Spring AI Why toolCallArguments cannot be null or empty?

Bug description A simple toolcalling for asking current datetime, but i receive toolCallArguments cannot be null or empty

Environment 1.0.0-RC1 spring-ai-starter-model-openai

Steps to reproduce

@Slf4j
public class DatetimeTools {
  private static final DateTimeFormatter formatter = DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss");

  @Tool(description = "Query current date and time")
  public String getCurrentDatetime() {
    String currentDatetime = LocalDateTime.now().format(formatter);
    return currentDatetime;
  }
}

client.prompt()
        .user("What time is it now? ")
        .tools(new DatetimeTools())
        .stream()
        .content()
        .subscribe(log::info);

Expected behavior It's ok when i use sync calling, but fail on stream calling. Anything i missed?

Minimal Complete Reproducible example

Comment From: ThomasVitale

@lanyuanxiaoyao I tried running your same example, but it works correctly for me. Still, in order to be resilient to situations where a model implementation uses null toolCallArguments instead of {}, I relaxed the constraint and handled it internally in ToolCallingObservationContext. See https://github.com/spring-projects/spring-ai/pull/3236

Thanks for reporting this issue!

Comment From: lanyuanxiaoyao

@lanyuanxiaoyao I tried running your same example, but it works correctly for me. Still, in order to be resilient to situations where a model implementation uses null toolCallArguments instead of {}, I relaxed the constraint and handled it internally in ToolCallingObservationContext. See #3236我尝试运行你的相同示例，但它对我来说工作正常。尽管如此，为了能够适应模型实现使用 null toolCallArguments 而不是 {} 的情况，我放宽了约束并在 ToolCallingObservationContext 内部处理了它。参见 #3236

Thanks for reporting this issue!感谢您报告此问题！

Wow, thanks for your quick response to this issue. I originally thought this issue had nothing to do with the model. In fact, I am using vLLM + Qwen3 0.6B for test.

Comment From: ThomasVitale

@lanyuanxiaoyao could you elaborate on your setup? I can see you're using spring-ai-starter-model-openai. So, you use the OpenAI module to call vLLM? Thanks!

Comment From: lanyuanxiaoyao

@ThomasVitale OK

docker run \
  --privileged=true \
  --shm-size=4g \
  --rm \
  -p 8000:8000 \
  -v /root/llm/models:/models \
  -e VLLM_CPU_KVCACHE_SPACE=5 \
  -e VLLM_CPU_OMP_THREADS_BIND=0-11 \
  -e VLLM_CPU_MOE_PREPACK=0 \
  public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.8.5 \
  --model /models/Qwen3-0.6B \
  --port 8000 \
  --host 0.0.0.0 \
  --dtype bfloat16 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --uvicorn-log-level debug \
  --enable-reasoning \
  --reasoning-parser deepseek_r1

And if this is your first time using vLLM + Spring AI, this question may save your time :) https://github.com/spring-projects/spring-ai/issues/3124

Comment From: ThomasVitale

@lanyuanxiaoyao thanks for the additional context

The change I linked in my previous comment is part of 1.0.0, which was released today.