ToolResult represents structured and multi-part tool outputs while keeping application-visible data separate from model-facing content.

• output is the original application value. Maps and lists remain available to local code and provider adapters without a text round trip.
• content is the content sent back to the model. It may contain text or ReqLLM.Message.ContentPart values for images and files. When omitted, ReqLLM.Context derives model-facing text from output.
• metadata carries supplementary application or provider-native data. It is not a substitute for model-facing error or result content.

These fields may intentionally differ. For example, an application can keep a rich JSON result in output while returning a concise explanation, an error, or a multimodal file/image payload in content:

%ReqLLM.ToolResult{
  output: %{document_id: "doc_123", page_count: 4},
  content: [
    ReqLLM.Message.ContentPart.text("The requested document is attached."),
    ReqLLM.Message.ContentPart.file_id("file_123")
  ],
  metadata: %{provider_native: %{request_id: "req_123"}}
}

Tool callbacks retain their existing contract: return {:ok, result} for a successful result or {:error, result} for an error. Either result may be a plain text/JSON value or a ToolResult.