What problem does it solve? Coordinating large coding epics with subagents is expensive and slow when every chunk triggers multiple fresh reviewer agents and oversized models. This Skill runs an experimental nightly orchestration workflow that cuts cost and wall-clock time by using build-only workers on cheaper models, batching reviews every two chunks, and routing fixes to a cheap fixer agent. ## Core Features & Use Cases - Phased state machine: Drives work through Gate, Intake, Explore, Refine, Design, Decompose, Plan, Execute, Review, and Converge phases with explicit transition checkpoints and a resumable markdown store in .orchestration/. - Cost-optimized dispatch: Assigns haiku, sonnet, or opus models per chunk based on complexity and risk tags, batches reviews two chunks at a time, and sends fix-lists to a cheap md-fixer instead of the original builder. - Measured experimentation: Compares nightly cost and wall-clock against the stable orchestrate workflow using bench/orchestration-cost.cjs, reported in every close-out brief. - Use Case: You ask the agent to "orchestrate nightly" on a multi-file feature epic; it decomposes the work into 2-4 chunks, dispatches sonnet builders, reviews them in batches while building continues, and lands the work only after your approval. ## Quick Start Say "orchestrate nightly" or run /orchestrate-nightly and describe the epic you want built so the orchestrator can gate, decompose, and execute it.
