cub
Vision → Tasks → Code
The autonomous loop for AI coding agents. Prep your tasks, hit run, walk away.
curl -LsSf https://lavallee.github.io/cub/install.sh | bash
cub init
Set up cub in your project directory
cub prep
Describe your goal — cub breaks it into atomic tasks
cub run
AI executes tasks one by one. Walk away, come back to commits.
Built on the Ralph Wiggum technique — running AI agents in a while loop, but with guardrails.
🎯 The Gap
Too Hands-On
Sitting in IDE, approving every tool call, babysitting the agent
Too Hands-Off
Agents run amok with vague instructions, chaos ensues
🔄 Two Main Events
PREP
Turn your ideas into agent-sized chunks through a guided interview.
- Triage — What are we building?
- Architect — How should it work?
- Plan — Break it into tasks
- Bootstrap — Write to backend
RUN
Let AI execute autonomously with dependency ordering and retries.
- Pick ready tasks
- Generate prompts
- Invoke AI harness
- Commit, log, repeat
✨ Why Cub?
The Ralph Wiggum technique is elegantly simple: run an AI agent in a while loop until the work is done. The trick is making that loop reliable.
cub Pick Ready Task
Find the next task with no blockers. Respects priority, epics, and dependency graph.
cub Generate Prompt
Build structured prompt with context, acceptance criteria, and project guardrails.
agent Run AI Agent
Claude Code, Codex, Gemini, or OpenCode does the work. Cub waits.
cub Verify & Commit
Check git state, commit changes, log to JSONL, mark complete. Loop back.
you Walk away. Cub loops until blocked, budget exhausted, or done.