Read tool enters infinite retry loop when .jpg file contains HTML (400 'Could not process image')

Bug Description

When a file has a .jpg extension but contains HTML content, the Read tool attempts to process it as an image, the API returns 400 invalid_request_error: Could not process image, and Claude Code enters an infinite retry loop that the user cannot escape from within the same session.

Minimal Reproducible Example

No external services needed — just create a fake image locally:

echo '<!doctype html><html><head><title>Login</title></head><body><h1>Please log in</h1></body></html>' > /tmp/fake_image.jpg

Then ask Claude Code to read it:

"Read the file /tmp/fake_image.jpg"

Result: Claude sends it to the image API → 400 Could not process image → infinite retry loop.

How This Happens in Practice

Any URL that returns HTML instead of an image (auth redirects, login walls, CDN errors) when downloaded with curl and saved with a .jpg extension will trigger this. Example:

# URL requires auth → server returns HTML login page → saved as .jpg
curl -sL "https://some-service.com/protected/image.jpg" -o /tmp/img.jpg
file /tmp/img.jpg  # → HTML document, ASCII text

Expected Behavior

1. Read tool should detect actual file content type (magic bytes, not extension) before sending to image API
2. If API returns 400 Could not process image, Claude Code should surface the error once and stop — no retries

Actual Behavior

• API Error: 400 {"type":"invalid_request_error","message":"Could not process image"} repeated multiple times
• Claude keeps retrying with the same broken file
• Session gets stuck — user has to abandon the conversation

Environment

• Platform: Linux (Ubuntu), zsh
• Claude Code version: claude/main-NiROL