Pi

Opinionated config for Pi. Strong model, small harness, strict workflow.

• Engine: openai-codex/gpt-5.5, high thinking.
• Taste: suckless, A Philosophy of Software Design, The Pragmatic Programmer.
• Human owns direction. Agent owns evidence.

Behavior contract: agent/SYSTEM.md.

Install

git clone https://github.com/drsh4dow/.pi ~/.pi
bun add -g @mariozechner/pi-coding-agent   # or: npm i -g
pi install npm:pi-web-access npm:pi-questions npm:pi-delegate

Loop

1. /do-plan <task> — read-only. Explores, researches, interviews, emits a plan.
2. Approve or revise.
3. /do-work — small slices, each verified before claiming done.
4. Review diff + evidence. Commit only on accept.

/grill-me for stress-testing a design first. The concepts behind the /grill-me prompt applied in /do-plan, so this prompt is useful only when there is still doubts in the plan on both sides.

Prompts (agent/prompts/)

We use prompts, not skills. Skills are injected into the context, so it makes sense for things that must be triggered automatically. But for deterministic things, it makes a lot more sense to use prompts.

do-plan · do-work · grill-me · improve-codebase-architecture · write-prd · prd-to-issues · write-skill

Skills (agent/skills/)

verification-before-completion · context7 · agent-browser

Packages

• pi-web-access — search + URL/GitHub/PDF/video extraction.
• pi-questions — ask_questions TUI.
• pi-delegate — delegate to isolated child sessions. The only subagent primitive.

Runtime state (auth.json, sessions/, run-history.jsonl, caches, .env) is gitignored and those will get populated with your data by running the agent.