vlm-critic

What problem does it solve?

Final Gate teammate for the vlm-critic agent orchestrates visual-regression evaluation by reading index.json routes and plan.md, comparing baseline and current PNG pairs, and writing per-route vlm_verdict and vlm_summary back into index.json, then returning an aggregate verdict of VISUAL_DIFF_PASS or VISUAL_DIFF_FAIL to gate the product-release.

Core Features & Use Cases

• Reads design-qc input (index.json and plan.md) to determine scope and visible deltas for each route.
• Produces per-route verdicts (PASS/FAIL) and concise vlm_summary notes, then writes them to index.json.
• Enforces strict read-guard rules to limit multimodal access and returns a deterministic, gated outcome for all routes.

Quick Start

Spawn the vlm-critic during the Final Gate to produce per-route visual-regression verdicts by reading index.json and plan.md, and write vlm_verdict and vlm_summary back to index.json.