wehub-resource-sync
336 lines
12 KiB
Python
336 lines
12 KiB
Python
"""
|
||
GenericAgentV2 - 使用 <think></think> <answer></answer> 格式的纯视觉 GUI Agent。
|
||
|
||
特点：
|
||
- 输出格式为 <think>思考过程</think> <answer>JSON动作</answer>
|
||
- 不包含当前路由信息（模拟真实的纯视觉手机 GUI Agent）
|
||
- 支持环境的完整动作空间（除 NOOP 外）
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import re
|
||
from typing import Any, ClassVar, Optional
|
||
|
||
from bench_env.agent.base import BaseAgent, AgentConfig, ActionMapping, AgentStepRecord
|
||
from bench_env.env.base import Action, ActionType, Observation
|
||
from bench_env.llm import LLMClient
|
||
|
||
|
||
class GenericAgentV2(BaseAgent):
|
||
"""
|
||
纯视觉 GUI Agent，使用 think/answer 格式。
|
||
|
||
特点：
|
||
- 输出格式：<THINK>思考过程</THINK> <ANSWER>JSON动作</ANSWER>
|
||
- 适用于评估模型的纯视觉 GUI 操作能力
|
||
"""
|
||
|
||
SYSTEM_PROMPT: ClassVar[str] = """你是一个手机 GUI-Agent 操作专家。你需要根据用户下发的任务、手机屏幕截图以及历史操作记录，分析当前界面并输出一个动作来与手机交互，从而完成任务。
|
||
|
||
坐标系：左上角为原点，x 向右，y 向下，取值范围均为 0-1000（归一化坐标）。
|
||
|
||
可用动作（JSON 格式）：
|
||
|
||
1. 点击：{"action": "CLICK", "point": [x, y]}
|
||
2. 双击：{"action": "DOUBLE_TAP", "point": [x, y]}
|
||
3. 长按：{"action": "LONGPRESS", "point": [x, y]}
|
||
4. 输入：{"action": "TYPE", "value": "文本内容"} // 可选 "point": [x, y] 指定输入位置；可选 "clear": true 先清空输入框再输入（默认追加到已有文本后面）
|
||
5. 滑动：{"action": "SWIPE", "point1": [x1, y1], "point2": [x2, y2]}
|
||
6. 拖拽：{"action": "DRAG", "point1": [x1, y1], "point2": [x2, y2]} // 按住起点拖动到终点
|
||
7. 返回：{"action": "BACK"}
|
||
8. 回到桌面：{"action": "HOME"}
|
||
9. 打开最近任务：{"action": "RECENT"}
|
||
10. 输入回车：{"action": "ENTER"}
|
||
11. 等待：{"action": "WAIT", "value": 秒数}
|
||
12. 打开应用：{"action": "AWAKE", "value": "应用名称"}
|
||
13. 提交答案：{"action": "ANSWER", "value": "纯答案文本"}
|
||
14. 任务完成：{"action": "COMPLETE", "return": "完成说明"} // 所有任务完成后使用，给出简短的说明
|
||
15. 中止任务：{"action": "ABORT", "value": "中止原因"} // 任务无法完成时使用，需要说明原因
|
||
|
||
|
||
你必须按以下格式输出：
|
||
|
||
<THINK>
|
||
在这里描述你对当前屏幕的理解、分析和决策过程。
|
||
包括：
|
||
1. 当前屏幕显示的内容是什么
|
||
2. 为了完成任务，下一步应该做什么
|
||
3. 具体要点击/操作哪个元素
|
||
</THINK>
|
||
<ANSWER>
|
||
{
|
||
"action": "动作类型",
|
||
// 根据动作类型填写相应参数
|
||
}
|
||
</ANSWER>
|
||
|
||
|
||
要求：
|
||
- 坐标必须为数字，范围 0-1000
|
||
- JSON 必须是有效格式
|
||
- 仔细观察屏幕截图，根据视觉信息做出判断
|
||
- 需要回答问题时，必须使用 ANSWER 提交答案
|
||
- COMPLETE 只用于结束任务，需要在执行完任务后使用
|
||
"""
|
||
|
||
ACTION_MAP: ActionMapping = {
|
||
"CLICK": (ActionType.CLICK, lambda p: {"point": p.get("point")}),
|
||
"TAP": (ActionType.CLICK, lambda p: {"point": p.get("point")}),
|
||
"DOUBLE_TAP": (ActionType.DOUBLE_TAP, lambda p: {"point": p.get("point")}),
|
||
"DOUBLETAP": (ActionType.DOUBLE_TAP, lambda p: {"point": p.get("point")}),
|
||
"LONGPRESS": (ActionType.LONG_PRESS, lambda p: {"point": p.get("point")}),
|
||
"LONG_PRESS": (ActionType.LONG_PRESS, lambda p: {"point": p.get("point")}),
|
||
"TYPE": (ActionType.TYPE, lambda p: {"value": p.get("value", p.get("text", "")), "point": p.get("point"), "clear": p.get("clear", False)}),
|
||
"SLIDE": (ActionType.SWIPE, lambda p: {"point1": p.get("point1", p.get("start")), "point2": p.get("point2", p.get("end"))}),
|
||
"SWIPE": (ActionType.SWIPE, lambda p: {"point1": p.get("point1", p.get("start")), "point2": p.get("point2", p.get("end"))}),
|
||
"DRAG": (ActionType.DRAG, lambda p: {"point1": p.get("point1", p.get("start")), "point2": p.get("point2", p.get("end"))}),
|
||
"BACK": (ActionType.BACK, lambda p: {}),
|
||
"HOME": (ActionType.HOME, lambda p: {}),
|
||
"RECENT": (ActionType.RECENT, lambda p: {}),
|
||
"ENTER": (ActionType.ENTER, lambda p: {}),
|
||
"WAIT": (ActionType.WAIT, lambda p: {"value": float(p.get("value", p.get("duration", 1.0)))}),
|
||
"AWAKE": (ActionType.AWAKE, lambda p: {"value": p.get("value", p.get("app", ""))}),
|
||
"LAUNCH": (ActionType.AWAKE, lambda p: {"value": p.get("value", p.get("app", ""))}),
|
||
"ANSWER": (ActionType.ANSWER, lambda p: {"value": p.get("value", p.get("text", ""))}),
|
||
"COMPLETE": (ActionType.COMPLETE, lambda p: {"return": p.get("return", p.get("message", ""))}),
|
||
"FINISH": (ActionType.COMPLETE, lambda p: {"return": p.get("return", p.get("message", ""))}),
|
||
"ABORT": (ActionType.ABORT, lambda p: {"value": p.get("value", p.get("reason", ""))}),
|
||
}
|
||
|
||
DEFAULT_MODEL_ARGS: ClassVar[dict[str, Any]] = {
|
||
"temperature": 0.1,
|
||
"top_p": 0.95,
|
||
"frequency_penalty": 0.0,
|
||
"max_tokens": 8192,
|
||
# "reasoning_effort": "none",
|
||
# "extra_body": {
|
||
# "chat_template_kwargs": {"enable_thinking": True}, # Qwen3.x via chat template ✓
|
||
# "enable_thinking": False, # 旧版 vLLM 扁平参数,对 Qwen3.6-35B-A3B 无效
|
||
# "reasoning_effort": "none", # OpenAI o1/o3 系
|
||
# "reasoning": {"effort": "none"}, # OpenAI GPT-5 系
|
||
# },
|
||
}
|
||
|
||
# ==================== 初始化 ====================
|
||
|
||
def __init__(self, llm: LLMClient, config: Optional[AgentConfig] = None):
|
||
super().__init__(config)
|
||
self.llm = llm
|
||
|
||
merged_args = dict(self.DEFAULT_MODEL_ARGS)
|
||
merged_args.update(self.config.model_args or {})
|
||
self.config.model_args = merged_args
|
||
|
||
@property
|
||
def name(self) -> str:
|
||
return "GenericAgentV2"
|
||
|
||
def reset(self, task: str) -> None:
|
||
self._task = task
|
||
self._history = []
|
||
|
||
# ==================== 响应解析 ====================
|
||
|
||
@staticmethod
|
||
def _extract_think_answer(text: str) -> tuple[str, str]:
|
||
"""
|
||
从文本中提取 <think> 和 <answer> 内容。
|
||
|
||
Returns:
|
||
(think_content, answer_content)
|
||
"""
|
||
think_content = ""
|
||
answer_content = ""
|
||
|
||
# 提取 <think>...</think>（不区分大小写，兼容 <THINK>/<think>）
|
||
think_match = re.search(r'<think>(.*?)</think>', text, re.DOTALL | re.IGNORECASE)
|
||
if think_match:
|
||
think_content = think_match.group(1).strip()
|
||
|
||
# 提取 <answer>...</answer>（不区分大小写，兼容 <ANSWER>/<answer>）
|
||
answer_match = re.search(r'<answer>(.*?)</answer>', text, re.DOTALL | re.IGNORECASE)
|
||
if answer_match:
|
||
answer_content = answer_match.group(1).strip()
|
||
|
||
return think_content, answer_content
|
||
|
||
@staticmethod
|
||
def _extract_first_json(text: str) -> Optional[str]:
|
||
"""从文本中提取第一个 JSON 对象"""
|
||
s = text
|
||
start = s.find("{")
|
||
if start < 0:
|
||
return None
|
||
in_str = False
|
||
esc = False
|
||
depth = 0
|
||
for i in range(start, len(s)):
|
||
ch = s[i]
|
||
if in_str:
|
||
if esc:
|
||
esc = False
|
||
elif ch == "\\":
|
||
esc = True
|
||
elif ch == '"':
|
||
in_str = False
|
||
else:
|
||
if ch == '"':
|
||
in_str = True
|
||
elif ch == "{":
|
||
depth += 1
|
||
elif ch == "}":
|
||
depth -= 1
|
||
if depth == 0:
|
||
return s[start : i + 1]
|
||
return None
|
||
|
||
def _parse_llm_output(self, response_text: str) -> tuple[str, dict[str, Any]]:
|
||
"""
|
||
解析 LLM 输出为 (thought, action_dict)。
|
||
|
||
Returns:
|
||
(thought, action_dict)
|
||
"""
|
||
raw = str(response_text or "").strip()
|
||
if not raw:
|
||
return "", {"_error": "empty_response"}
|
||
|
||
# 提取 think 和 answer 部分
|
||
think_content, answer_content = self._extract_think_answer(raw)
|
||
|
||
# 如果没有找到 <answer> 标签，尝试直接解析整个文本
|
||
if not answer_content:
|
||
answer_content = raw
|
||
|
||
# 解析 answer 中的 JSON
|
||
parsed: Any = None
|
||
try:
|
||
parsed = json.loads(answer_content)
|
||
except Exception:
|
||
extracted = self._extract_first_json(answer_content)
|
||
if extracted:
|
||
try:
|
||
parsed = json.loads(extracted)
|
||
except Exception:
|
||
pass
|
||
|
||
if not isinstance(parsed, dict):
|
||
return think_content, {"_error": "invalid_json", "raw": raw}
|
||
|
||
return think_content, parsed
|
||
|
||
def parse_response(self, response_text: str) -> Action:
|
||
"""解析 LLM 响应为 Action"""
|
||
thought, parsed = self._parse_llm_output(response_text)
|
||
|
||
if "_error" in parsed:
|
||
return Action(
|
||
action_type=ActionType.ABORT,
|
||
data={"value": parsed.get("_error", "parse_error")},
|
||
raw_response=response_text,
|
||
)
|
||
|
||
action_name = str(parsed.get("action") or parsed.get("action_type") or "").strip().upper()
|
||
explain = str(parsed.get("explain", "") or "")
|
||
|
||
return self.parse_action(
|
||
action_name,
|
||
parsed,
|
||
thought=thought,
|
||
explain=explain,
|
||
raw_response=response_text,
|
||
)
|
||
|
||
# ==================== 消息构建 ====================
|
||
|
||
def build_messages(self, obs: Observation) -> list[dict]:
|
||
"""
|
||
构建发送给 LLM 的消息。
|
||
|
||
使用完整多轮对话历史（类似 AutoGLMAgent）：
|
||
- 历史步骤：user（任务/步骤标记） + assistant（模型响应）
|
||
- 当前步骤：user（截图+提示）
|
||
|
||
注意：不包含当前路由信息，作为纯视觉 GUI Agent。
|
||
"""
|
||
messages: list[dict] = [
|
||
{"role": "system", "content": self.SYSTEM_PROMPT}
|
||
]
|
||
|
||
# 构建历史对话
|
||
for i, record in enumerate(self._history):
|
||
# 历史 user 消息（不含截图）
|
||
if i == 0:
|
||
user_text = f"[任务]\n{self._task}"
|
||
else:
|
||
user_text = f"[Step {i + 1}]"
|
||
|
||
messages.append({
|
||
"role": "user",
|
||
"content": [{"type": "text", "text": user_text}],
|
||
})
|
||
|
||
# 历史 assistant 响应
|
||
messages.append({
|
||
"role": "assistant",
|
||
"content": record.llm_response,
|
||
})
|
||
|
||
# 当前步骤（带截图）
|
||
step_num = len(self._history) + 1
|
||
if len(self._history) == 0:
|
||
user_text = f"[任务]\n{self._task}"
|
||
else:
|
||
user_text = f"[Step {step_num}]"
|
||
|
||
messages.append({
|
||
"role": "user",
|
||
"content": [
|
||
{"type": "image_url", "image_url": {"url": obs.image_data_url}},
|
||
{"type": "text", "text": user_text},
|
||
],
|
||
})
|
||
|
||
return messages
|
||
|
||
# ==================== 核心逻辑 ====================
|
||
|
||
def act(self, obs: Observation) -> Action:
|
||
"""生成动作"""
|
||
messages = self.build_messages(obs)
|
||
|
||
if self.config.verbose:
|
||
print(f"\n[GenericAgentV2] Step {obs.step_idx}, sending prompt...")
|
||
|
||
response = self.llm.chat(
|
||
messages=messages,
|
||
args={
|
||
**self.config.model_args,
|
||
"stream": self.config.stream,
|
||
"stream_print": self.config.stream,
|
||
},
|
||
)
|
||
|
||
if self.config.verbose and not self.config.stream:
|
||
print(f"\n[LLM Response]\n{response.content}\n")
|
||
|
||
action = self.parse_response(response.content)
|
||
|
||
self._history.append(AgentStepRecord(
|
||
step_idx=obs.step_idx,
|
||
observation=obs,
|
||
action=action,
|
||
llm_response=response.content,
|
||
llm_prompt=messages,
|
||
))
|
||
|
||
# 内存瘦身：历史仅用文本，保留最近 2 条完整记录
|
||
self._evict_old_records(keep_recent=2)
|
||
|
||
if self.config.verbose:
|
||
print(f"[GenericAgentV2] Action: {action.action_type}, Data: {action.data}")
|
||
|
||
return action
|