visual-regression · shenyimings - Agent Skill

01 前言 Front Matter

name: visual-regression
description: Review Cloverhollow rendered Scenario Runner captures against intentional visual baselines. Use for pixel-art, UI, camera, scene, or art regression work.
compatibility: Scenario Runner, ImageMagick comparison helpers, image reader/vision

03 SKILL.md 正文

Visual regression

Run the repository's existing ./tools/ci/run-scenario-rendered.sh <scenario_id> and inspect the capture directory, trace.json, and run.log before comparing images. A passing shell exit is not enough: require passed: true, no errors or noop events, expected capture labels, and visual inspection of PNGs.

Use canonical baselines under baselines/visual/<id>. Promotion requires an explicit reviewed command after native image inspection: <scenario> <capture_dir> --reviewed. Missing or extra captures, dimension mismatches, and pixel diffs fail closed. Baseline updates are intentional review decisions, never an automatic repair. Do not approve an image because it was quantized or because a diff script swallowed an error. Fail closed for missing artifacts, missing baselines, or unreadable captures. Never use OS-level window control.

Exact dialogue/UI claims require native attachment inspection, trace/diff review, and OCR corroboration. Use tools/agents/review-capture.ts with the current opencode2 models catalog, check its transcription against the PNG, and fail closed on model unavailability. Do not trust prior approvals, generic model approval, or the disabled built-in vision helper.