Реализация технических планов с оркестрацией агентов

orchestration

implement_plan is a Claude Code skill that executes approved technical plans stored in thoughts/shared/plans/, supporting two modes: direct implementation for plans with three or fewer tasks, and agent orchestration for larger plans where each agent runs with a fresh context and writes handoff files to survive context compaction. Before any work begins, the skill runs a deep pre-mortem via /premortem deep, checking for technical, integration, process, and testing risks — HIGH-severity findings block execution until the user explicitly accepts them or adds mitigations. Progress is tracked by checking off items directly in the plan file, and after each phase the skill pauses to report automated verification results and request manual testing from the user. When the actual codebase diverges from the plan, execution stops with a structured mismatch report rather than guessing. It is designed for teams running multi-phase implementations in long Claude Code sessions where context preservation and human oversight matter.