이것은 실제, 매일 구동 구성 파이 코딩 에이전트 (@earendil-works/pi-coding-agent]) 실행 100% 로컬에 대한 LM Studio에 대한 맥북. 클라우드 LLM 없음. API 수수료 없음. 인터넷은 없습니다. workhorse 모델은 Apple Silicon의 Qwen3.5 122B (MLX)이며, 이 하네스는 제로 도구 오류를 가진 헤드리스 모드에서 완료, 확인된 웹 사이트와 로컬 모델 간의 차이입니다.

이 repo의 모든 것은 소스 기계에 ~/.pi/agent/에 살고있다. 그것을 복제, 당신이 원하는 조각을 훔쳐, 또는 자신의 pi 구성에 전체 것을 드롭.

** 영수증.** 이 정확한 설정으로 로컬 호스트 122B는 pi -p]를 통해 전체 생산 웹 사이트 끝을 구축했습니다 (클래스에서 인간 없음) : 디자인 시스템, 복사, GSAP 애니메이션, 비디오 영웅, 자체 브라우저 기반 시각적 검증 및 전달 생산 빌드. 4개의 연속 헤드리스 빌드를 통해 세션 감사는 0개의 툴 콜 오류와 0개의 편집 복원 루프를 발견했습니다.

---

내용 표

1. 왜 지역, 그리고 왜 마구
2. 모두 함께 적합
3. 연장
4. 모든 것을 변경하는 검증 루프
5. 기술
6. 헤드리스 모드, 그것은 모든 지불 떨어져
7. 명세.

Mark’s Pi Harness

This is the real, daily-driven config for the pi coding agent (@earendil-works/pi-coding-agent) running 100% local against LM Studio on a MacBook. No cloud LLM. No API bill. No internet required. The workhorse model is Qwen3.5 122B (MLX) on Apple Silicon, and this harness is the difference between a local model that flails and a local model that one-shots complete, verified websites in headless mode with zero tool errors.

Everything in this repo lives in ~/.pi/agent/ on the source machine. Clone it, steal the pieces you want, or drop the whole thing into your own pi config.

The receipts. With this exact setup, a locally-hosted 122B has built full production websites end to end through pi -p (headless, no human in the loop): design system, copy, GSAP animation, video heroes, its own browser-based visual verification, and a passing production build. A session audit across four consecutive headless builds found zero tool-call errors and zero edit-retry loops.

---

Table of contents

1. Why local, and why a harness
2. How it all fits together
3. The extensions
4. The verification loop that changes everything
5. The skills
6. Headless mode, where it all pays off
7. Hard-won lessons
8. Design memory
9. Repo layout
10. Setup

---

Why local, and why a harness

A frontier cloud model tolerates a sloppy setup. A local 122B does not. It will write CSS for HTML it never created, declare victory from log output, retry a failing command forever, and hang for 35 minutes on brew install when the wifi drops on a plane. None of that is fixed by a bigger prompt. It is fixed by infrastructure: tools that force the model to look at its work, gates it cannot talk its way past, guards that turn environmental chaos (offline, port conflicts, giant tool outputs) into structured signals it can act on.

That is what this harness is. Three layers, all included here:

• AGENTS.md, the standing rules. Working style, verification requirements, server discipline, code discipline, and the care rules for destructive actions.
• extensions/, 15 TypeScript modules pi loads at startup. They add tools (browser_check, bash_background, task, web_search, facts_recall), enforce discipline (gates, freshness, output caps, budget awareness), and adapt to the machine (offline detection, per-model prompt tuning).