What problem does it solve? Applying the same refactor across dozens of files sequentially is slow and error-prone, while pure codemods cannot handle natural-language references in comments, docstrings, and log messages. This skill coordinates a swarm of cheap, fast per-file agents so each file is refactored independently with light judgment, in two controlled waves. ## Core Features & Use Cases - Two-wave parallel execution: Launch one agent per file on the harness's cheapest fast model tier, source files first, then test files, with a checkpoint between waves. - Structured agent prompt template: Each agent receives a self-contained brief with explicit boundary constraints (what not to touch) and authorized judgment (what it may decide per file). - Verification protocol: Grep for leftover identifiers after literal renames, and run the test suite, type checker, and linter for judgment-involved refactors. - Use Case: Rename a parameter across 20 source and test files, including docstring and log-message references, by dispatching one agent per file and verifying no stale occurrences remain. ## Quick Start Ask the agent to apply the same refactor across all matching files using the refactor-swarm pattern, with one cheap-model agent per file in source and test waves.
