fix: 兼容旧模型内联思考标签

- 在结构化 reasoning 为空时解析正文开头的 think/thinking 标签

- 将思考与正文映射为现有流式事件并覆盖 Agent 与 Bot 链路

- 补充跨分片和旁路条件测试

easyflow-modules/easyflow-module-ai/src/main/java/tech/easyflow/ai/easyagents/listener/ChatStreamListener.java

@@ -23,6 +23,7 @@ import tech.easyflow.core.runtime.ChatRuntimeContext;
 import tech.easyflow.core.runtime.ChatRuntimeExtKeys;
 import tech.easyflow.core.runtime.ChatRuntimeManager;
 import tech.easyflow.core.runtime.ChatRuntimeMessage;
+import tech.easyflow.core.runtime.LegacyThinkingTagParser;
 
 import java.math.BigInteger;
 import java.util.Date;
@@ -43,6 +44,7 @@ public class ChatStreamListener implements StreamResponseListener {
     private final ChatRuntimeManager chatRuntimeManager;
     private final ChatRuntimeContext runtimeContext;
     private final ChatAssistantAccumulator assistantAccumulator;
+    private final LegacyThinkingTagParser legacyThinkingTagParser = new LegacyThinkingTagParser();
     // 核心标记：是否允许执行onStop业务逻辑（仅最后一次无后续工具调用时为true）
     private boolean canStop = true;
     // 辅助标记：是否进入过工具调用（避免重复递归判断）
@@ -65,6 +67,7 @@ public class ChatStreamListener implements StreamResponseListener {
 
     @Override
     public void onStart(StreamContext context) {
+        legacyThinkingTagParser.reset();
         StreamResponseListener.super.onStart(context);
     }
 
@@ -80,6 +83,7 @@ public class ChatStreamListener implements StreamResponseListener {
                 return;
             }
             if (aiMessage.isFinalDelta() && aiMessageResponse.hasToolCalls()) {
+                flushLegacyThinkingSegments();
                 this.canStop = false; // 工具调用期间，禁止执行onStop
                 this.hasToolCall = true; // 标记已进入过工具调用
                 List<ToolCall> toolCalls = aiMessage.getToolCalls();
@@ -103,20 +107,7 @@ public class ChatStreamListener implements StreamResponseListener {
                 if (this.hasToolCall) {
                     this.canStop = true;
                 }
-                String reasoningContent = aiMessage.getReasoningContent();
-                if (reasoningContent != null && !reasoningContent.isEmpty()) {
-                    assistantAccumulator.appendReasoning(reasoningContent);
-                    chatRuntimeManager.recordAssistantDelta(runtimeContext, buildAssistantDeltaMessage(reasoningContent, ChatType.THINKING));
-                    sendChatEnvelope(sseEmitter, reasoningContent, ChatType.THINKING);
-                } else {
-                    String delta = aiMessage.getContent();
-                    if (delta != null && !delta.isEmpty()) {
-                        assistantAccumulator.appendContent(delta);
-                        chatRuntimeManager.recordAssistantDelta(runtimeContext, buildAssistantDeltaMessage(delta, ChatType.MESSAGE));
-                        sendChatEnvelope(sseEmitter, delta, ChatType.MESSAGE);
-                    }
-                }
-
+                handleAssistantDelta(aiMessage);
             }
         } catch (Exception e) {
             LOG.error("Chat stream onMessage failed, conversationId={}, message={}, exception={}",
@@ -137,6 +128,7 @@ public class ChatStreamListener implements StreamResponseListener {
                 sendSystemError(sseEmitter, context.getThrowable().getMessage(), context.getThrowable());
                 return;
             }
+            flushLegacyThinkingSegments();
             memoryPrompt.addMessage(context.getFullMessage());
             chatRuntimeManager.recordAssistantCompleted(runtimeContext, buildAssistantCompletedMessage(context));
             chatRuntimeManager.recordCompleted(runtimeContext);
@@ -186,6 +178,49 @@ public class ChatStreamListener implements StreamResponseListener {
         }
     }
 
+    /**
+     * 归一化并发送单个模型增量。
+     *
+     * @param aiMessage 模型增量消息
+     */
+    private void handleAssistantDelta(AiMessage aiMessage) {
+        String reasoningContent = aiMessage.getReasoningContent();
+        if (StringUtil.hasText(reasoningContent)) {
+            emitAssistantSegments(legacyThinkingTagParser.acceptReasoning(reasoningContent));
+            return;
+        }
+        emitAssistantSegments(legacyThinkingTagParser.acceptContent(aiMessage.getContent()));
+    }
+
+    /**
+     * 发送兼容解析后的思考与正文片段。
+     *
+     * @param segments 解析片段
+     */
+    private void emitAssistantSegments(List<LegacyThinkingTagParser.Segment> segments) {
+        for (LegacyThinkingTagParser.Segment segment : segments) {
+            String text = segment.getText();
+            if (segment.getType() == LegacyThinkingTagParser.SegmentType.REASONING) {
+                assistantAccumulator.appendReasoning(text);
+                chatRuntimeManager.recordAssistantDelta(runtimeContext,
+                        buildAssistantDeltaMessage(text, ChatType.THINKING));
+                sendChatEnvelope(sseEmitter, text, ChatType.THINKING);
+                continue;
+            }
+            assistantAccumulator.appendContent(text);
+            chatRuntimeManager.recordAssistantDelta(runtimeContext,
+                    buildAssistantDeltaMessage(text, ChatType.MESSAGE));
+            sendChatEnvelope(sseEmitter, text, ChatType.MESSAGE);
+        }
+    }
+
+    /**
+     * 收口并发送旧标签解析器仍保留的少量前缀。
+     */
+    private void flushLegacyThinkingSegments() {
+        emitAssistantSegments(legacyThinkingTagParser.finish());
+    }
+
     private void sendToolCallEnvelope(ToolCall toolCall) {
         if (toolCall == null) {
             return;
@@ -330,7 +365,10 @@ public class ChatStreamListener implements StreamResponseListener {
         message.setRole("assistant");
         message.setContentType("TEXT");
         String fullContent = context != null && context.getFullMessage() != null ? context.getFullMessage().getContent() : null;
-        message.setContentText(StringUtil.hasText(fullContent) ? fullContent : assistantAccumulator.getContent());
+        String normalizedContent = assistantAccumulator.getContent();
+        message.setContentText(legacyThinkingTagParser.isLegacyFormatDetected()
+                ? normalizedContent
+                : (StringUtil.hasText(fullContent) ? fullContent : normalizedContent));
         message.setContentPayload(assistantAccumulator.buildPayload(message.getContentText()));
         message.setCreatedAt(new Date());
         message.setSenderId(runtimeContext.getAssistantId());