376 lines
13 KiB
Python
376 lines
13 KiB
Python
"""
|
||
真正的 ReAct Agent 实现
|
||
LLM 是大脑，全程参与决策！
|
||
|
||
ReAct 循环:
|
||
1. Thought: LLM 思考当前状态和下一步
|
||
2. Action: LLM 决定调用哪个工具
|
||
3. Observation: 执行工具，获取结果
|
||
4. 重复直到 LLM 决定完成
|
||
"""
|
||
|
||
import json
|
||
import logging
|
||
import re
|
||
from typing import List, Dict, Any, Optional, Tuple
|
||
from dataclasses import dataclass
|
||
|
||
from .base import BaseAgent, AgentConfig, AgentResult, AgentType, AgentPattern
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
REACT_SYSTEM_PROMPT = """你是 DeepAudit 安全审计 Agent，一个专业的代码安全分析专家。
|
||
|
||
## 你的任务
|
||
对目标项目进行全面的安全审计，发现潜在的安全漏洞。
|
||
|
||
## 你的工具
|
||
{tools_description}
|
||
|
||
## 工作方式
|
||
你需要通过 **思考-行动-观察** 循环来完成任务：
|
||
|
||
1. **Thought**: 分析当前情况，思考下一步应该做什么
|
||
2. **Action**: 选择一个工具并执行
|
||
3. **Observation**: 观察工具返回的结果
|
||
4. 重复上述过程直到你认为审计完成
|
||
|
||
## 输出格式
|
||
每一步必须严格按照以下格式输出：
|
||
|
||
```
|
||
Thought: [你的思考过程，分析当前状态，决定下一步]
|
||
Action: [工具名称]
|
||
Action Input: [工具参数，JSON 格式]
|
||
```
|
||
|
||
当你完成分析后，输出：
|
||
```
|
||
Thought: [总结分析结果]
|
||
Final Answer: [JSON 格式的最终发现]
|
||
```
|
||
|
||
## Final Answer 格式
|
||
```json
|
||
{{
|
||
"findings": [
|
||
{{
|
||
"vulnerability_type": "sql_injection",
|
||
"severity": "high",
|
||
"title": "SQL 注入漏洞",
|
||
"description": "详细描述",
|
||
"file_path": "path/to/file.py",
|
||
"line_start": 42,
|
||
"code_snippet": "危险代码片段",
|
||
"suggestion": "修复建议"
|
||
}}
|
||
],
|
||
"summary": "审计总结"
|
||
}}
|
||
```
|
||
|
||
## 审计策略建议
|
||
1. 先用 list_files 了解项目结构
|
||
2. 识别关键文件（路由、控制器、数据库操作）
|
||
3. 使用 search_code 搜索危险模式（eval, exec, query, innerHTML 等）
|
||
4. 读取可疑文件进行深度分析
|
||
5. 如果有 semgrep，用它进行全面扫描
|
||
|
||
## 重点关注的漏洞类型
|
||
- SQL 注入 (query, execute, raw SQL)
|
||
- XSS (innerHTML, document.write, v-html)
|
||
- 命令注入 (exec, system, subprocess, child_process)
|
||
- 路径遍历 (open, readFile, path concatenation)
|
||
- SSRF (requests, fetch, http client)
|
||
- 硬编码密钥 (password, secret, api_key, token)
|
||
- 不安全的反序列化 (pickle, yaml.load, eval)
|
||
|
||
现在开始审计！"""
|
||
|
||
|
||
@dataclass
|
||
class AgentStep:
|
||
"""Agent 执行步骤"""
|
||
thought: str
|
||
action: Optional[str] = None
|
||
action_input: Optional[Dict] = None
|
||
observation: Optional[str] = None
|
||
is_final: bool = False
|
||
final_answer: Optional[Dict] = None
|
||
|
||
|
||
class ReActAgent(BaseAgent):
|
||
"""
|
||
真正的 ReAct Agent
|
||
|
||
LLM 全程参与决策，自主选择工具和分析策略
|
||
"""
|
||
|
||
def __init__(
|
||
self,
|
||
llm_service,
|
||
tools: Dict[str, Any],
|
||
event_emitter=None,
|
||
agent_type: AgentType = AgentType.ANALYSIS,
|
||
max_iterations: int = 30,
|
||
):
|
||
config = AgentConfig(
|
||
name="ReActAgent",
|
||
agent_type=agent_type,
|
||
pattern=AgentPattern.REACT,
|
||
max_iterations=max_iterations,
|
||
system_prompt=REACT_SYSTEM_PROMPT,
|
||
)
|
||
super().__init__(config, llm_service, tools, event_emitter)
|
||
|
||
self._conversation_history: List[Dict[str, str]] = []
|
||
self._steps: List[AgentStep] = []
|
||
|
||
def _get_tools_description(self) -> str:
|
||
"""生成工具描述"""
|
||
descriptions = []
|
||
|
||
for name, tool in self.tools.items():
|
||
if name.startswith("_"):
|
||
continue
|
||
|
||
desc = f"### {name}\n"
|
||
desc += f"{tool.description}\n"
|
||
|
||
# 添加参数说明
|
||
if hasattr(tool, 'args_schema') and tool.args_schema:
|
||
schema = tool.args_schema.schema()
|
||
properties = schema.get("properties", {})
|
||
if properties:
|
||
desc += "参数:\n"
|
||
for param_name, param_info in properties.items():
|
||
param_desc = param_info.get("description", "")
|
||
param_type = param_info.get("type", "string")
|
||
desc += f" - {param_name} ({param_type}): {param_desc}\n"
|
||
|
||
descriptions.append(desc)
|
||
|
||
return "\n".join(descriptions)
|
||
|
||
def _build_system_prompt(self, project_info: Dict, task_context: str = "") -> str:
|
||
"""构建系统提示词"""
|
||
tools_desc = self._get_tools_description()
|
||
prompt = self.config.system_prompt.format(tools_description=tools_desc)
|
||
|
||
if project_info:
|
||
prompt += f"\n\n## 项目信息\n"
|
||
prompt += f"- 名称: {project_info.get('name', 'unknown')}\n"
|
||
prompt += f"- 语言: {', '.join(project_info.get('languages', ['unknown']))}\n"
|
||
prompt += f"- 文件数: {project_info.get('file_count', 'unknown')}\n"
|
||
|
||
if task_context:
|
||
prompt += f"\n\n## 任务上下文\n{task_context}"
|
||
|
||
return prompt
|
||
|
||
def _parse_llm_response(self, response: str) -> AgentStep:
|
||
"""解析 LLM 响应"""
|
||
step = AgentStep(thought="")
|
||
|
||
# 提取 Thought
|
||
thought_match = re.search(r'Thought:\s*(.*?)(?=Action:|Final Answer:|$)', response, re.DOTALL)
|
||
if thought_match:
|
||
step.thought = thought_match.group(1).strip()
|
||
|
||
# 检查是否是最终答案
|
||
final_match = re.search(r'Final Answer:\s*(.*?)$', response, re.DOTALL)
|
||
if final_match:
|
||
step.is_final = True
|
||
try:
|
||
# 尝试提取 JSON
|
||
answer_text = final_match.group(1).strip()
|
||
# 移除 markdown 代码块
|
||
answer_text = re.sub(r'```json\s*', '', answer_text)
|
||
answer_text = re.sub(r'```\s*', '', answer_text)
|
||
step.final_answer = json.loads(answer_text)
|
||
except json.JSONDecodeError:
|
||
step.final_answer = {"raw_answer": final_match.group(1).strip()}
|
||
return step
|
||
|
||
# 提取 Action
|
||
action_match = re.search(r'Action:\s*(\w+)', response)
|
||
if action_match:
|
||
step.action = action_match.group(1).strip()
|
||
|
||
# 提取 Action Input
|
||
input_match = re.search(r'Action Input:\s*(.*?)(?=Thought:|Action:|Observation:|$)', response, re.DOTALL)
|
||
if input_match:
|
||
input_text = input_match.group(1).strip()
|
||
# 移除 markdown 代码块
|
||
input_text = re.sub(r'```json\s*', '', input_text)
|
||
input_text = re.sub(r'```\s*', '', input_text)
|
||
try:
|
||
step.action_input = json.loads(input_text)
|
||
except json.JSONDecodeError:
|
||
# 尝试简单解析
|
||
step.action_input = {"raw_input": input_text}
|
||
|
||
return step
|
||
|
||
async def _execute_tool(self, tool_name: str, tool_input: Dict) -> str:
|
||
"""执行工具"""
|
||
tool = self.tools.get(tool_name)
|
||
|
||
if not tool:
|
||
return f"错误: 工具 '{tool_name}' 不存在。可用工具: {list(self.tools.keys())}"
|
||
|
||
try:
|
||
self._tool_calls += 1
|
||
await self.emit_tool_call(tool_name, tool_input)
|
||
|
||
import time
|
||
start = time.time()
|
||
|
||
result = await tool.execute(**tool_input)
|
||
|
||
duration_ms = int((time.time() - start) * 1000)
|
||
await self.emit_tool_result(tool_name, str(result.data)[:200], duration_ms)
|
||
|
||
if result.success:
|
||
# 截断过长的输出
|
||
output = str(result.data)
|
||
if len(output) > 4000:
|
||
output = output[:4000] + "\n\n... [输出已截断，共 {} 字符]".format(len(str(result.data)))
|
||
return output
|
||
else:
|
||
return f"工具执行失败: {result.error}"
|
||
|
||
except Exception as e:
|
||
logger.error(f"Tool execution error: {e}")
|
||
return f"工具执行错误: {str(e)}"
|
||
|
||
async def run(self, input_data: Dict[str, Any]) -> AgentResult:
|
||
"""
|
||
执行 ReAct Agent
|
||
|
||
LLM 全程参与，自主决策！
|
||
"""
|
||
import time
|
||
start_time = time.time()
|
||
|
||
project_info = input_data.get("project_info", {})
|
||
task_context = input_data.get("task_context", "")
|
||
config = input_data.get("config", {})
|
||
|
||
# 构建系统提示词
|
||
system_prompt = self._build_system_prompt(project_info, task_context)
|
||
|
||
# 初始化对话历史
|
||
self._conversation_history = [
|
||
{"role": "system", "content": system_prompt},
|
||
{"role": "user", "content": "请开始对项目进行安全审计。首先了解项目结构，然后系统性地搜索和分析潜在的安全漏洞。"},
|
||
]
|
||
|
||
self._steps = []
|
||
all_findings = []
|
||
|
||
await self.emit_thinking("🤖 ReAct Agent 启动，LLM 开始自主分析...")
|
||
|
||
try:
|
||
for iteration in range(self.config.max_iterations):
|
||
if self.is_cancelled:
|
||
break
|
||
|
||
self._iteration = iteration + 1
|
||
|
||
await self.emit_thinking(f"💭 第 {iteration + 1} 轮思考...")
|
||
|
||
# 🔥 调用 LLM 进行思考和决策
|
||
response = await self.llm_service.chat_completion_raw(
|
||
messages=self._conversation_history,
|
||
temperature=0.1,
|
||
max_tokens=2048,
|
||
)
|
||
|
||
llm_output = response.get("content", "")
|
||
self._total_tokens += response.get("usage", {}).get("total_tokens", 0)
|
||
|
||
# 发射思考事件
|
||
await self.emit_event("thinking", f"LLM: {llm_output[:500]}...")
|
||
|
||
# 解析 LLM 响应
|
||
step = self._parse_llm_response(llm_output)
|
||
self._steps.append(step)
|
||
|
||
# 添加 LLM 响应到历史
|
||
self._conversation_history.append({
|
||
"role": "assistant",
|
||
"content": llm_output,
|
||
})
|
||
|
||
# 检查是否完成
|
||
if step.is_final:
|
||
await self.emit_thinking("✅ LLM 完成分析，生成最终报告")
|
||
|
||
if step.final_answer and "findings" in step.final_answer:
|
||
all_findings = step.final_answer["findings"]
|
||
break
|
||
|
||
# 执行工具
|
||
if step.action:
|
||
await self.emit_thinking(f"🔧 LLM 决定调用工具: {step.action}")
|
||
|
||
observation = await self._execute_tool(
|
||
step.action,
|
||
step.action_input or {}
|
||
)
|
||
|
||
step.observation = observation
|
||
|
||
# 添加观察结果到历史
|
||
self._conversation_history.append({
|
||
"role": "user",
|
||
"content": f"Observation: {observation}",
|
||
})
|
||
else:
|
||
# LLM 没有选择工具，提示它继续
|
||
self._conversation_history.append({
|
||
"role": "user",
|
||
"content": "请继续分析，选择一个工具执行，或者如果分析完成，输出 Final Answer。",
|
||
})
|
||
|
||
duration_ms = int((time.time() - start_time) * 1000)
|
||
|
||
await self.emit_event(
|
||
"info",
|
||
f"🎯 ReAct Agent 完成: {len(all_findings)} 个发现, {self._iteration} 轮迭代, {self._tool_calls} 次工具调用"
|
||
)
|
||
|
||
return AgentResult(
|
||
success=True,
|
||
data={
|
||
"findings": all_findings,
|
||
"steps": [
|
||
{
|
||
"thought": s.thought,
|
||
"action": s.action,
|
||
"action_input": s.action_input,
|
||
"observation": s.observation[:500] if s.observation else None,
|
||
}
|
||
for s in self._steps
|
||
],
|
||
},
|
||
iterations=self._iteration,
|
||
tool_calls=self._tool_calls,
|
||
tokens_used=self._total_tokens,
|
||
duration_ms=duration_ms,
|
||
)
|
||
|
||
except Exception as e:
|
||
logger.error(f"ReAct Agent failed: {e}", exc_info=True)
|
||
return AgentResult(success=False, error=str(e))
|
||
|
||
def get_conversation_history(self) -> List[Dict[str, str]]:
|
||
"""获取对话历史"""
|
||
return self._conversation_history
|
||
|
||
def get_steps(self) -> List[AgentStep]:
|
||
"""获取执行步骤"""
|
||
return self._steps
|