Tools are called twice on streaming requests for Anthropic API

[youssefwadie]

Bug description

When using the Anthropic API in Spring AI with streaming responses, tools (functions) are called twice. This behavior does not occur in blocking calls (i.e., ChatClient.call), where each function is invoked only once, as expected.

Environment

• Spring Boot version: 3.3.3
• Spring AI version: 1.0.0-M2
• Java version: 17
• OS: Windows 11

Steps to reproduce

1. Register a custom function using the Spring AI Anthropic API for a streaming request.
2. Trigger a request that involves streaming function calling.
3. Observe that the function is invoked twice during the stream, even though it should only be triggered once.

Expected behavior
The function should be called only once during the streaming response, just like the behavior observed in blocking calls.

Minimal Complete Reproducible example
I have replicated Dan Vega's example from his "Spring into AI" tutorial, which you can find here: Dan Vega's Spring into AI example.
But, I replaced the OpenAI client with Anthropic's API. You can find my version here: Spring AI Anthropic Function Calling.

The logs for both streaming and non-streaming requests demonstrate the issue. The tool is invoked twice in the streaming scenario but only once in the non-streaming scenario, as expected.

Non-streaming request

---

Streaming request

[mmazurkevich]

After some time for debugging, I actually found the root cause of the problem.

The problem located in org.springframework.ai.anthropic.api.AnthropicApi#chatCompletionStream(org.springframework.ai.anthropic.api.AnthropicApi.ChatCompletionRequest, org.springframework.util.MultiValueMap<java.lang.String,java.lang.String>) and it constantly reproduced when the tool call is in the end of the message.
This is the problematic row.

.map(event -> this.streamHelper.eventToChatCompletionResponse(event, chatCompletionReference))

For the StreamEvent with type MESSAGE_STOP method returns tool call again

ChatResponse [metadata={ id: msg_01WxuBJqX4mgUymqwEFhxtJU, usage: DefaultUsage{promptTokens=540, completionTokens=274, totalTokens=814}, rateLimit: org.springframework.ai.chat.metadata.EmptyRateLimit@5336b09a }, generations=[Generation[assistantMessage=AssistantMessage [messageType=ASSISTANT, toolCalls=[], textContent=null, metadata={messageType=ASSISTANT}], chatGenerationMetadata=DefaultChatGenerationMetadata[finishReason='tool_use', filters=0, metadata=0]], Generation[assistantMessage=AssistantMessage [messageType=ASSISTANT, toolCalls=[ToolCall[id=toolu_01BRmF2TkLvFEaQCJEVESTMm, type=function, name=getCurrentWeather, arguments={"unit":"C","location":"Paris, France"}]], textContent=, metadata={messageType=ASSISTANT}], chatGenerationMetadata=DefaultChatGenerationMetadata[finishReason='tool_use', filters=0, metadata=0]]]]

which already was send to the stream before (when we handle message MessageDeltaEvent[type=MESSAGE_DELTA, delta=MessageDelta[stopReason=**tool_use**, stopSequence=null], usage=MessageDeltaUsage[outputTokens=274]])

So the problem is that the same ChatResponse sent to the stream. It actually lead to the irrational token consumption.
Also, one thing I want to mention that if tool call in the middle of the message all is ok; only in the end of the message ChatResponse duplicates

[ThomasVitale]

Fixed in #2005