Thanks! Really glad you noticed this feature. Mixing client-side and server-side tool calls was something we spent a lot of time thinking through.
The current SOTA, e.g. OpenAI’s Responses API or Anthropic’s Computer Use API, basically mandates that server-side tool results return directly, while client-side tool results have to be manually parsed and executed by the user (for obvious security reasons). As a result, it was extremely unclear how a user would be able to chain together tool calls that mixed local and remote tools.
We wanted to close this DX gap, which surprisingly had no real incumbent solution. Users should be able to just define tools and get back clean responses. For power users, we still support manual JSON parsing for full low-level control, but our belief is simple: developers should spend their time building, not doing plumbing work like post-processing tool results.