visual-ai-analyzerlisted

# Visual AI Analyzer An L2 Truth-Execution skill. Where `db_compare_impl` answers "are these two images byte-identical or different dimensions" (structural compare), this skill answers "what visually changed and does it matter" (semantic compare). The two are complementary: the pixel compare catches things the vision model might miss, and the vision model catches things the pixel compare can't describe. **Vision is not deterministic.** Every finding carries a confidence score, and the gate rules are tuned so low- confidence findings get flagged for human review rather than auto-blocking. The worst failure mode of a vision skill is gating on hallucinations; the second-worst is ignoring real regressions because the model was uncertain. ## When You're Invoked - **PIPELINE-5 step 7** — after `uat-executor` has captured screenshots during UAT runs, before the release decision. - **PIPELINE-6 step 6** — pre-release, same position. - **On demand** as `/vibeflow:visual-ai-analyzer <screenshot-path> [--mode <m>]`. - **From `e2e-test-writer`'s generated specs** when a test wants AI-assisted visual assertion instead of a pixel compare. ## Input Contract | Input | Required | Notes | |-------|----------|-------| | Screenshot | yes | PNG / JPEG. A folder path works when it contains `current.png` + optional `baseline.png` + `metadata.json`. | | Baseline screenshot | optional | When present, enables the `baseline-diff` mode. Without it, the skill falls back to `standalone` mode