fix(agent): 增强 LLM 流式回复兜底

review_agent/llm.py

@@ -1,4 +1,5 @@
 import json
+import logging
 from urllib import error, request
 
 from django.conf import settings
@@ -12,6 +13,9 @@ class LLMRequestError(RuntimeError):
     """Raised when the remote LLM provider call fails."""
 
 
+logger = logging.getLogger(__name__)
+
+
 def generate_reply(conversation, user_message: str) -> str:
     """Calls the SiliconFlow OpenAI-compatible chat endpoint and returns assistant text."""
 
@@ -130,7 +134,11 @@ def stream_reply(conversation, user_message: str):
                 data = line[5:].strip()
                 if data == "[DONE]":
                     break
-                payload = json.loads(data)
+                try:
+                    payload = json.loads(data)
+                except json.JSONDecodeError:
+                    logger.warning("Skipping malformed LLM stream data", extra={"data": data[:200]})
+                    continue
                 delta = (
                     payload.get("choices", [{}])[0]
                     .get("delta", {})

review_agent/services.py

@@ -219,26 +219,52 @@ def stream_message(conversation: Conversation, content: str):
         )
         return
 
+    stream_failed = False
+    stream_error = ""
     try:
         for chunk in stream_reply(conversation, content):
             assistant_parts.append(chunk)
             yield sse_event("chunk", {"delta": chunk})
     except (LLMConfigurationError, LLMRequestError) as exc:
-        fallback = f"模型调用失败：{exc}"
-        assistant_parts = [fallback]
+        stream_failed = True
+        stream_error = str(exc)
         logger.warning(
             "LLM stream failed",
             extra={"conversation_id": conversation.pk, "error": str(exc)},
         )
-        yield sse_event("error", {"message": fallback})
     except Exception as exc:
-        fallback = f"回复生成中断：{exc}"
-        assistant_parts.append("\n\n" + fallback)
+        stream_failed = True
+        stream_error = str(exc)
         logger.exception(
             "Unexpected stream failure",
             extra={"conversation_id": conversation.pk, "error": str(exc)},
         )
-        yield sse_event("error", {"message": fallback})
+
+    if stream_failed:
+        try:
+            fallback_reply = generate_reply(conversation, content)
+            assistant_parts = [fallback_reply]
+            logger.info(
+                "Non-stream fallback reply succeeded",
+                extra={"conversation_id": conversation.pk, "content_length": len(fallback_reply)},
+            )
+            yield sse_event("replace", {"content": fallback_reply})
+        except (LLMConfigurationError, LLMRequestError) as exc:
+            fallback = f"模型调用失败：{exc}"
+            assistant_parts = [fallback]
+            logger.warning(
+                "Non-stream fallback reply failed",
+                extra={"conversation_id": conversation.pk, "error": str(exc), "stream_error": stream_error},
+            )
+            yield sse_event("error", {"message": fallback})
+        except Exception as exc:
+            fallback = f"回复生成中断：{stream_error or exc}"
+            assistant_parts.append("\n\n" + fallback)
+            logger.exception(
+                "Non-stream fallback crashed",
+                extra={"conversation_id": conversation.pk, "error": str(exc), "stream_error": stream_error},
+            )
+            yield sse_event("error", {"message": fallback})
 
     assistant_message = append_assistant_message(conversation, "".join(assistant_parts).strip())