Hermes Agent Community Reports Screenshot Size Bug That May Cause Claude Desktop Actions to Fail After Resending Oversized Images

On October 6, the Hermes Agent community reported that computer_use on macOS may send an oversized desktop screenshot directly to Claude, then resend the same content after hitting a size limit, eventually interrupting the task. The case used an experimental DirectSDK provider with a Claude subscription. Logs showed a Base64 image of about 13.3 MB, while the endpoint reported a 5 MB limit. All three attempts failed, and no thumbnail recovery message appeared. Community report

This is a core capability of desktop agents. Hermes’s official documentation says its built-in tools connect to cua-driver through MCP standard input and output, supporting clicks, typing, and scrolling. A typical operation first captures the screen, then retrieves the screenshot and indexed interface elements. If the screenshot cannot be sent to the model, subsequent interpretation and actions may stall at the observation stage. Computer Use documentation

After inspecting the code, the reporter said the native vision branch embeds screenshots directly without applying an outbound image byte budget. The error classifier also did not match the size error text from this provider, so thumbnail recovery was not triggered. This is the reporter’s root cause analysis and has yet to be confirmed by maintainers. Issue analysis

The limits should not be generalized to every Claude interface. Current official documentation lists a 10 MB per-image limit for the direct API after Base64 encoding, and 5 MB for Amazon Bedrock and Google Cloud. The 5 MB limit returned by the experimental provider in this case should not be treated as a general rule for the direct API. The documentation also warns that screenshots exceeding the model’s size limit when using native computer use tools will be rejected, and should be resized by the application in advance. Claude vision documentation

For engineers, this case highlights the need to validate image budgets and error recovery together. Based on the mechanisms described above, retrying unchanged oversized content cannot resolve a size limit. Resizing also requires preserving coordinate mapping to avoid a mismatch between what the model sees and the actual click position. Follow-up should track whether a fix covers high-detail screenshots, resending historical images, and provider switching. Current evidence comes from a single macOS development environment, so it does not establish that all production versions or platforms are affected. Case and proposed fixes