What problem does it solve? Executing a multi-lane phase plan by hand means manually dispatching agents, tracking lane dependencies, verifying gates, and merging work without clobbering parallel changes. This Skill automates that entire loop: it parses a plan doc produced by claude-plan-phase, dispatches worktree-isolated Claude teammates per lane in DAG order, verifies gates, and auto-merges on green with retry-once failure handling. ## Core Features & Use Cases - DAG-ordered lane dispatch: Parses the plan's Lane Index, topologically sorts lanes, and dispatches ready lanes in parallel waves capped by EXECUTE_MAX_PARALLEL_LANES, each in its own git worktree. - Model-tier routing and retry escalation: Assigns fast/strong/frontier Claude models per lane complexity and escalates tiers on retry, halting all lanes after a second failure. - Merge safety audits: Runs pre-merge destructiveness checks, cross-lane file-touch audits, parent-tree leakage detection, and post-merge import smoke tests before any merge lands. - Resumable state machine: Persists lane and gate state to a JSON state file so interrupted phases resume with --resume. - Use Case: After running /claude-plan-phase to produce plans/phase-plan-v2-p1.md, invoke this Skill to execute all lanes full-auto, merging each into the target branch as its verification gates go green. ## Quick Start Run /claude-execute-phase with the path to your plan doc (or let it auto-detect plans/phase-plan-<VERSION>-<PHASE_ALIAS>.md) on a clean working tree to dispatch and merge all lanes automatically.
