three-man-teamlisted

## When to Use - **Tasks flagged CRITICAL** by complexity-classifier - **Estimated >2 hours** of work - **Multi-phase features** (3+ distinct phases: design, build, test) - **Large PRs** touching >5 files - **Cross-module changes** requiring architectural review before implementation ## When NOT to Use - Simple bug fixes (<30 min) - Single-file changes - Refactoring with clear, linear steps - Anything already well-scoped in an existing plan --- ## How It Works Three agents run in parallel, each with a specific role: ### Agent 1: Architect (model: opus) - **Input**: task description, codebase context, current branch state - **Job**: - Read the relevant codebase modules and map module interdependencies; identify breaking-change risks and flag hidden coupling - Identify phases, boundaries, and architectural decisions - Write a structured plan to `~/.claude/plans/<task>.md` - **Output**: `~/.claude/plans/<task>.md` with: - Objective & scope - Phase breakdown (Phase 1, Phase 2, etc.) - Acceptance criteria per phase - Implementation boundaries (what Builder should NOT touch) - Known constraints & gotchas - Links to relevant code/docs ### Agent 2: Builder (model: sonnet) - **Input**: Architect's plan, task description - **Job**: - Implement based on the plan - Write code, tests, commit messages - Stage all changes in a worktree or branch - Prepare PR description - **Output**: PR-ready code (or staged changes + PR branch) ### Agent 3: Reviewer (mod