feat: support images in mcp tool responses

[bcheung]

Vision-capable models can now consume images from mcp tool responses.

https://ai-sdk.dev/docs/ai-sdk-core/tools-and-tool-calling#multi-modal-tool-results

This resolves the merge conflict in the original PR: #1589

Closes: #1563

[bcheung]

One thing to point out: this only supports images in mcp tools, not tools from the registry.
To support tool responses from the registry, need to parse the attachments from the result here in toModelOutput: https://github.com/sst/opencode/blob/c450a5707ce8446e3ce9f2f3fa4099bf426f8b38/packages/opencode/src/session/prompt.ts#L531-L595

However, I realize that extracting base64 data from the url might get a bit messy and don't want to block this PR bc of that so leaving it out.

[rekram1-node]

gonna merge today sorry for delays

[rekram1-node]

/oc revert the toModelOutput changes but keep the rest as is. If you don't know what im saying read the previous commits.

DO NOT commit your changes

[opencode-agent]

Perfect! I've reverted the toModelOutput changes back to the original simple implementation that just returns the text output, while keeping all the other changes from the commit (including the execute function modifications that support image attachments).

opencode session  |  github run

[rekram1-node]

Im gonna merge it like this, I mentioned on discord I dont think the toModelOutput change was needed? And tbh I think we can delete that block altogether but maybe im missing something

[bcheung]

remember to remove this when removing toModelOutput

[bcheung]

Im gonna merge it like this, I mentioned on discord I dont think the toModelOutput change was needed? And tbh I think we can delete that block altogether but maybe im missing something

took a quick skim, looks like toModelOutput isn't used as it was replaced with MessageV2.toModelMessage()
I'm not entirely sure though, but you know more about the internals of opencode

just left a note to not forget to remove this as it was directly used for toModelOutput when I thought this was the function that sent messages to the model
#4100 (comment)