AnthropicApi does not invoke nullary function tools when using streaming

[asw12]

Bug description
When using an Anthropic ChatModel with a streaming chat response (StreamResponseSpec#chatResponse into Flux<ChatResponse>), tools that take in zero parameters are not invoked.

The JSON content returned by Anthropic's text/event-stream response looks like:

{"type":"message_start","message":{"id":"msg_01CNBhjoXJNRjoDHX5jLhWa5","type":"message","role":"assistant","model":"claude-3-7-sonnet-20250219","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":3154,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":3}}   }
...
{"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_01Pi8emvHyK6qoLUVjcnmDCr","name":"hello_world","input":{}}            }
{"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":""}       }
{"type":"content_block_stop","index":1   }

When parsing the event with type EventType.CONTENT_BLOCK_STOP, the ToolUseAggregationEvent is not returned due to the partialJson being an empty String. In my IDE, this looks like:

where you can see the !eventAggregator.isEmpty() only passes if partialJson has a non-empty value. That's resulting in StreamHelper#mergeToolUseEvents returning the ContentBlockStopEvent instead of the ToolUseAggregationEvent being finished and returned.

In my opinion, either the isEmpty check should still return true if the partialJson is empty, or if we need to avoid the squashIntoContentBlock() call, a different definition should be checked if eventAggregatgion applies on ContentBlockStop (this would require the getToolContentBlocks() check to be modified downstream as well).

Environment
spring-ai 1.0.0-M7
Amazon Corretto JDK 21

Steps to reproduce
Using Anthropic AutoConfiguration, invoke:

chatClient.prompt().user("Call the hello_world tool")).stream().chatResponse();

with a simple nullary tool defined:

@Tool(name = "hello_world")
String helloWorld() {
	return "Hi spring-ai!";
}

Set a breakpoint in that tool, or in StreamHelper as per the above screenshot. You will see that the tool is never invoked.

Expected behavior
Nullary tools should still be invoked when streaming -- I note that the tool call works normally the same as a function with positive arity (still running into #2670) if I manually set the partialJson to "{}" in the debugger

This seems similar to this same issue involving Suppliers with Bedrock Converse (#1878), though the code is separate. While AnthropicApi has its own StreamHelper class; Bedrock Converse has a similar check for isEmpty() in ConverseApiUtils with its own forked ToolUseAggregationEvent class.

[bnutzer]

I'd like to confirm this issue. The same behavior is exhibited (again?) when using anthropic models in AWS Bedrock. Non-Anthropic in AWS work correctly.

This is a minimized test class:

@RestController
public class ChatBotController {

    private final ChatClient chatClient;

    public ChatBotController(ChatClient.Builder chatClientBuilder) {
        this.chatClient = chatClientBuilder.build();
    }

    static class MyTools {

        @Tool(description = "This tool reveals the secret.")
        public String getSecret() {
            return "42 is NOT the answer";
        }
    }

    @GetMapping("/streamchat")
    public Flux<String> getMessageStream(@RequestParam String query) {

        return chatClient
                .prompt()
                .advisors(new SimpleLoggerAdvisor())
                .tools(new MyTools())
                .user(query)
                .stream()
                .content();
    }

As soon as I add a @ToolParam for my tool, it is called successfully.

[MrLys]

I can confirm this, but would also like to add that adding:

    static class MyTools {

        @Tool(description = "This tool reveals the secret.")
        public String getSecret(
                @ToolParam(description = "The secret to reveal", required = false) String secret
        ) {
            return "42 is NOT the answer, it is " + secret;
        }
    }

also does not work. That is ToolParam with required=false prevents the tool from being invoked when streaming.