleanly. Or send your clanker here to argue about which ones you actually want.

General setup: extensions plus shareable skills. Excludes Codex conversion and Omarchy because those depend on your model/workstation setup.

pi install npm:@howaboua/pi-stuff

General extension packages. Excludes pi-codex-conversion; install that separately if you run Codex/GPT models and want native-tool adaptation.

pi install npm:@howaboua/pi-extensions

Shareable skill packages. Excludes omarchy-help because not everyone is running my kind of desktop setup.

pi install npm:@howaboua/pi-skills

Codex-style tools for Pi: exec_command, write_stdin, apply_patch, image tools, native Codex web search, and prompt/tool adaptation.

pi install npm:@howaboua/pi-codex-conversion

Gives the agent a change_reasoning tool so it can raise/lower reasoning level when the work changes shape.

pi install npm:@howaboua/pi-auto-reasoning-tool

Adds /marker and /end for long sessions. Set a useful return point, summarize what was accomplished, then keep going.

pi install npm:@howaboua/pi-auto-trees

Adds /review, an isolated review subagent that checks the right branch/range and returns findings for the main agent to address.

pi install npm:@howaboua/pi-subagent-review

Adds semantic_grep, a meaning-based code/docs search tool backed by local SQLite indexes and OpenAI-compatible embeddings.

pi install npm:@howaboua/pi-semantic-grep

Adds vent, a small tool for logging repeated workflow friction into VENT.md.

pi install npm:@howaboua/pi-vent

Adds explore_subagent, discovery-only shallow/deep subagents for reading and summarizing code without editing files.

pi install npm:@howaboua/pi-explore-subagents

Adds /skills, /workflows, workflow capture, /learn, and nested AGENTS.md context loading.

pi install npm:@howaboua/pi-markdown-workflows

Side-session questions with explicit injection back into the main chat. Useful when you want a tangent without derailing the main thread.

pi install npm:@howaboua/pi-smart-btw

KISS local memory for Pi based on global AGENTS.md.

pi install npm:@howaboua/pi-memories

Refactor/audit posture for agent-built code: fewer godfiles, clearer ownership, less duplication, better traversability.

pi install npm:@howaboua/pi-skill-agent-native-hardening

Rewrites text so it sounds specific, human, and less like a polite SaaS brochure.

pi install npm:@howaboua/pi-skill-anti-ai-copy

Browser inspection/control through Chrome DevTools Protocol. Based on pasky/chrome-cdp-skill, with local Pi packaging changes.

pi install npm:@howaboua/pi-skill-chrome-cdp

A generic GitHub issue/PR workflow with gh, branches, validation, PR bodies, and review triage.

pi install npm:@howaboua/pi-skill-gh-issue-pr-flow

Writes, audits, and trims scoped AGENTS.md files without turning them into repo summaries.

pi install npm:@howaboua/pi-skill-agents-md

Helps design compact Pi tool names, descriptions, schemas, snippets, and guidelines that models call correctly.

pi install npm:@howaboua/pi-skill-model-facing-api-design

Looks up external or local repos as reference context, then returns evidence-backed findings.

pi install npm:@howaboua/pi-skill-project-reference-research

Helps design, write, package, and tighten reusable agent skills.

pi install npm:@howaboua/pi-skill-skill-creator

Generic Arch + Omarchy workstation maintenance. Install separately and customize it for your own machine.

pi install npm:@howaboua/pi-skill-omarchy-help

A normal session is not fancy:

1. Start Pi and ask the agent to familiarise itself with the repo, usually with discovery subagents. (pi-explore-subagents, pi-semantic-grep)
2. Pull in GitHub issues, or ask the agent to group open issues into something PR-sized. (pi-skill-gh-issue-pr-flow)
3. Use /marker once the agent has enough baseline context. (pi-auto-trees)
4. Work on one issue or feature. (pi-codex-conversion, pi-auto-reasoning-tool)
5. Run /review, fix what is actually worth fixing, then review again if needed. (pi-subagent-review)
6. Open a PR and use external review feedback as another pass. (pi-skill-gh-issue-pr-flow)
7. Do manual QA and try to break the feature. (your eyes, sadly still required)
8. Use /end to summarize what changed and advance the marker. (pi-auto-trees)
9. Continue from the new marker for the next feature. (pi-auto-trees)
10. After larger changes, ask for a hardening pass to modularise the result and remove obvious slop. (pi-skill-agent-native-hardening)

The point is not loops, worker swarms, or pretending the agent is magic. It is a few raw Pi sessions, clear context boundaries, review passes, and enough tooling to stop long sessions from turning into archaeology.

For UI work, I usually give the agent a reference frame first: apps to inspect, screenshots, bits of interface I like. The agent builds a mock, then I iterate with browser control, screenshots, and human taste. One-shotting a good frontend is mostly a party trick.

A few useful skills are intentionally not in the bundles because they are either very taste-specific, UI/design-specific, or better installed from their own source:

• impeccable by Peter Bakaus: high-quality frontend/interface generation.
• make-interfaces-feel-better by Jakub Krehel: UI microdetails, interaction polish, spacing, shadows, motion, all the fiddly bits.
• design-md: useful when working with Google Labs’ DESIGN.md spec and @google/design.md CLI.
• agent-pages: useful for rich local proposal/report/mockup surfaces, but not something I want to force into the default bundle.
• React Grab by Aiden Bai is not a Pi skill here, but it is extremely useful for React UI iteration.

See CHANGELOG.md. Package-level changelogs remain next to their packages where they exist.

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