ROACH PI – An Open-Source Extension That Puts Engineering Discipline into AI Coding Agents

To solve the reliability problems that arise when AI coding agents work without a plan or fail to verify the code they wrote themselves, it applies an engineering lifecycle through multi-agent orchestration.

It goes through the stages of Clarify → Plan → Execute → Validate → Cleanup, and in particular separates the execution agent (Worker) and the validation agent (Validator) into separate processes to pursue quality assurance through information isolation.

Key features are as follows:

• 12 specialized agents: Includes exploration, execution, planning, validation, and 5 independent reviewers (architecture, risk, user value, etc.).
• Ultraplan: Decomposes complex tasks into a dependency DAG in milestone units.
• Long Run: Repeats large-scale tasks based on checkpoints to address context loss.
• Behavioral guardrails: Automatically injects engineering principles into agents, such as 'no writing before reading' and 'no optimization without measurement.'
• AI Slop Cleaner: Cleans up LLM-specific code smells over 6 stages after execution.

All prompts, agent definitions, and skills are publicly available in markdown format, providing complete transparency, allowing all behavior to be checked without hidden system prompts.