Optimize AI SDK Agent

Optimize an existing TypeScript or JavaScript AI SDK agent through one evidence loop:

baseline -> protect behavior -> audit -> prioritize -> implement -> measure -> keep or revert

Change the repository by default. Return an audit without edits only when the user explicitly requests review-only or no-code work. Connect every retained change to a repository-grounded problem, a hypothesis, a comparable before/after result, and a regression guard.

Scope and hard guards