vision input fails

Summary

Vision input currently works only through interactive TUI clipboard-paste and drag-drop (detectDroppedImageFile / detectClipboardImage → Jimp → base64 → {type:"image"}).
Environment

• cmd v0.38.2
• Installed via npm i -g command-code (mise)
• macOS (darwin)
• Model: Claude Sonnet 4 (claude-4-sonnet-20250514) — confirmed inputModalities: ["text","image"]
  Reproduce

1. Put any JPEG or PNG in the working directory

2. Run:

cmd -p "Describe what you see in test.jpg"
Result: Model responds that it cannot visually inspect the image. Same result for PNG, GIF, WEBP.
Root cause (from inspecting dist/index.mjs)

1. read_file returns text-only output for all binary files
   src/tools/read-file/read-file.ts → formatOutput:
   function formatOutput(e) {
   if ("binary" === e.contentType)
   return [..., Binary file: ${e.fileType}].join("\n");
   // Model only ever sees: "Binary file: image/jpeg"
   ...
   }
   No {type: "image"} content block is ever created from a file read. The raw bytes are read into a Buffer but discarded.
2. @filepath references silently skip binary files
   src/utils/file-references.ts → processFileReferences:
   if ("binary" === i.contentType) continue; // images vanish entirely
3. AVIF and other modern formats fall through to the text path
   getFileType uses a hardcoded extension map (Ib) with no .avif entry:
   Ib = { ".png":"image/png", ".jpg":"image/jpeg", ".jpeg":"image/jpeg",
   ".gif":"image/gif", ".webp":"image/webp", ".svg":"image/svg+xml",
   ".bmp":"image/bmp", ".ico":"image/x-icon", ... }
   // No .avif, .heic, .heif, .tiff
   Since getFileType(".avif") returns "unknown", and isBinaryFile("unknown") returns false (it only matches known MIME values), AVIF files are read as UTF-8 text — producing garbled output rather than being flagged as binary.
4. Vision content blocks only exist in the TUI paste/drag-drop path
   src/utils/clipboard-image.ts → detectDroppedImageFile → Jimp resize → base64 → {type:"image"}. This path is a React/Ink interactive handler and is not reachable in headless mode.

Expected Behavior

..

Actual Behavior

..

Steps to reproduce the issue

..

Command Code Version

latest

Operating System

macOS

Terminal/IDE

No response

Shell

No response

Additional context

No response

Start with src/tools/read-file/read-file.ts and src/utils/file-references.ts, then compare their binary handling with src/utils/clipboard-image.ts. Reproduce with cmd -p against a local image and trace how the file reaches the model; done means supported image files, including the formats named in the issue, produce vision content in headless mode instead of text-only or being skipped.