Spring AI Anthropic Client fails to parse "thinking" content block type from Claude 3.7 response

[dinodanger]

trafficstars

Description

The application is failing to process responses from Claude 3.7 Sonnet when using the thinking parameter. We're encountering a JSON parsing exception because Spring AI's Anthropic client doesn't recognize the thinking content block type in the streaming response.

Error Details

java.lang.RuntimeException: Failed to json: {"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":"","signature":""}}

Root cause:

com.fasterxml.jackson.databind.exc.InvalidTypeIdException: Could not resolve type id 'thinking' as a subtype of `org.springframework.ai.anthropic.api.AnthropicApi$ContentBlockStartEvent$ContentBlockBody`: known type ids = [text, tool_use]

My Code

    val optionsBuilder = AnthropicChatOptions.builder().model(model).temperature(temperature)
        if (thinking) {
            optionsBuilder.thinking(AnthropicApi.ThinkingType.ENABLED, 16000)
            optionsBuilder.maxTokens(24192)
            optionsBuilder.temperature(1.0)
        }

        val prompt = Prompt(messages, optionsBuilder.build())

        return this.chatClient.prompt(prompt).stream().content().collectList().awaitSingle()
            .joinToString(separator = "")

Where client is

private val chatClient: ChatClient = ChatClient.create(chatModel)

Steps to Reproduce

1. Configure Spring AI with Anthropic client using Claude 3.7 model
2. Enable the "thinking" parameter in AnthropicChatOptions using optionsBuilder.thinking(AnthropicApi.ThinkingType.ENABLED, 16000)
3. Send a request to the model

Expected Behavior

The Spring AI Anthropic client should properly parse and handle the "thinking" content block type in the streaming response.

Environment

• Claude model: claude-3-7-sonnet-20250219
• Java version: 21

Possible Solution

Update the type registry in Spring AI's Anthropic client to include the "thinking" content block type as a recognized subtype of ContentBlockBody. This likely requires adding a class to handle this block type and updating the Jackson deserializer configuration.