(Beta) The ultimate AI coding harness on top of pi.dev.

Goal

Ship a production-grade coding harness where the agent:

1. uses compressed context before raw tool spam,
2. stays concise and deterministic in output style,
3. remains easy to extend with skills and extensions.

Getting started

1. Install from npm:

   npm install -g ultimate-pi

2. Install package into PI from your project folder:

   pi install ultimate-pi -l

3. Reload PI to activate skills and extensions:

   /reload

4. Optional: install from GitHub Packages instead:

   npm install -g @aryaniyaps/ultimate-pi --registry=https://npm.pkg.github.com

Included skills (upstream repositories)

Skill	Upstream repo	What it does
caveman	juliusbrussee/caveman	Ultra-compressed response style for token-efficient, high-signal communication.
compress	juliusbrussee/caveman	Compresses memory/context files into caveman format while preserving technical content.
lean-ctx	yvgude/lean-ctx	Context runtime and compression toolkit for reads/search/shell/edit flows.
context7-cli	upstash/context7	Pulls current library docs and manages Context7 skills/config from CLI.
scrapling-official	D4Vinci/Scrapling	Official Scrapling skill for web extraction/scraping with adaptive parsing, JS rendering, stealth fetchers, and anti-bot handling.

Runtime guardrail extension

• extensions/lean-ctx-enforce.ts
  • Detects lean-ctx availability (cached check).
  • Overrides built-in read tool to execute through lean-ctx read when available.
  • Overrides built-in bash tool and auto-wraps commands with lean-ctx -c <command> when needed.
  • Falls back to default built-in tools if lean-ctx is unavailable.
  • Adds status command:
    • /lean-ctx-status

Design choices (concise)

1. Lean-ctx-first execution
   Reason: cut context/token waste and keep tool usage structured.

2. Caveman-by-default response style
   Reason: short, exact, low-noise outputs for coding loops.

3. Runtime enforcement in extension code
   Reason: executable extension logic is the single source of truth for lean-ctx routing.

4. Skill-based composition
   Reason: easy to swap/upgrade capabilities without rewiring core package.

5. Model-agnostic compression path
   Reason: compress skill calls PI CLI, so provider/model can change per environment.

Use locally

From repo root:

pi install . -l

Then restart PI or run /reload.

Terminal-Bench evaluation (local only)

Use Harbor custom-agent flow from docs:

npm run bench:terminal

Scaffold files live in bench/terminal-bench/ and are excluded from npm publish via .npmignore.