skill-benchmark

# Skill benchmark ## Contract | Field | Bound contract | |---|---| | Trigger | The user runs /skill-benchmark | | Authority | Human-only. Preview the benchmark target, judge or candidate models, rubric or task set, and estimated spend before any LLM call. No skill, code, credential, or remote mutation. | | Side effect | Writes benchmark artifacts under .gstack/benchmark-reports/ and incurs LLM inference spend. | | Done | A scored skill-quality report or model-comparison table is written and returned to the human. | ## Inputs ### Skill-quality mode - `--baseline`: capture a scored baseline before changes. Run first on a clean branch. - `--quick`: single-pass scoring without baseline comparison. - `--skills <name1>,<name2>`: score only named skills. Omit to auto-discover from the skill directory. - `--diff`: score only skills whose files changed on the current branch. - `--trend`: show score trends from historical baseline files. - Judge model and rubric must be supplied or confirmed by the user before scoring begins. ### Model-comparison mode - The task or task set to run against every candidate model (required). - The candidate model list (required): two or more models to compare. - Per-model run count or spend budget cap (optional; defaults to one run per model per task). - Output path for the comparison table (optional; defaults to a local artifact under .gstack/benchmark-reports/). ## Procedure 1. Determine the benchmark target from the request. If the user names ...