OpenHands

I don't love that we have python code in the Canvas repo now: `canvas_ui_tool.py`

For the custom tool calling, couldn't we just pass custom tool names/descriptions into `POST /conversations`, and then wait for the agent to call them over the websocket? we shouldn't really need custom python code

Seems reasonable, <https://github.com/OpenHands/software-agent-sdk/issues/3305> created this issue to track it

I thought the executor is not serializable :thinking_face:

yeah that's my point though--you don't need a custom executor!

Yes... it's weird. Technically required, but it does about nothing here

if the frontend just says

```POST /conversations
{
  "custom_tools": {
     "open_files_ui": "Opens the files UI. Do this every time you edit an HTML file"
  }
}```
then the agent server will call that tool and get a null observation

the frontend will see the tool call come over the websocket and do the right thing

you could call them "remote tools"--the agent server doesn't do anything with them python-wise, it just lets the agent call them and sends em over the websocket

hypothetically we could let observations get pushed back over the websocket but we don't need that for MVP