eval-driven-dev

Enforces a 7-stage development workflow: Discover (Socratic questioning) → Explore (codebase scan) → Spec (acceptance criteria) → Eval (write tests FIRST) → Implement → Verify → Iterate. Supports deterministic tests (Playwright, unit), fuzzy evals (screenshot + LLM-as-judge), integration evals, QA dogfood, and regression suites. Strong nudge enforcement — agents are pressured to follow each stage but can override with justification. Trigger on "eval driven," "test first," "write tests first," "TDD," "eval," "acceptance criteria," "spec first," "workflow," "development process," "how should I build this," or any new feature request. Hooks into my-stack for skill routing and dogfood-complete for QA/test generation.