ralphai

ralphai

Put your AI coding agent on autopilot.

Point Ralphai at a GitHub issue. It drives your AI coding agent through fresh-session iterations — with branch isolation, build/test feedback loops, and stuck detection — until the PR is ready.

Why Ralphai?

AI coding agents get worse the longer they run. As the conversation grows, the model drops older context: it forgets what it tried, repeats mistakes, and drifts.

Ralphai avoids this by starting each iteration with a fresh agent session: just the plan and a progress log. No conversation history to lose. No drift.

• No context rot — iteration 10 is as sharp as iteration 1
• Fresh feedback — real build output every cycle, never recalled from memory
• Stuck detection — stops burning tokens when progress stalls

How it works →

Try It Now

npx ralphai init --yes       # auto-detect agent and project setup
ralphai run 42               # run GitHub issue #42

Ralphai creates an isolated worktree, drives your agent through build/test feedback loops, and opens a draft PR when done. init --yes auto-detects installed agents (checking Claude Code and OpenCode first) and your project's build/test commands.

Install

Requires Node.js 18+ (or Bun/Deno) and a supported CLI agent.

npm install -g ralphai                    # install the CLI
npx skills add mfaux/ralphai -g           # install agent skills (recommended)

Included skills

• write-a-prd — create a product requirements document through interactive interview
• prd-to-issues — decompose a PRD into vertical-slice GitHub sub-issues
• triage-issue — investigate bugs and create TDD fix plans
• tdd — test-driven development with red-green-refactor loops
• improve-codebase-architecture — find and propose module-deepening refactors
• request-refactor-plan — plan structural changes with tiny, verifiable commits
• ralphai-planning — write local plan files for autonomous execution

How It Works

1. Plan with your agent

Use the included skills to turn ideas into GitHub issues your agent can execute:

• Features — ask your agent to write-a-prd, then prd-to-issues to decompose it into labeled sub-issues. Ralphai processes them sequentially on one branch and opens a single aggregate PR.
• Bugs & small tasks — ask your agent to triage-issue to investigate and create a standalone issue. Each gets its own branch and PR.

Both skills label the issues automatically (ralphai-prd / ralphai-standalone). You can also label issues by hand. Labels are configurable.

2. Run

ralphai run 42               # run a specific issue (PRD or standalone)
ralphai run --drain           # process all eligible issues until the queue is empty
ralphai run --dry-run         # preview without changing anything

Each run creates an isolated worktree on a conventional <type>/<slug> branch, iterates the agent with fresh build/test feedback, and opens a draft PR when done. Stuck sub-issues are skipped so progress continues.

3. Review the PR

Ralphai surfaces extracted learnings in the draft PR — patterns the agent discovered during implementation. Promote useful ones to AGENTS.md or skill docs. More on learnings →

Interactive mode

Running bare ralphai opens a TUI to browse the pipeline, pick issues, and launch runs without memorizing subcommands.

Local plan files

You can also drive Ralphai with local markdown files instead of GitHub issues — see Workflows → Local plan files.

Day-to-Day

ralphai                  # open the interactive menu
ralphai status           # see what's queued, running, and completed
ralphai stop             # stop the active runner (or --all)
ralphai run --resume     # commit dirty state and continue
ralphai doctor           # validate setup (agent, feedback, config, git)
ralphai reset            # reset stuck plans
ralphai clean            # remove archived plans and orphaned worktrees

Press Ctrl-C during a headless run to stop cleanly after the current iteration. Work is preserved and ralphai run picks up where it left off.

Supported Agents

Ralphai works with any CLI agent that accepts a prompt argument. Claude Code and OpenCode are actively tested.

Agent commands

Agent	Command	Status
Claude Code	claude -p	Tested
OpenCode	opencode run --agent build	Tested
Codex	codex exec	Untested
Gemini CLI	gemini -p	Untested
Aider	aider --message	Untested
Goose	goose run -t	Untested
Kiro	kiro-cli chat --no-interactive	Untested
Amp	amp -x	Untested

Reference

• CLI Reference — all commands, flags, and configuration
• How Ralphai Works — context rot, feedback loops, stuck detection
• Hooks, Gates, and Prompt Controls — advanced customization
• Workflows — common patterns and recipes
• Worktrees — parallel runs in isolated directories
• Troubleshooting — common issues and fixes

Acknowledgements

• Ralph by Geoffrey Huntley — creator of the technique behind the loop
• Getting Started With Ralph by Matt Pocock
• mattpocock/skills — inspiration for the planning and TDD skills
• Sandcastle by AI Hero — demonstrated the ephemeral-container-per-invocation pattern for agent sandboxing
• Caveman by Julius Brussee — inspiration for the caveman communication skill
• Vercel CLI for CLI DX inspiration

License

MIT