vision

What problem does it solve? Text-only orchestrator models like GLM 5.2 or DeepSeek cannot see screenshots, dropped images, or visual tool results, so they cannot answer layout, color, alignment, or visual-verification questions. This Skill routes those visual tasks to a dynamically registered vision-capable subagent and returns structured JSON findings the orchestrator can relay. ## Core Features & Use Cases - Visual Intent Detection: Recognizes explicit visual language, image attachments in user messages, screenshot tool results (chrome-devtools, Playwright, cua-driver), and gaps between text output and visual criteria. - Dynamic Model Discovery: Runs a bundled script that lists image-capable models from your configured OpenCode providers, lets the user pick one, and persists the choice across sessions. - Task-Specific JSON Contracts: Designs a minimal response template per visual task, delegates to a vision-* subagent, and parses/validates the returned JSON with retry on malformed output. - Use Case: After a browser tool returns /tmp/dashboard.png and the user asks "does the dashboard look right", the Skill delegates the screenshot to a vision subagent with a layout-focused JSON template and reports evidence-backed findings. ## Quick Start Ask the agent to visually verify that the page in the latest screenshot matches the design, and it will route the image to a vision subagent.