The Problem
Coding agents like pi can drift into unstructured, undisciplined workflows—skipping planning, reviewing their own work superficially, and producing inconsistent results. Teams need a way to enforce engineering rigor (planning, multi-perspective review, verification) without manually orchestrating every step.
What This Does
roach-pi is a set of TypeScript extensions for the pi coding agent that adds a structured engineering workflow layer. The core is extensions/agentic-harness/, which contains the orchestration logic: index.ts is the entry point, agents.ts defines the subagent roles, and discipline.ts enforces workflow phases. The extension provides commands like /plan, /ultraplan, /review, and /ultrareview that delegate to specialized reviewer agents defined in extensions/agentic-harness/agents/ (e.g., reviewer-architecture.md, reviewer-security.md).
The repo also includes extensions/autonomous-dev/, which adds a higher-level orchestrator (orchestrator.ts) for autonomous development runs, and extensions/fff-search/ for filesystem-based fuzzy search. Documentation is extensive—74 files under docs/engineering-discipline/ covering plans, milestone checkpoints, and post-implementation reviews.
How To Use It
Setup — This is an extension for the pi coding agent, not a standalone tool. Install with:
pi install git:github.com/tmdgusya/roach-pi
Then run /setup first. It modifies ~/.pi/agent/settings.json to set quietStartup: true. The README warns that a superpowers skill conflicts with bundled skills and must be removed.
Configuration — No environment variables are required. The extension is configured via the /setup command and per-command options. Each extension subdirectory (agentic-harness, autonomous-dev, fff-search) has its own package.json and tsconfig.json; the code is TypeScript and presumably compiled via npm, though no root-level build script or Makefile is present.
Running it — Invoke commands inside a pi session: /clarify, /plan, /ultraplan, /review [target], /ultrareview [target]. The README documents all commands. Review outputs are saved to docs/engineering-discipline/reviews/.
Real-World Use
A developer working on a non-trivial feature would run /plan to get a structured implementation plan, then /ultraplan to have five reviewer perspectives (feasibility, architecture, risk, dependency, user value) run in parallel and synthesize a milestone DAG. After implementation, /ultrareview dispatches 10 subagents (5 reviewer roles × 2 seeds) to catch bugs and design flaws before the PR is opened. This is a local, open-source alternative to cloud-based review pipelines like claude-code's bughunter.
Code Health & Issues
Med — No LICENSE file. Usage and redistribution rights are unclear. This is a blocker for commercial adoption. Med — No root-level build or CI for the extensions themselves. The GitHub Actions workflows (pages.yml, release.yml) handle docs pages and releases, but there's no test runner or linting configured at the repo root. Tests exist (39 files under extensions/agentic-harness/tests/), but the CI doesn't appear to run them. Low — Single-maintainer risk. Forked from tmdgusya/roach-pi, with 0 stars on this fork. The extensive documentation suggests active development, but the project's trajectory depends on one person. Low — Skill conflicts. The README documents a conflict with the superpowers skill. Duplicate skill names could cause nondeterministic behavior.
The Bottom Line
roach-pi is a serious, well-documented attempt to bring engineering discipline to AI-assisted coding. The multi-agent review pipeline and the emphasis on observable, transparent behavior are genuinely useful. It's worth evaluating for teams already using pi, but the missing license and lack of automated test execution in CI are real concerns for production adoption.