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
+                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

@@ -219,25 +219,51 @@ 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}"
+        stream_failed = True
-        assistant_parts = [fallback]
+        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}"
+        stream_failed = True
-        assistant_parts.append("\n\n" + fallback)
+        stream_error = str(exc)
         logger.exception(
             "Unexpected stream failure",
             extra={"conversation_id": conversation.pk, "error": str(exc)},
         )
+
+    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())

tests/test_llm_streaming.py

@@ -0,0 +1,41 @@
+import io
+from urllib import request
+
+import pytest
+
+from review_agent.llm import stream_reply
+from review_agent.models import Conversation
+
+
+pytestmark = pytest.mark.django_db
+
+
+class FakeStreamingResponse:
+    def __iter__(self):
+        return iter(
+            [
+                b'data: {"choices":[{"delta":{"content":"A"}}]}\n\n',
+                b"data: not-json\n\n",
+                b'data: {"choices":[{"delta":{"content":"B"}}]}\n\n',
+                b"data: [DONE]\n\n",
+            ]
+        )
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc, traceback):
+        return False
+
+
+def test_stream_reply_skips_malformed_sse_data(monkeypatch, settings, django_user_model):
+    settings.LLM_API_KEY = "key"
+    settings.LLM_MODEL = "model"
+    settings.LLM_BASE_URL = "https://example.test/v1"
+    monkeypatch.setattr(request, "urlopen", lambda req, timeout: FakeStreamingResponse())
+    user = django_user_model.objects.create_user(username="owner", password="pass")
+    conversation = Conversation.objects.create(user=user, title="会话")
+
+    chunks = list(stream_reply(conversation, "你好"))
+
+    assert chunks == ["A", "B"]