Optimization of thinking content acquisition in thinking models

Spring AI's retrieval of reasoning content from deep-thinking model responses is overly cumbersome. Can you provide a unified solution for obtaining reasoning content?

Currently, processing thinking content is very cumbersome, as follows:
`
AtomicBoolean atomicBoolean = new AtomicBoolean(false);
clientRequestSpec
.stream()
.chatResponse()
.doOnNext(response -> {
AssistantMessage output = response.getResult().getOutput();
String text = output.getText();
if(StringUtils.hasText(text)){
atomicBoolean.set(true);
}
boolean noThink = atomicBoolean.get();
if(!noThink){
String reasoningContent ;
if(output instanceof DeepSeekAssistantMessage deepSeekAssistantMessage){
reasoningContent = deepSeekAssistantMessage.getReasoningContent();
}else{
Map<String, Object> metadata = output.getMetadata();
Object o = metadata.get("reasoningContent");
reasoningContent = o != null ? o.toString() : "";
}

                }
              
            })
            .doOnError(error->{
                String errorMsg = error.getMessage();
            })
            .doOnComplete(() -> {})
            .subscribe();`

In LangChain4j, a dedicated interface for acquiring thinking content is provided, which is very convenient, as follows:
`public interface TokenStream {

TokenStream onPartialResponse(Consumer<String> partialResponseHandler);

/**
 * The provided consumer will be invoked every time a new partial thinking/reasoning text (usually a single token)
 * from a language model is available.
 */
default TokenStream onPartialThinking(Consumer<PartialThinking> partialThinkingHandler) {
    throw new UnsupportedOperationException("not implemented");
}

TokenStream onCompleteResponse(Consumer<ChatResponse> completeResponseHandler);


TokenStream onError(Consumer<Throwable> errorHandler);

}
`

I hope to get your help. thanks!

Start by tracing clientRequestSpec.stream().chatResponse() and how AssistantMessage, DeepSeekAssistantMessage, and response metadata currently expose reasoning content. Compare this flow with the cited LangChain4j TokenStream callbacks; done should mean a documented, consistent way to obtain thinking content across supported model responses.