my-pi-setup

Personal configuration repo for my Pi coding-agent setup.

This repo snapshots the custom Pi extensions and themes I use across projects so the setup can be reviewed, versioned, and restored on a new machine.

What is included

extensions/
  agentyc-mcp/          MCP bridge for agentyc browser automation (66 tools: navigate, click,
                        type, extract, screenshots). Auto-connects on startup. `/agentyc` to status.
  agentmemory/          Cross-session memory tools backed by a local agentmemory server
  contextia/            Contextro/Contextia codebase indexing and semantic search tools
  context-cat/          Animated context-window cat widget, heatmap, and handoff helper
  subagents/            Scout/researcher/reviewer/worker subagent orchestration
  code-improve/         /code-improve and /code-review commands — deep multi-axis code quality review
                        dispatching sub-agents (Architect, Reviewer, Minimalist, Security)
  codegraph.ts          @colbymchenry/codegraph integration tools
  copy-all.ts           Copy/export helper command
  crawl4ai-search.ts    Local Crawl4AI web search and scraping tools
  diff.ts               Diff display helper
  flow-title.ts         Session/title helper
  learn-anything.ts     /learn and /learn-codebase commands for learning workflows
  obsidian-brain.ts     Durable Obsidian wiki memory tools
  opencode-zen-login.ts OpenCode Zen login helper
  design-polish/        /design-polish command, skill, audit helpers, and Remotion video templates
  tps-tracker.ts        Tokens-per-second status tracker
  usage.ts              Usage/cost analysis helper
  yeet.ts               Fast command/helper extension
  zsh-user-bash.ts      Shell/user environment helper

themes/
  github-dark-default.json

What is intentionally excluded

Secrets and machine-local runtime state are not committed:

• .env / .env.*
• auth.json
• node_modules/
• session logs
• provider tokens or API keys

Browser automation (agentyc MCP)

This setup includes an MCP bridge extension that spawns agentyc mcp as a child process, discovers all 66 tools via the Model Context Protocol, and registers them with Pi's tool system so the LLM can call them directly.

Commands:

• /agentyc — show connection status and tool list
• /agentyc restart — restart the MCP server

Prerequisite: uv tool install agentyc (or pip install agentyc)

Install dependencies: cd ~/.pi/agent/extensions/agentyc-mcp && bun install

Context Cat

This setup includes context-cat, an animated Pi widget that watches context usage, shows green/yellow/red heatmap cells for compact/noisy/bad context, and can start a lightly cleaned handoff session before the window gets too risky. The heatmap uses a research-informed feature set: clear intent, slots/constraints, evidence boundaries, output format, and verification anchors are good; length, ambiguity, distractors, repetition, missing slots, weak evidence rules, logs, diffs, errors, oversized entries, and filler-heavy user text become context pressure.

/context-cat
/context-cat-explain
/context-cat-handoff

Install dependencies: cd ~/.pi/agent/extensions/context-cat && bun install

Code quality commands

This setup includes a code-improve extension that runs deep multi-axis code quality review. It combines candid-review methodology with ambitious structural improvement (code judo, anti-spaghetti, 1k-line boundary) and dispatches specialized sub-agents in parallel for a complete audit.

/code-improve [path] [options]
/code-review  (shorthand alias)

Options:

• --harsh / --constructive — tone control
• --focus security|performance|architecture|edge-case — narrow the review
• --auto-commit — auto-commit fixes after user selects them
• --exclude <pattern> — skip matching files

Sub-agents dispatched (parallel):

Agent	Focus
👁️ Code Reviewer	Correctness, security, maintainability, performance
🏛️ Software Architect	Structure, boundaries, code judo, anti-spaghetti, 1k-line gate
🪡 Minimal Change Engineer	Scope discipline, diff size, no premature abstraction
🔒 Security Engineer	Auto-dispatched when auth/data changes

Learning commands

This setup includes a small learn-anything.ts extension that routes learning requests into the Understand-Anything workflow when those skills are installed.

/learn <topic, repo, folder, or concept>
/learn-codebase [path]

Use it when you want Pi to teach you a codebase, architecture, domain model, or knowledge base instead of just making edits.

Install / restore

Clone the repo, then run:

./scripts/install.sh

By default the script installs into:

~/.pi/agent

To restore into a different Pi agent directory:

PI_AGENT_DIR=/path/to/.pi/agent ./scripts/install.sh

After installing, restart Pi or run:

/reload

The context-cat and design-polish extensions have their own local packages. If you want their optional package dependencies available after install, run:

cd ~/.pi/agent/extensions/context-cat && bun install
cd ~/.pi/agent/extensions/design-polish && bun install

How Pi discovers these files

Pi auto-discovers global extensions from:

~/.pi/agent/extensions/*.ts
~/.pi/agent/extensions/*/index.ts

Pi auto-discovers global themes from:

~/.pi/agent/themes/*.json

Update this repo from the local Pi setup

From this repo:

rsync -a --delete --exclude 'node_modules/' ~/.pi/agent/extensions/ extensions/
rsync -a --delete ~/.pi/agent/themes/ themes/
git status

Review the diff before committing so secrets or local state never get added.

License

This repository is licensed under the MIT License.

You can use, copy, modify, and share this setup, but it is provided as-is with no warranty. Third-party tools, package names, trademarks, and services referenced by these extensions remain under their own licenses and terms.