feat: 增强工具调用代理功能，支持多工具调用和消息历史转换

主要改进:
- 新增 convert_tool_calls_to_content 函数，将消息历史中的 tool_calls 转换为 LLM 可理解的 XML 格式
- 修复 response_parser 支持同时解析多个 tool_calls
- 优化响应解析逻辑，支持 content 和 tool_calls 同时存在
- 添加完整的测试覆盖，包括多工具调用、消息转换和混合响应

技术细节:
- services.py: 实现工具调用历史到 content 的转换
- response_parser.py: 使用非贪婪匹配支持多个 tool_calls 解析
- main.py: 集成消息转换功能，确保消息历史正确传递给 LLM

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

app/services.py

@@ -39,6 +39,70 @@ def _parse_sse_data(chunk: bytes) -> Optional[Dict[str, Any]]:
 # --- End Helper ---
 
 
+def convert_tool_calls_to_content(messages: List[ChatMessage]) -> List[ChatMessage]:
+    """
+    Converts assistant messages with tool_calls into content format using XML tags.
+
+    This function processes the message history and converts any assistant messages
+    that have tool_calls into a format that LLMs can understand. The tool_calls
+    are converted to <invoke>...</invoke> tags in the content field.
+
+    Args:
+        messages: List of ChatMessage objects from the client
+
+    Returns:
+        Processed list of ChatMessage objects with tool_calls converted to content
+
+    Example:
+        Input:  [{"role": "assistant", "tool_calls": [...]}]
+        Output: [{"role": "assistant", "content": "<invoke>{...}</invoke>"}]
+    """
+    from .response_parser import TOOL_CALL_START_TAG, TOOL_CALL_END_TAG
+
+    processed_messages = []
+
+    for msg in messages:
+        # Check if this is an assistant message with tool_calls
+        if msg.role == "assistant" and msg.tool_calls and len(msg.tool_calls) > 0:
+            # Convert each tool call to XML tag format
+            tool_call_contents = []
+            for tc in msg.tool_calls:
+                tc_data = tc.get("function", {})
+                name = tc_data.get("name", "")
+                arguments_str = tc_data.get("arguments", "{}")
+
+                # Parse arguments JSON to ensure it's valid
+                try:
+                    arguments = json.loads(arguments_str) if isinstance(arguments_str, str) else arguments_str
+                except json.JSONDecodeError:
+                    arguments = {}
+
+                # Build the tool call JSON
+                tool_call_json = {"name": name, "arguments": arguments}
+
+                # Wrap in XML tags
+                tool_call_content = f'{TOOL_CALL_START_TAG}{json.dumps(tool_call_json, ensure_ascii=False)}{TOOL_CALL_END_TAG}'
+                tool_call_contents.append(tool_call_content)
+
+            # Create new message with tool calls in content
+            # Preserve original content if it exists
+            content_parts = []
+            if msg.content:
+                content_parts.append(msg.content)
+
+            content_parts.extend(tool_call_contents)
+            new_content = "\n".join(content_parts)
+
+            processed_messages.append(
+                ChatMessage(role=msg.role, content=new_content)
+            )
+        else:
+            # Keep other messages as-is
+            processed_messages.append(msg)
+
+    return processed_messages
+
+
 def inject_tools_into_prompt(messages: List[ChatMessage], tools: List[Tool]) -> List[ChatMessage]:
     """
     Injects a system prompt with tool definitions at the beginning of the message list.