What problem does it solve? Plans written by a single AI session often carry unchallenged assumptions and blind spots. This Skill convenes an adversarial review council over a spec-kit feature's plan.md so design flaws are caught before any code is written or tokens are spent on implementation. ## Core Features & Use Cases - Multi-agent plan defense: Dispatches a configurable bench of 1-8 Sonnet reviewer agents (each with a distinct lens like correctness, risk, simplicity, testability, sequencing) to produce independent opinions in parallel, followed by anonymized peer review and an Opus chairman synthesis. - Tiered ceremony and cost control: Supports full and standard council tiers via profile.yaml, scaling session counts, context loading (eager vs lazy), output caps, and graph-query ceilings to balance review depth against token cost. - Resumable, auditable rounds: Writes classified suggestions to council/round-N/suggestions.md with stable finding IDs, per-session traces, graphify grounding checks, and support for --reopen delta or full to re-review revised plans. - Use Case: After running /speckit-plan on a new feature, invoke /speckit-council to have five independent reviewers attack the plan, then feed the resulting suggestions.md into /speckit-council-triage before writing tasks. ## Quick Start Run /speckit-council in a spec-kit repository with a completed plan.md to convene the plan-defense council and generate a classified suggestions report for the current feature.
