atmos-long-task-impllisted

# Atmos Long-Task Implementation Parent session = **lead agent**. Feature code is written only by **impl subagents**; **review subagents** are read-only; spec ambiguity is escalated only to **HUMAN**. Hard boundaries are in [roles.md](roles.md). This skill **replaces the parent writing code**, not the coding rules in `atmos-specs-impl`. Each impl subagent still follows that skill (layer order, WebSocket-first, regression gate). ## When to use Use this skill if **any** of these hold: - The user wants parallel work, wave splitting, kanban tracking, subagent dispatch, or long-task implementation - One TECH phase spans multiple crates or apps and a single session would lose invariants from the first half - Two workflows would edit the same files if started at once ## When not to use - A small change fits in one `atmos-specs-impl` session with regression green - PRD/TECH/TEST are still templates or TECH does not admit a single architecture (stop, ask HUMAN, or return to `atmos-specs-tech`) - The user only wants a few files edited in the parent session Scenario testing still belongs to `atmos-specs-test-run`. Full spec review belongs to `atmos-specs-review`. This skill's review subagent is a **slice-level, checklist-driven gate**, not a replacement for those skills. ## Architecture (optimal cut vs “everyone parallel”) Four roles are enough. Do not add planner / integrator / tester. ``` HUMAN ←—— sole product / contract / ambiguity arbiter ▲ Lead agent wave split · co