Vision Sandbox

**Judgement:** A strong, specialized vision skill for agentic, code-verified image analysis—particularly valuable for UI auditing, spatial grounding, and visual math—provided you are comfortable depending on Gemini’s API and sandbox. **What it does well:** - Uses Gemini’s native Python sandbox to *verify* visual inferences (coordinates, bounding boxes, counts) instead of relying purely on the model’s internal reasoning. - Comes with clear pattern prompts for spatial grounding, counting/logic, visual math, and layout checking, making it immediately useful without heavy prompt engineering. - Produces structured data (e.g., coordinates in a fixed scale, UI metadata JSON) that can be piped into coding agents like OpenCode for automated front-end fixes or regression checks. **Key risks / limitations:** - Hard dependency on `GEMINI_API_KEY` and Google-hosted sandbox; outages, rate limits, or model regressions will directly impact reliability. - Latency and cost characteristics are tied to `gemini-3-flash-preview`, which may change over time and may not be ideal for very high-throughput workloads. - Handling of screenshots and UI images may involve sensitive data; safe use requires your own privacy/access controls and data-handling policies. **Recommended scenarios:** - Automated UI test pipelines that need robust element localization and overlap detection from screenshots. - Agentic coding workflows where a coding agent (e.g., OpenCode) must react to visual UI state—fixing CSS/HTML based on extracted visual metadata. - Visual counting and simple visual math tasks where code-verified results (rather than heuristic model guesses) are important. - Auditing or monitoring of front-end layouts across versions or environments using image snapshots plus programmatic analysis.