You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Tool results have no image channel, so the model cannot see screenshots it captures #824
tools.Result (internal/tools/types.go:92-98) carries Output string and no image or content-item field, so no tool can hand the model an image. There is also no view_image tool: grep -rni view_image internal/ returns nothing.
The sharp edge is our own screenshot tooling. internal/tools/local_capture.go implements browser_screenshot, desktop_screenshot and browser_pdf. They write a PNG to disk, set Meta["artifact_path"], and return the sentence "Artifact captured: ". Nothing reads artifact_path afterwards. So the model asks for a screenshot, is told one exists, and cannot look at it. From the model's side the tool may as well have done nothing.
Worth being clear that Zero is not image-blind. The input pipeline is complete and reasonably mature: zeroruntime.ImageBlock, Message.Images, internal/imageinput for pasted and attached images, media-type normalisation, and ACP image blocks. A user can hand the model an image today. What is missing is the model-initiated leg: a tool result cannot contain one.
Two pieces:
An image channel on tools.Result that the agent loop turns into Message.Images on the tool-result message.
A view_image tool so the model can open a path it has been given, which also covers files the user mentions but does not attach.
Once (1) exists, the screenshot tools become useful with a one-line change each, and MCP image blocks have somewhere to go.
Activity
feat(tools): give tool results an image channel and add view_image (T…