Skill evaluation (CodeCraft)

Description

Test and promote skills in the CodeCraft LLC agent-skills library — writing realistic eval prompts per skill (task evals and trigger evals with near-miss negatives), objective assertions, running each prompt with and without the skill in fresh agent sessions, grading outputs, recording results in the standard evals/ format, comparing against baseline, human review, deciding when a skill moves from drafted to tested in TODO.md, and keeping live-model costs approved and bounded. Use whenever the user asks to test, evaluate, benchmark, or validate a CodeCraft skill, promote skills from drafted to done, check which skills have evals, or tune a skill's description for triggering.