Opening images with read_file

Overview

read_file opens a picture instead of reading it as text. When the path ends in a picture extension, the tool attaches the file's pixels to the tool result as an image, so a vision model actually sees the screenshot, diagram, or photo — and the chat card paints the same picture back to the user. Text files keep their existing behavior; .svg stays on the text path because its markup is what a model can use.

Usage

Ask the assistant to look at an image, or call the tool directly:

read_file { "path": "docs/features/images/inspector-panel.png" }

Supported extensions: .png, .jpg, .jpeg, .gif, .webp, .bmp, .ico. .svg is read as text (markup).

startLine / endLine do not apply to a picture and are ignored.

Finding an image

list_files lists pictures alongside text files and marks them, so the model can find one without a shell round trip:

# Listing: <all text and image files>
# Count: 2

# docs/features/images/
  inspector-panel.png (image)
  notes.md

The model-facing result

An image read returns a three-line header — no base64 payload — plus the picture itself, which arrives as a separate vision message part right after the tool result:

# File: docs/features/images/inspector-panel.png
# Kind: image (image/png, 184320 bytes)
# The picture is attached to this tool result as an image part.

The chat card

The tool card shows the path, the MIME type, the size, and the picture. Tap the thumbnail to open it full screen; tap outside the picture, tap ✕, or press Escape to close. A result that reached the UI as plain text (a replayed transcript row, a nested subagent result) has no bytes left, so the card says Image bytes are not part of this result. instead of painting an empty frame.

Behavior

• Caps. An image larger than fileReadMaxImageBytes (default 4 MB) is refused with ETOOL_CAP. The model can downscale with the shell tool and read again. Set the knob at app level:

{
  "fileReadMaxLines": 10000,
  "fileReadMaxImageBytes": 4194304
}

• Vision or text model. The picture is sent as an image part regardless of the model. A text-only model rejects image input, so the provider returns its own error; use a vision-capable model (the model picker's provider catalog notes which OpenRouter models accept image input).

• Authorization is unchanged. read_file is gated by the same File tools authorization mode (off / ask / allowlist / allow) as every other file tool. See tool-authorization.md.

• Path safety is unchanged. Images are subject to the same project-root containment rules as text reads (EOUTSIDE_PROJECT).

• Hidden content does not apply. hide-file-content rules mark lines; a picture has none, so an image read is not redacted.

• search_files is unchanged. It still searches text only — a byte search over pictures would be noise.

• Transcript size. The pixels are persisted with the tool result (the same way an MCP image result is), so a 4 MB picture adds roughly 5.3 MB of base64 to that chat's transcript. Keep the cap low if a project holds large screenshots.

Related

• Native file tools: file-tools.md.

• Tool authorization: tool-authorization.md.

• Hidden line ranges: hide-file-content.md.

• MCP image results (the same content-block path): mcp.md.

• Image preview in the file editor: files-modal-text-and-images.md.