The cursor cloud agent is unable to access images uploaded via the ‘Add file’ icon.
To quote the agent directly:
“Thanks for the clarification. I still don’t have programmatic access to chat-uploaded files in this runtime unless they’re materialized as filesystem paths, and they aren’t showing up in the accessible directories.”
Steps to Reproduce
Upload an image in a chat message inside a cloud agent. Ask the agent what it can see
Expected Behavior
It should be able to see the image. In our case these are screenshots of bugs, which are obviously important when fixing issues.
I just tested this and I’m able to share images with Cloud Agents and have them described correctly (see attached screenshot). My test was started directly from cursor.com/agents.
How are you starting your agents? Through Slack, GitHub, or cursor.com/agents? If you try starting one from cursor.com/agents and attaching the same images, is it able to see them?
I can’t reproduce this. Just in case it makes a difference, could you try with GPT 5.4 Codex? Can you also confirm that the images you’re sending are .png / .jpg, and not another file type?
Thanks for the detailed repro with both bcIDs – that was really helpful.
I dug into the logs for both runs, and found something interesting: the failing agent (bc-2a45d8f0) isn’t actually routing to GPT 5.3 Code, but being routed to a default model (auto) that doesn’t support vision.
Could you share how you’re selecting the model for these Cloud Agent runs? Any specific syntax you’re using in Slack?
To be clear, your billing looks correct and you’re being charged appropriately for the model that actually serves the request, not for a more expensive model you didn’t receive.
Please try running @Cursor settings in the Slack channel where you’re launching Cloud Agents, click Change Settings, and try toggling to a different model and then back to 5.3 Codex.
