Simplify `createChatCompletion`

[kamath]

Right now, you can theoretically use the stagehand.llmClient object to access the underlying LLM for Stagehand.

This can be helpful for times when you want to ask the LLM custom questions that can guide your next step -- for example, if you're writing a Stagehand script, you might want to ask the LLM for what word to input next. This is possible right now, but it's really cumbersome:

It would be awesome to have stagehand.llmClient.generateText() and stagehand.llmClient.generateObject() like the Vercel AI SDK. Bonus points if you add streamText :)

[yash1744]

Yeah @kamath, I think we can wrap our new functionality around the vercel ai's generateText() ,generateObject() ,streamText(), streamObject() methods. can i work on this issue?

[kamath]

yeah have at it!

[Nishant142k2]

can we expose commonly used SDK-style utilities directly on the LLMClient, like so :
stagehand.llmClient.generateText({ ... });
stagehand.llmClient.generateObject({ ... });
stagehand.llmClient.streamText({ ... });

This mimics the Vercel AI SDK pattern and makes developer usage much more intuitive.
These methods are already being imported from ai package so this would just involve exposing them clearly and documenting intending usage.

We could also include a typed wrapper like
structuredAsk<T>(prompt: string, schema: ZodType<T>): Promise<T>
to support structured JSON output via Zod, e.g.

await stagehand.llmClient.structuredAsk("Return the next UI action", z.object({ action: z.string() }));