sdlc-agent-evalslisted

# Test the configuration, not just the code `CLAUDE.md`, skills, hooks and subagent definitions steer the agent as surely as source code steers the program. They're edited by hand, they interact, and their effects are invisible until something goes wrong in a diff nobody expected. That makes them exactly the kind of thing regression tests exist for. A suite answers one question: **after this change, does the agent still do the work to the standard it did before?** ## What you need first Something worth measuring — a real `CLAUDE.md` (`sdlc-claude-md`) and a working verification story (`sdlc-feedback-loop`), because an eval's checks are that same verification run headlessly. Plus CI that can run the agent non-interactively, and budget for the runs. ## Assembling the suite **Take 20–50 tasks from work you actually did**, each with the outcome you accepted at the time. Real tasks. Invented ones test your imagination, and they're systematically easier than reality because you wrote them knowing the answer. **Each eval is a prompt plus its checks** — tests green, types clean, behavior unchanged, the relevant rule followed. The checks have to be mechanical; a human grading 40 outputs per config change will not keep doing it. **Every incident earns an eval**, written by whoever owned the incident, and it stays in the suite permanently. This is how the suite gets good: it accumulates the specific ways this codebase goes wrong. **Gate config changes on the result.** A skill e