Compound Engineering for Pi

Compound Engineering adapted for Pi. 84 skills, 9 legacy workflow prompts for compatibility, and a native compatibility layer, all installable in one command.

Install

Use this package as the Pi install/distribution layer:

pi install npm:compound-engineering-pi -l

Reload your session, then you're ready:

/reload

Quick start

Legacy prompt aliases still work:

pi -p "/workflows-plan Build a todo app"
pi -p "/workflows-work"
pi -p "/workflows-review"
pi -p "/workflows-compound"

And the newer upstream workflow skills are bundled too:

pi -p "/skill:ce:plan Build a todo app"
pi -p "/skill:ce:work"
pi -p "/skill:ce:review"
pi -p "/skill:ce:compound"

These follow the core loop: Plan, Work, Review, Compound.

What's included

Legacy workflow prompts

| Prompt | What it does | |--------|-------------| | /workflows-plan | Create a structured project plan | | /workflows-work | Execute work from a plan | | /workflows-review | Multi-agent code review | | /workflows-compound | Document learnings for future sessions | | /workflows-brainstorm | Explore ideas before planning | | /deepen-plan | Add depth to an existing plan | | /test-browser | Run browser tests on your pages | | /feature-video | Record a video walkthrough of a feature | | /resolve_todo_parallel | Resolve all pending todos in parallel |

84 skills

This package now bundles the newer upstream Compound Engineering skill set for Pi, including the ce:* workflow skills, reviewer personas, planning/document-review personas, browser/video workflows, and the older specialized research/review skills that still power the legacy prompts.

Run pi skills after installing to see the full list.

Pi compatibility extension

Four tools that bridge common agent workflows into Pi:

• ask_user_question - prompt the user for input
• subagent - orchestrate tasks (single, parallel, chain)
• mcporter_list - list available MCP servers
• mcporter_call - call MCP server tools

Subagent output behavior:

• single runs return the full subagent output
• chain runs return the final step output plus a step summary
• parallel runs stay compact by default, but support includeOutputs: true
• if you install a richer pi-subagents package, this package automatically defers to it

Subagent model selection (compatibility tool):

• pass model: "haiku" or a provider/modelId (optionally with a :thinking suffix) for a single task
• set a top-level model as the default for tasks[] or chain[]; each entry can override it
• omitted models keep the child Pi process's configured default, not necessarily the parent session's active model
• an empty or whitespace-only entry model clears the top-level default

Example subagent tool arguments for a cheaper scout and a stronger reviewer:

{
  "model": "haiku",
  "tasks": [
    { "agent": "repo-research-analyst", "task": "Find the relevant code" },
    { "agent": "kieran-typescript-reviewer", "task": "Review correctness", "model": "sonnet" }
  ]
}

Use models available and authenticated in your Pi installation. See the Pi guide for all three modes.

Conversion and sync

For converter usage, prefer the upstream package:

bunx @every-env/compound-plugin install compound-engineering --to pi

This repo’s bundled package should mainly be used via pi install npm:compound-engineering-pi.

The vendored plugins/compound-engineering snapshot and bundled Pi skills can be refreshed from the main plugin repo with:

bun run sync:upstream

By default this syncs from ../compound-engineering-plugin. Override with COMPOUND_PLUGIN_SOURCE=/path/to/compound-engineering-plugin if needed.

Optional: MCP support via MCPorter

For MCP interoperability, install MCPorter:

npm i -g mcporter

Maintainer workflow

This repo now follows an upstream-first model:

• make converter/content changes in compound-engineering-plugin first
• refresh this package with bun run sync:upstream
• keep this repo focused on Pi packaging, Pi docs, and Pi-specific compatibility behavior

See MAINTENANCE.md for the exact boundary and workflow.

Credits

Built on the Compound Engineering plugin by Every and Kieran Klaassen. Read the original writeup.

License

MIT