Look At - Multimodal File Analysis

Multi-backend vision router for images, PDFs, video, diagrams and other media. Defaults to agy -p on gemini-3.7-flash-high — Gemini via Antigravity OAuth, unmetered — which reads images, PDFs and video natively. Audio auto-routes to the metered api backend, the only one that handles it. Three further unmetered CLI backends (claude-code -p, codex exec, Copilot on GPT-5.4) give independent second opinions.

Tool Selection Enforcement

Tool Routing Facts

• Read on a media file loads the full content into context regardless of how briefly you look at it — a "quick glance" costs the same thousands of tokens as a full read. Content type, not file size, determines the tool.
• Read on a PDF extracts raw text and loses table structure and visual information; look_at returns it as structured data.
• The point is context economy, not vision capability. Read pulls the whole image into this session's context; look_at spends a subprocess's context instead and returns text. Better vision models do not change that arithmetic — they make the cheap backends sufficient.
• Backend extraction is accurate for most use cases — start with look_at, escalate to Read only if the extraction is insufficient. Defaulting to Read "for exact text" wastes the context this skill exists to save.
• The claude backend spawns a child claude-code -p. look_at.sh sets LOOK_AT_NESTED=1 so image-read-guard.ts stands down inside that child — without it the guard denies the child's Read and points it back at look_at.sh, which spawns another child. That is unbounded recursion, not a slow call.

Red Flags

• Passing an image, PDF, or screenshot path to Read → use look_at.
• A text-based PDF with structure/tables/charts → still look_at, not Read.