I took an image and dropped it in and asked if it could see the screenshot. It responded with:

I don't see any screenshot attached to your message. There's no image, file, or paste in what came through to me — only your text question.

Using SaaS Prod and its happening on glm 5.2 but Sonnet was able to see it. GLM 5.2 isn't a multimodal model, so we should ensure that the agent comes back and tells the user that so it's not being perceived as being broken.

@acceptable-analyst-7859 there a couple of ways we could approach this. 1 - just tell the user that the model is not multimodal 2 - check if the user has access to multimodal models and suggests a model to describe the image 3 - proactively picking a multimodal model, describing the image, then falling back to the original model. would love to hear some thoughts on the best approach

3 would be interesting, but maybe is more complex? I tend to lean towards 2.

created an issue, will keep you posted github.com/OpenHands/software-agent-sdk/issues/3892