autonomous-refactor

Test-driven autonomous refactoring against project design principles. Captures a green test baseline, audits code against README.md principles, applies changes in git worktree isolation (reverts on test failure), and produces a before/after report — entirely without human input in the loop.

Read-only analysis agent for autonomous-refactor. Reads target source files and the project README.md, extracts stated design principles, identifies concrete violations or improvement opportunities, and returns a ranked JSON list with a numeric alignment score. Called in Phase 2 (initial audit) and after each successful Phase 3 change (re-audit).

Write-capable Phase 3 agent for autonomous-refactor. Receives a single refactoring opportunity and a git worktree path, reads the target files inside that worktree, implements the minimal change needed to address the opportunity, and returns a structured summary. Does NOT run tests — the orchestrator runs them after this agent returns.

Read-only Phase 4 agent for autonomous-refactor. Reads the session state, baseline and final metrics files, and populates the final-report.md template with before/after comparisons. Optionally runs the metrics.ts diff utility for per-file diff summaries. Returns the fully populated report markdown to the orchestrator.

Phase 1 agent for autonomous-refactor. Reads target source files, generates a comprehensive behavioural test suite covering all exported symbols, runs the tests to confirm a green baseline, and returns structured results. Called once per refactor session before any code changes are made.