Build AI SDK Agent Evals

Treat the eval suite as the agent's executable behavioral contract: what it should do, what it should avoid, and the budgets within which it must finish. Inspect, design, implement, run, calibrate, and report; leave runnable artifacts rather than a strategy document.

Scope and hard guards

Work on an existing TypeScript or JavaScript AI SDK agent. Add or change eval, test, fixture, config, and CI files, plus the smallest behavior-preserving production seams needed to inject dependencies. Use an existing agent-plan.md as evidence when present, but verify it against the implementation.

Keep these guards active throughout the run: