Anthropic interleaved/extended thinking isn't captured properly by the default Anthropic instrumentation

[ldorigo]

Description

(C.f. discussion on Logfire's slack: https://pydanticlogfire.slack.com/archives/C06EDRBSAH3/p1771351911189739 )

The code looks like:

# Build request kwargs with thinking enabled
            kwargs: dict[str, Any] = {
                "model": self.model_name_map[self.model_params.model_version],
                "messages": anthropic_messages,
                "temperature": 1,  # Only temperature of 1 is supported for thinking
                "max_tokens": self._resolve_max_tokens(default_max_tokens=16384),
                "stream": True,
                "tools": anthropic_tools,
            }

            if system_message:
                kwargs["system"] = system_message

            # Add thinking parameter if budget specified
            if thinking_budget is not None:
                kwargs["thinking"] = {"type": "enabled", "budget_tokens": thinking_budget}

            # Call Anthropic API with beta features for interleaved thinking
            logger.debug(f"Starting agentic loop iteration {iteration_num}")
            stream = await self.client.beta.messages.create(
                betas=["interleaved-thinking-2025-05-14"], **kwargs
            )

Outputs aren't shown and instead the 'tool results' (user message sent back with tool results) are shown as outputs.

Python, Logfire & OS Versions, related packages (not required)