What problem does it solve? Multi-agent coding workflows often rely on subagent self-reports of success, leading to unverified code, missed spec requirements, and silent failures. This Skill replaces linear implement-then-review flows with a rigorous 4-phase execution loop where the orchestrator independently validates every work unit and a fresh adversarial reviewer verifies compliance against the written spec contract. ## Core Features & Use Cases - 4-Phase Execution Loop: Every work unit passes through IMPLEMENT, VALIDATE, ADVERSARIAL REVIEW, and COMMIT phases with blocking quality gates (tests, coverage thresholds, type checking, lint, file scope verification) and a max-3-retry escalation protocol. - Adversarial Review with Fresh Reviewers: A separate review subagent checks each Definition of Done item with file:line evidence, returning binary PASS/FAIL verdicts; re-reviews always use a fresh reviewer instance to prevent anchoring bias. - Work Unit Decomposition & Parallel Execution: Plans are decomposed into a DAG of work units with explicit dependencies, file scopes, and human checkpoints, with fan-out/fan-in patterns for parallel execution and sequential commits. - Context Persistence & Recovery: Plans, project context, and execution state are persisted to docs/ so orchestration can resume after context compaction or session interruption. - Use Case: When implementing a multi-feature spec across several subagents, use this Skill to decompose the spec into work units, run each through the gated loop, pause at human checkpoints, and produce a final comprehensive review before PR creation. ## Quick Start Ask the AI to execute the approved implementation plan using the orchestrated execution 4-phase loop with adversarial review for each work unit.
