rpi

AI Agent: Research-Propose-Plan-Implement Flow

AI coding agents are capable -- the challenge is steering them. Without structure, you end up re-running prompts hoping the output lands closer to what you actually need. RPI gives you a framework to direct that capability: staged decisions, reviewable artifacts, and behavioral specs that keep work on track.

Each stage produces a document you can read, edit, and approve before the next one starts. A compiled Go CLI handles the bookkeeping so the LLM spends its tokens on thinking, not parsing. Built for Claude Code and OpenCode, but the methodology works with any AI coding tool.

Quick Start

In Claude Code:

/plugin marketplace add A-NGJ/rpi
/plugin install rpi@rpi
/rpi:rpi-setup

That's it. The first command adds this repo as a plugin marketplace; the second installs the rpi plugin from it (skills, hooks, MCP server); /rpi:rpi-setup fetches the matching rpi binary from GitHub Releases into ~/.rpi/bin/rpi. Re-running /rpi:rpi-setup upgrades the binary. First-time only: restart Claude Code after /rpi:rpi-setup so the rpi MCP server can launch — it tried to start at session-open but the binary wasn't there yet. See the full Installation section for OpenCode, standalone CLI, and from-source paths.

See It in Action

Add rate limiting to an API in four commands:

/rpi-research How does the API middleware chain work?

Claude explores your codebase conversationally. You discuss findings, no artifact required.

/rpi-propose Add per-endpoint rate limiting for authenticated and anonymous users

Claude presents 2-3 options with pros/cons tied to your codebase. You pick one. Writes .rpi/designs/.

/rpi-plan .rpi/designs/2026-03-04-api-rate-limiting.md

Breaks work into phases (core module, middleware integration, configuration), each with file changes, verification commands, and success criteria. Writes .rpi/plans/.

/rpi-implement .rpi/plans/2026-03-04-api-rate-limiting.md

Implements phase-by-phase -- runs tests, commits, and advances automatically when checks pass. Pauses only when manual verification is needed or something diverges from the plan.

Install (Claude Code)

Add this repo as a plugin marketplace, install the rpi plugin from it, then run the one-step setup to fetch the matching rpi binary:

/plugin marketplace add A-NGJ/rpi
/plugin install rpi@rpi
/rpi:rpi-setup

/plugin marketplace add A-NGJ/rpi registers this repo's .claude-plugin/marketplace.json. /plugin install rpi@rpi installs the plugin from it (skills, hooks, MCP server). /rpi:rpi-setup downloads the binary from GitHub Releases, verifies its SHA256 against checksums.txt, and installs it to ~/.rpi/bin/rpi. It writes nothing outside that directory. Re-running /rpi:rpi-setup upgrades the binary in place; no other state is touched. See the plugin README for the marketplace listing.

Command names for plugin users. Skill folders carry the rpi- prefix so the trigger surface stays unambiguous even when Claude Code's slash-command picker strips the plugin namespace. Triggers are /rpi:rpi-plan, /rpi:rpi-implement, /rpi:rpi-verify, etc. (paralleling the standalone /rpi-plan, /rpi-implement, …). The MCP server name (rpi) and tool prefix (mcp__rpi__*) are unchanged.

If you previously installed RPI via rpi init --global, run rpi uninstall --global before /rpi:rpi-setup — the plugin refuses to overwrite a standalone install.

Alternative install (opencode / manual / non-Claude-Code)

For environments where the Claude Code plugin marketplace is unavailable (locked-down systems, opencode users, manual setups), use the standalone CLI:

curl -sSfL https://raw.githubusercontent.com/A-NGJ/rpi/main/install.sh | bash
rpi init                                         # current directory, Claude Code
rpi init /path/to/project                        # different directory
rpi init /path/to/project --target opencode      # OpenCode

This creates:

• .claude/ (or .opencode/) -- Agent Skills
• .rpi/ -- Directory for all pipeline artifacts (tracked in git by default)
• CLAUDE.md (or AGENTS.md) -- Project-level instructions for the AI
• MCP server registration (Claude Code only) -- auto-registers rpi serve so the AI calls typed tools instead of shelling out

To update the rpi binary itself:

rpi upgrade         # download and install the latest release