fix _parse_response 解析llm返回内容错误

[wang0618]

问题复现

Open-AutoGLM/phone_agent/model/client.py

Line 176 in 86f5538

def _parse_response(self, content: str) -> tuple[str, str]:

在遇到

<think>先思考</think>
<answer>do(action="Launch", app="知乎")</answer>

这种格式的llm响应时，会将action错误解析成 do(action="Launch", app="知乎")</answer>

复现代码：

def parse_response(content: str) -> tuple[str, str]:
    # Rule 1: Check for finish(message=
    if "finish(message=" in content:
        parts = content.split("finish(message=", 1)
        thinking = parts[0].strip()
        action = "finish(message=" + parts[1]
        return thinking, action

    # Rule 2: Check for do(action=
    if "do(action=" in content:
        parts = content.split("do(action=", 1)
        thinking = parts[0].strip()
        action = "do(action=" + parts[1]
        return thinking, action

    # Rule 3: Fallback to legacy XML tag parsing
    if "<answer>" in content:
        parts = content.split("<answer>", 1)
        thinking = parts[0].replace("<think>", "").replace("</think>", "").strip()
        action = parts[1].replace("</answer>", "").strip()
        return thinking, action

    # Rule 4: No markers found, return content as action
    return "", content

raw_content = """
<think>用户需要打开知乎查看首页第一条消息，当前在系统桌面，首先需要启动知乎应用，因此执行Launch操作打开知乎。</think>
<answer>
do(action="Launch", app="知乎")
</answer>
"""
thinking, action = parse_response(raw_content)
print("thinking:", thinking)
print("action:", action)

本PR修复了此问题，并补充测试用例