What problem does it solve? Long multi-phase AI sessions outlive their context window, and compaction summaries lose critical state: which diff a reviewer approved, what evidence was recorded, and which decisions are settled. This Skill persists that state to files so resumed sessions do not re-litigate settled decisions or treat stale approvals as current. ## Core Features & Use Cases - Four-file state model: Maintains plan.md, decisions.md, evidence.md, and gates.md per task, updated in place at phase boundaries. - Approval ledger with staleness checks: Records HEAD commit and diff summary per reviewer verdict so approvals can be invalidated when the code changes. - Resume protocol: Defines an authoritative re-read procedure after compaction, restarts, or handoffs, with artifacts taking precedence over conversation memory. - Use Case: During a long swarm-mode implementation with reviewer and critic gates, record each APPROVE verdict with its git HEAD; after a context compaction, compare the current HEAD against the ledger and re-run only the gates whose approvals went stale. ## Quick Start Ask the agent to persist the current task's plan, decisions, evidence, and gate verdicts to durable files before the context grows large.
