I'm using the OpenAI `computer-use-preview` agent, and often get this error: `Failed to execute task: 400 Invalid type for 'input[0].output': expected a computer call output, but got a string instead.` Initialized this way: `agent = stagehand.agent(model="computer-use-preview", ...)` It's great when it works! But seems there's some kind of invalid model output that presumably Stagehand could heal better under the hood.