My personal pi configuration — agents, skills, extensions, and prompts that shape how pi works for me.
Clone this repo directly to ~/.pi/agent/ — pi auto-discovers everything from there (extensions, skills, agents, AGENTS.md, mcp.json). No symlinks, no manual wiring.
# 1. Install pi (https://github.com/badlogic/pi)
# 2. Clone this repo as your agent config
mkdir -p ~/.pi
git clone [email protected]:umgbhalla/pi-config ~/.pi/agent
# 3. Run setup (installs packages + extension deps)
cd ~/.pi/agent && ./setup.sh
# 4. Add your API keys to ~/.pi/agent/auth.json
# 5. Restart picd ~/.pi/agent && git pullThis config uses subagents — visible pi sessions spawned in cmux terminals. Each subagent is a full pi session with its own identity, tools, and skills. The user can watch agents work in real-time and interact when needed.
- Subagents — visible cmux terminals running pi. Autonomous agents self-terminate via
subagent_done. Interactive agents wait for the user. - Agent definitions (
agents/*.md) — one source of truth for model, tools, skills, and identity per role. - Plan workflow —
/planspawns an interactive planner subagent, then orchestrates workers and reviewers. - Iterate pattern —
/iterateforks the session into a subagent for quick fixes without polluting the main context.
Specialized roles with baked-in identity, workflow, and review rubrics.
| Agent | Model | Purpose |
|---|---|---|
| planner | Opus 4.6 | Interactive brainstorming — clarify, explore, validate design, write plan, create todos |
| scout | Haiku 4.5 | Fast codebase reconnaissance — gathers context without making changes |
| worker | Opus 4.6 | Implements tasks from todos, commits with polished messages |
| reviewer | Codex 5.3 | Reviews code for quality, security, correctness (review rubric baked in) |
| researcher | Opus 4.6 | Deep research using parallel.ai tools + local repo inspection |
| synthesizer | Opus 4.6 | Fan-in role — merges multiple artifacts into one concise, source-grounded memo |
| debugger | Codex 5.3 | Reproduces bugs, isolates root cause, applies the smallest verified fix |
| visual-tester | Opus 4.6 | Visual QA — navigates web UIs via Chrome CDP, spots issues, produces reports |
| auditor | Codex 5.3 | Deep codebase audit — security, architecture, dependencies, operational risk |
| autoresearch | GPT-5.4 | Autonomous experiment loop — runs, measures, and optimizes iteratively |
Loaded on-demand when the context matches.
| Skill | When to Load |
|---|---|
| commit | Making git commits (mandatory for every commit) |
| code-simplifier | Simplifying or cleaning up code |
| github | Working with GitHub via gh CLI |
| iterate-pr | Iterating on a PR until CI passes |
| learn-codebase | Onboarding to a new project, checking conventions |
| session-reader | Reading and analyzing pi session JSONL files |
| skill-creator | Scaffolding new agent skills |
| cmux | Managing terminal sessions via cmux |
| presentation-creator | Creating data-driven presentation slides |
| add-mcp-server | Adding MCP server configurations |
| Extension | What it provides |
|---|---|
| answer/ | /answer command + Ctrl+. — extracts questions into interactive Q&A UI |
| bg-sessions/ | Background session management — keep sessions running while switching |
| cmux/ | cmux integration — notifications, sidebar, workspace tools |
| cost/ | /cost command — API cost summary |
| pi-mono/ | Dev tooling — diff viewer, file browser, prompt URL widget, redraws, TPS counter |
| reload-and-resume/ | reload_and_resume_runtime tool — reload pi, then automatically continue in the same session |
| restart/ | /restart command — restart the current session |
| todos/ | /todos command + todo tool — file-based todo management |
| watchdog/ | Monitors agent behavior |
| Command | Description |
|---|---|
/plan <description> |
Start a planning session — spawns planner subagent, then orchestrates execution |
/subagent <agent> <task> |
Manual one-off subagent spawn when you explicitly want it |
/iterate [task] |
Fork session into interactive subagent for quick fixes |
/iterate --agent <agent> [task] |
Typed self-fork — keep current context but adopt a named agent role |
agent_group(...) |
Standard orchestration primitive — launch one or many subagents and collect one grouped result |
active_subagents(...) |
Inspect running subagents from the main session |
message_subagent(...) |
Send a nudge/follow-up into a running subagent |
/switch |
Switch sessions while optionally keeping current one running in background |
/bg [list|kill|attach|logs|clear] |
Manage background sessions |
/answer |
Extract questions into interactive Q&A |
/todos |
Visual todo manager |
/cost |
API cost summary |
/resume |
Switch to a previous session |
Installed via pi install, managed in settings.json.
| Package | Description |
|---|---|
| pi-interactive-subagents | Subagent runtime + agent_group(...), /plan, /subagent, /iterate |
| pi-parallel | Parallel web search, extract, research, and enrich tools |
| pi-smart-sessions | AI-generated session names |
| pi-autoresearch | Autonomous experiment loop with dashboard |
| pi-mcp-adapter | MCP server integration |
| glimpse | Native macOS UI — dialogs, forms, visualizations |
| chrome-cdp-skill | Chrome DevTools Protocol CLI for visual testing |
Extensions from mitsuhiko/agent-stuff: answer, todos
Skills from mitsuhiko/agent-stuff: commit, github
Skills from getsentry/skills: code-simplifier