fix：修复回复器格式问题，记录完整空回复请求

src/llm_models/exceptions.py

@@ -54,27 +54,30 @@ class RespNotOkException(Exception):
             return f"未知的异常响应代码：{self.status_code}"
 
 
-class RespParseException(Exception):
-    """响应解析错误，常见于响应格式不正确或解析方法不匹配"""
+class ResponseContextException(Exception):
+    """携带原始响应上下文的异常基类。"""
 
-    def __init__(self, ext_info: Any, message: str | None = None):
+    default_message: str = "请求失败"
+
+    def __init__(self, ext_info: Any = None, message: str | None = None):
         super().__init__(message)
         self.ext_info = ext_info
         self.message = message
 
     def __str__(self):
-        return self.message or "解析响应内容时发生未知错误，请检查是否配置了正确的解析方法"
+        return self.message or self.default_message
 
 
-class EmptyResponseException(Exception):
+class RespParseException(ResponseContextException):
+    """响应解析错误，常见于响应格式不正确或解析方法不匹配"""
+
+    default_message = "解析响应内容时发生未知错误，请检查是否配置了正确的解析方法"
+
+
+class EmptyResponseException(ResponseContextException):
     """响应内容为空"""
 
-    def __init__(self, message: str = "响应内容为空，这可能是一个临时性问题"):
-        super().__init__(message)
-        self.message = message
-
-    def __str__(self):
-        return self.message
+    default_message = "响应内容为空，这可能是一个临时性问题"
 
 
 class ModelAttemptFailed(Exception):

src/llm_models/model_client/gemini_client.py

@@ -552,7 +552,7 @@ def _build_stream_api_response(
 
     _warn_if_max_tokens_truncated(last_response, response.content, response.tool_calls)
     if not response.content and not response.tool_calls and not response.reasoning_content:
-        raise EmptyResponseException()
+        raise EmptyResponseException(last_response)
     return response
 
 
@@ -627,7 +627,7 @@ def _default_normal_response_parser(
     usage_record = _extract_usage_record(response)
     _warn_if_max_tokens_truncated(response, api_response.content, api_response.tool_calls)
     if not api_response.content and not api_response.tool_calls and not api_response.reasoning_content:
-        raise EmptyResponseException("响应中既无文本内容也无工具调用")
+        raise EmptyResponseException(response, "响应中既无文本内容也无工具调用")
     return api_response, usage_record
 
 

src/llm_models/model_client/openai_client.py

@@ -587,7 +587,7 @@ def _build_api_status_message(error: APIStatusError) -> str:
         message_parts.append(str(error.message))
     response_text = getattr(getattr(error, "response", None), "text", None)
     if response_text:
-        message_parts.append(str(response_text)[:300])
+        message_parts.append(str(response_text))
     if message_parts:
         return " | ".join(message_parts)
     return f"上游接口返回状态码 {error.status_code}"
@@ -750,7 +750,7 @@ class _OpenAIStreamAccumulator:
         response.raw_data = {"model": self.model_name} if self.model_name else None
 
         if not response.content and not response.tool_calls:
-            raise EmptyResponseException()
+            raise EmptyResponseException(response.raw_data)
 
         return response
 
@@ -834,7 +834,7 @@ def _default_normal_response_parser(
     """
     choices = getattr(resp, "choices", None)
     if not choices:
-        raise EmptyResponseException("响应解析失败，choices 为空或缺失")
+        raise EmptyResponseException(resp, "响应解析失败，choices 为空或缺失")
 
     api_response = APIResponse()
     message_part = choices[0].message
@@ -875,7 +875,7 @@ def _default_normal_response_parser(
     _log_length_truncation(finish_reason, getattr(resp, "model", None))
 
     if not api_response.content and not api_response.tool_calls:
-        raise EmptyResponseException()
+        raise EmptyResponseException(resp)
 
     return api_response, usage_record
 

src/llm_models/request_snapshot.py

@@ -58,6 +58,42 @@ def _json_friendly(value: Any) -> Any:
     return str(value)
 
 
+def extract_error_response_body(error: Exception) -> Any | None:
+    """尽量从异常对象中提取上游返回体，便于排查模型请求失败。"""
+    candidate_errors = [error, getattr(error, "__cause__", None)]
+
+    for candidate in candidate_errors:
+        if candidate is None:
+            continue
+
+        response = getattr(candidate, "response", None)
+        if response is not None:
+            response_json = getattr(response, "json", None)
+            if callable(response_json):
+                try:
+                    return _json_friendly(response_json())
+                except Exception:
+                    pass
+
+            response_text = getattr(response, "text", None)
+            if response_text not in (None, ""):
+                return str(response_text)
+
+            response_content = getattr(response, "content", None)
+            if response_content not in (None, b"", ""):
+                return _json_friendly(response_content)
+
+        response_body = getattr(candidate, "body", None)
+        if response_body not in (None, "", b""):
+            return _json_friendly(response_body)
+
+        ext_info = getattr(candidate, "ext_info", None)
+        if ext_info is not None:
+            return _json_friendly(ext_info)
+
+    return None
+
+
 def _sanitize_filename_component(value: str) -> str:
     """将任意字符串转换为适合文件名使用的片段。"""
     normalized_value = FILENAME_SAFE_PATTERN.sub("-", value.strip())
@@ -388,6 +424,10 @@ def save_failed_request_snapshot(
             "snapshot_version": SNAPSHOT_VERSION,
         }
 
+        response_body = extract_error_response_body(error)
+        if response_body is not None:
+            snapshot_payload["error"]["response_body"] = response_body
+
         snapshot_payload["replay"] = {
             "command": build_replay_command(snapshot_path),
             "file_uri": snapshot_path.as_uri(),