Refactors fail on hidden edges, not syntax
A model that aces a 200-line kata can still miss a cross-package contract. Opus 5’s lead on the agent board is the reason it belongs in the planning pass: it holds more of the call graph in working memory. You still need tests before merge.
The two-model pattern
Pay Opus 5 once for the design and file list. Switch to Sonnet 5 or GPT-5.6 Sol for the mechanical edits. Cursor Composer and Claude Code both keep a shadow of the repo so diffs stay reviewable. OpenHands is the open-source alternative if you want the loop in Docker.