[BUG] Model stuck in error loop after large image upload - clear/fork/rewind all fail

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When a large screenshot is uploaded and triggers "Image was too large. Try resizing the image or using a different approach.", the model enters an unrecoverable loop. Even after explicitly telling the model to ignore the image and move on, it keeps repeating the same error response. The model cannot follow the user's instruction to abandon the topic. clear conversation, fork, rewind, refresh none of the options work.

What Should Happen?

The model should acknowledge the user's instruction to ignore the failed image, drop that topic entirely, and respond to whatever the user asks next. Recovery options (clear conversation, fork, rewind, refresh) should produce a clean state.

Error Messages/Logs

Image was too large. Try resizing the image or using a different approach.

Steps to Reproduce

1. Upload a large screenshot in an active Claude Code conversation
2. Receive error: "Image was too large. Try resizing the image or using a different approach."
3. Tell the model: "ignore that, let's continue with X"
4. Model loops back to the same error response
5. Try each recovery option:

• Say "ignore it / move on" → model acknowledges but loops back
• Page refresh → same loop continues
• Clear conversation → stuck state persists
• Fork conversation → forked session inherits broken context
• Context rewind → model cannot reliably rewind to pre-error state

1. Only a completely new conversation fixes it - losing all prior context

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.81 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Related issue: the model also cannot reliably rewind in general. When told "go back to what you said before" or "undo that", it acknowledges but either repeats the same output or gets confused about its prior state. There is no reliable mechanism to truly discard context once something is in the window.

This is particularly destructive for long working sessions - a single failed image upload can poison the entire conversation with no escape path. All designed recovery mechanisms fail silently, giving the impression of a fix without actually providing one.