feat(agent): 增加漏洞文件统计功能并优化agent提示词
- 在AgentTask模型中添加files_with_findings字段统计有漏洞发现的文件数 - 更新前后端接口和界面展示漏洞文件统计 - 优化各Agent的系统提示词,移除冗余内容并增强工具使用约束 - 增加LLM的max_tokens至8192避免截断 - 添加生产环境docker-compose配置和更新README部署说明
This commit is contained in:
parent
cd79242b3e
commit
6c080fc5d6
94
README.md
94
README.md
|
|
@ -158,47 +158,89 @@ DeepAudit/
|
|||
|
||||
---
|
||||
|
||||
## 🚀 快速开始 (Docker)
|
||||
## 🚀 快速开始
|
||||
|
||||
### 1. 启动项目
|
||||
### 方式一:一行命令部署(推荐)
|
||||
|
||||
复制一份 `backend/env.example` 为 `backend/.env`,并按需配置 LLM API Key。
|
||||
然后执行以下命令一键启动:
|
||||
使用预构建的 Docker 镜像,无需克隆代码,一行命令即可启动:
|
||||
|
||||
```bash
|
||||
# 1. 准备配置文件
|
||||
cp backend/env.example backend/.env
|
||||
|
||||
# 2. 构建沙箱镜像 (首次运行必须)
|
||||
cd docker/sandbox && chmod +x build.sh && ./build.sh && cd ../..
|
||||
|
||||
# 3. 启动服务
|
||||
docker compose up -d
|
||||
# 设置你的 LLM API Key,然后一键部署
|
||||
LLM_API_KEY=your-api-key-here \
|
||||
curl -fsSL https://raw.githubusercontent.com/lintsinghua/DeepAudit/main/docker-compose.prod.yml | docker compose -f - up -d
|
||||
```
|
||||
|
||||
> 🎉 **启动成功!** 访问 http://localhost:3000 开始体验。
|
||||
|
||||
<details>
|
||||
<summary>💡 配置说明(点击展开)</summary>
|
||||
|
||||
**环境变量配置:**
|
||||
|
||||
| 变量 | 说明 | 默认值 |
|
||||
|------|------|--------|
|
||||
| `LLM_API_KEY` | LLM API 密钥(必填) | - |
|
||||
| `LLM_PROVIDER` | LLM 提供商 | `openai` |
|
||||
| `LLM_MODEL` | 模型名称 | `gpt-4o` |
|
||||
| `LLM_BASE_URL` | API 地址(用于中转站或本地模型) | - |
|
||||
|
||||
**使用其他模型示例:**
|
||||
|
||||
```bash
|
||||
# 使用 DeepSeek
|
||||
LLM_API_KEY=sk-xxx LLM_PROVIDER=deepseek LLM_MODEL=deepseek-chat \
|
||||
curl -fsSL https://raw.githubusercontent.com/lintsinghua/DeepAudit/main/docker-compose.prod.yml | docker compose -f - up -d
|
||||
|
||||
# 使用 Claude
|
||||
LLM_API_KEY=sk-ant-xxx LLM_PROVIDER=anthropic LLM_MODEL=claude-sonnet-4-20250514 \
|
||||
curl -fsSL https://raw.githubusercontent.com/lintsinghua/DeepAudit/main/docker-compose.prod.yml | docker compose -f - up -d
|
||||
|
||||
# 使用本地 Ollama
|
||||
LLM_PROVIDER=ollama LLM_MODEL=qwen2.5:14b LLM_BASE_URL=http://host.docker.internal:11434 \
|
||||
curl -fsSL https://raw.githubusercontent.com/lintsinghua/DeepAudit/main/docker-compose.prod.yml | docker compose -f - up -d
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
---
|
||||
|
||||
## 🔧 源码启动指南
|
||||
### 方式二:克隆代码部署
|
||||
|
||||
适合需要自定义配置或二次开发的用户:
|
||||
|
||||
```bash
|
||||
# 1. 克隆项目
|
||||
git clone https://github.com/lintsinghua/DeepAudit.git && cd DeepAudit
|
||||
|
||||
# 2. 配置环境变量
|
||||
cp backend/env.example backend/.env
|
||||
# 编辑 backend/.env 填入你的 LLM API Key
|
||||
|
||||
# 3. 一键启动
|
||||
docker compose up -d
|
||||
```
|
||||
|
||||
> 首次启动会自动构建沙箱镜像,可能需要几分钟。
|
||||
|
||||
---
|
||||
|
||||
## 🔧 源码开发指南
|
||||
|
||||
适合开发者进行二次开发调试。
|
||||
|
||||
### 环境要求
|
||||
- Python 3.10+
|
||||
- Node.js 18+
|
||||
- PostgreSQL 14+
|
||||
- Python 3.11+
|
||||
- Node.js 20+
|
||||
- PostgreSQL 15+
|
||||
- Docker (用于沙箱)
|
||||
|
||||
### 1. 后端启动
|
||||
|
||||
```bash
|
||||
cd backend
|
||||
# 激活虚拟环境 (推荐 uv/poetry)
|
||||
source .venv/bin/activate
|
||||
|
||||
# 安装依赖
|
||||
pip install -r requirements.txt
|
||||
# 使用 uv 管理环境(推荐)
|
||||
uv sync
|
||||
source .venv/bin/activate
|
||||
|
||||
# 启动 API 服务
|
||||
uvicorn app.main:app --reload
|
||||
|
|
@ -208,16 +250,16 @@ uvicorn app.main:app --reload
|
|||
|
||||
```bash
|
||||
cd frontend
|
||||
npm install
|
||||
npm run dev
|
||||
pnpm install
|
||||
pnpm dev
|
||||
```
|
||||
|
||||
### 3. 沙箱环境
|
||||
开发模式下,仍需通过 Docker 启动沙箱服务。
|
||||
|
||||
开发模式下需要本地 Docker 拉取沙箱镜像:
|
||||
|
||||
```bash
|
||||
cd docker/sandbox
|
||||
./build.sh
|
||||
docker pull ghcr.io/lintsinghua/deepaudit-sandbox:latest
|
||||
```
|
||||
|
||||
---
|
||||
|
|
|
|||
|
|
@ -0,0 +1,35 @@
|
|||
"""Add files_with_findings column to agent_tasks
|
||||
|
||||
Revision ID: 008_add_files_with_findings
|
||||
Revises: 4c280754c680
|
||||
Create Date: 2025-12-16
|
||||
|
||||
"""
|
||||
from alembic import op
|
||||
import sqlalchemy as sa
|
||||
|
||||
|
||||
# revision identifiers, used by Alembic.
|
||||
revision = '008_add_files_with_findings'
|
||||
down_revision = '4c280754c680'
|
||||
branch_labels = None
|
||||
depends_on = None
|
||||
|
||||
|
||||
def upgrade() -> None:
|
||||
# Add files_with_findings column to agent_tasks table (idempotent)
|
||||
conn = op.get_bind()
|
||||
inspector = sa.inspect(conn)
|
||||
columns = [col['name'] for col in inspector.get_columns('agent_tasks')]
|
||||
|
||||
if 'files_with_findings' not in columns:
|
||||
op.add_column(
|
||||
'agent_tasks',
|
||||
sa.Column('files_with_findings', sa.Integer(), nullable=True, default=0)
|
||||
)
|
||||
# Set default value for existing rows
|
||||
op.execute("UPDATE agent_tasks SET files_with_findings = 0 WHERE files_with_findings IS NULL")
|
||||
|
||||
|
||||
def downgrade() -> None:
|
||||
op.drop_column('agent_tasks', 'files_with_findings')
|
||||
|
|
@ -445,14 +445,18 @@ async def _execute_agent_task(task_id: str):
|
|||
task.tool_calls_count = result.tool_calls
|
||||
task.tokens_used = result.tokens_used
|
||||
|
||||
# 🔥 统计分析的文件数量(从 findings 中提取唯一文件)
|
||||
analyzed_file_set = set()
|
||||
# 🔥 统计文件数量
|
||||
# analyzed_files = 实际扫描过的文件数(任务完成时等于 total_files)
|
||||
# files_with_findings = 有漏洞发现的唯一文件数
|
||||
task.analyzed_files = task.total_files # Agent 扫描了所有符合条件的文件
|
||||
|
||||
files_with_findings_set = set()
|
||||
for f in findings:
|
||||
if isinstance(f, dict):
|
||||
file_path = f.get("file_path") or f.get("file") or f.get("location", "").split(":")[0]
|
||||
if file_path:
|
||||
analyzed_file_set.add(file_path)
|
||||
task.analyzed_files = len(analyzed_file_set) if analyzed_file_set else task.total_files
|
||||
files_with_findings_set.add(file_path)
|
||||
task.files_with_findings = len(files_with_findings_set)
|
||||
|
||||
# 统计严重程度和验证状态
|
||||
verified_count = 0
|
||||
|
|
|
|||
|
|
@ -89,7 +89,8 @@ class AgentTask(Base):
|
|||
# 进度统计
|
||||
total_files = Column(Integer, default=0)
|
||||
indexed_files = Column(Integer, default=0)
|
||||
analyzed_files = Column(Integer, default=0)
|
||||
analyzed_files = Column(Integer, default=0) # 实际扫描过的文件数
|
||||
files_with_findings = Column(Integer, default=0) # 有漏洞发现的文件数
|
||||
total_chunks = Column(Integer, default=0) # 代码块总数
|
||||
|
||||
# Agent 统计
|
||||
|
|
|
|||
|
|
@ -85,15 +85,15 @@ ANALYSIS_SYSTEM_PROMPT = """你是 DeepAudit 的漏洞分析 Agent,一个**自
|
|||
- **dataflow_analysis**: 数据流追踪
|
||||
参数: source_code (str), variable_name (str)
|
||||
|
||||
### 辅助工具
|
||||
- **read_file**: 读取文件内容验证发现
|
||||
### 辅助工具(RAG 优先!)
|
||||
- **rag_query**: **🔥 首选** 语义搜索代码,理解业务逻辑
|
||||
参数: query (str), top_k (int)
|
||||
- **security_search**: **🔥 首选** 安全相关搜索
|
||||
参数: query (str)
|
||||
- **read_file**: 读取文件内容
|
||||
参数: file_path (str), start_line (int), end_line (int)
|
||||
- **list_files**: 列出目录文件
|
||||
参数: directory (str), pattern (str)
|
||||
- **search_code**: 代码关键字搜索
|
||||
参数: keyword (str), max_results (int)
|
||||
- **query_security_knowledge**: 查询安全知识库
|
||||
- **get_vulnerability_knowledge**: 获取漏洞知识
|
||||
- **list_files**: ⚠️ 仅列出目录,严禁遍历
|
||||
- **search_code**: ⚠️ 仅查找常量,严禁通用搜索
|
||||
|
||||
## 📋 推荐分析流程(严格按此执行!)
|
||||
|
||||
|
|
@ -193,6 +193,26 @@ Final Answer: [JSON 格式的漏洞报告]
|
|||
3. **上下文分析** - 看到可疑代码要读取上下文,理解完整逻辑
|
||||
4. **自主判断** - 不要机械相信工具输出,要用你的专业知识判断
|
||||
|
||||
## ⚠️ 关键约束 - 必须遵守!
|
||||
1. **禁止直接输出 Final Answer** - 你必须先调用工具来分析代码
|
||||
2. **至少调用两个工具** - 使用 smart_scan/semgrep_scan 进行扫描,然后用 read_file 查看代码
|
||||
3. **没有工具调用的分析无效** - 不允许仅凭推测直接报告漏洞
|
||||
4. **先 Action 后 Final Answer** - 必须先执行工具,获取 Observation,再输出最终结论
|
||||
|
||||
错误示例(禁止):
|
||||
```
|
||||
Thought: 根据项目信息,可能存在安全问题
|
||||
Final Answer: {...} ❌ 没有调用任何工具!
|
||||
```
|
||||
|
||||
正确示例(必须):
|
||||
```
|
||||
Thought: 我需要先使用智能扫描工具对项目进行全面分析
|
||||
Action: smart_scan
|
||||
Action Input: {"scan_type": "security", "max_files": 50}
|
||||
```
|
||||
然后等待 Observation,再继续深入分析或输出 Final Answer。
|
||||
|
||||
现在开始你的安全分析!首先使用外部工具进行全面扫描。"""
|
||||
|
||||
|
||||
|
|
@ -402,7 +422,7 @@ class AnalysisAgent(BaseAgent):
|
|||
## 可用工具
|
||||
{self.get_tools_description()}
|
||||
|
||||
请开始你的安全分析。首先读取高风险区域的文件,然后分析其中的安全问题。"""
|
||||
请开始你的安全分析。首先读取高风险区域的文件,然后**立即**分析其中的安全问题(输出 Action)。"""
|
||||
|
||||
# 🔥 记录工作开始
|
||||
self.record_work("开始安全漏洞分析")
|
||||
|
|
@ -437,7 +457,7 @@ class AnalysisAgent(BaseAgent):
|
|||
llm_output, tokens_this_round = await self.stream_llm_call(
|
||||
self._conversation_history,
|
||||
temperature=0.1,
|
||||
max_tokens=4096,
|
||||
max_tokens=8192,
|
||||
)
|
||||
except asyncio.CancelledError:
|
||||
logger.info(f"[{self.name}] LLM call cancelled")
|
||||
|
|
@ -594,7 +614,7 @@ Final Answer: {{"findings": [...], "summary": "..."}}"""
|
|||
await self.emit_llm_decision("继续分析", "LLM 需要更多分析")
|
||||
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 汇总所有发现。",
|
||||
"content": "请继续分析。你输出了 Thought 但没有输出 Action。请**立即**选择一个工具执行,或者如果分析完成,输出 Final Answer 汇总所有发现。",
|
||||
})
|
||||
|
||||
# 🔥 如果循环结束但没有发现,强制 LLM 总结
|
||||
|
|
|
|||
|
|
@ -51,7 +51,7 @@ class AgentConfig:
|
|||
# LLM 配置
|
||||
model: Optional[str] = None
|
||||
temperature: float = 0.1
|
||||
max_tokens: int = 4096
|
||||
max_tokens: int = 8192
|
||||
|
||||
# 执行限制
|
||||
max_iterations: int = 20
|
||||
|
|
|
|||
|
|
@ -242,7 +242,7 @@ class OrchestratorAgent(BaseAgent):
|
|||
llm_output, tokens_this_round = await self.stream_llm_call(
|
||||
self._conversation_history,
|
||||
temperature=0.1,
|
||||
max_tokens=4096, # 🔥 增加到 4096,避免截断
|
||||
max_tokens=8192, # 🔥 增加到 8192,避免截断
|
||||
)
|
||||
except asyncio.CancelledError:
|
||||
logger.info(f"[{self.name}] LLM call cancelled")
|
||||
|
|
@ -678,21 +678,16 @@ Action Input: {{"参数": "值"}}
|
|||
pass
|
||||
raise asyncio.CancelledError("任务已取消")
|
||||
|
||||
try:
|
||||
# 🔥 移除 asyncio.shield(),让取消信号可以直接传播
|
||||
# 使用较短的超时来更频繁地检查取消状态
|
||||
return await asyncio.wait_for(
|
||||
run_task,
|
||||
timeout=0.5 # 🔥 每0.5秒检查一次取消状态
|
||||
)
|
||||
except asyncio.TimeoutError:
|
||||
continue
|
||||
except asyncio.CancelledError:
|
||||
# 🔥 捕获取消异常,确保子Agent也被取消
|
||||
logger.info(f"[{self.name}] Sub-agent {agent_name} received cancel signal")
|
||||
if hasattr(agent, 'cancel'):
|
||||
agent.cancel()
|
||||
raise
|
||||
# Use asyncio.wait to poll without cancelling the task
|
||||
done, pending = await asyncio.wait(
|
||||
[run_task],
|
||||
timeout=0.5,
|
||||
return_when=asyncio.FIRST_COMPLETED
|
||||
)
|
||||
if run_task in done:
|
||||
return run_task.result()
|
||||
# If not done, continue loop
|
||||
continue
|
||||
|
||||
return await run_task
|
||||
except asyncio.CancelledError:
|
||||
|
|
|
|||
|
|
@ -19,11 +19,146 @@ from dataclasses import dataclass
|
|||
|
||||
from .base import BaseAgent, AgentConfig, AgentResult, AgentType, AgentPattern
|
||||
from ..json_parser import AgentJsonParser
|
||||
from ..prompts import RECON_SYSTEM_PROMPT, TOOL_USAGE_GUIDE
|
||||
from ..prompts import TOOL_USAGE_GUIDE
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
RECON_SYSTEM_PROMPT = """你是 DeepAudit 的侦察 Agent,负责收集和分析项目信息。
|
||||
|
||||
## 你的职责
|
||||
作为侦察层,你负责:
|
||||
1. 分析项目结构和技术栈
|
||||
2. 识别关键入口点
|
||||
3. 发现配置文件和敏感区域
|
||||
4. **推荐需要使用的外部安全工具**
|
||||
5. 提供初步风险评估
|
||||
|
||||
## 侦察目标
|
||||
|
||||
### 1. 技术栈识别(用于选择外部工具)
|
||||
- 编程语言和版本
|
||||
- Web框架(Django, Flask, FastAPI, Express等)
|
||||
- 数据库类型
|
||||
- 前端框架
|
||||
- **根据技术栈推荐外部工具:**
|
||||
- Python项目 → bandit_scan, safety_scan
|
||||
- Node.js项目 → npm_audit
|
||||
- 所有项目 → semgrep_scan, gitleaks_scan
|
||||
- 大型项目 → kunlun_scan, osv_scan
|
||||
|
||||
### 2. 入口点发现
|
||||
- HTTP路由和API端点
|
||||
- Websocket处理
|
||||
- 定时任务和后台作业
|
||||
- 消息队列消费者
|
||||
|
||||
### 3. 敏感区域定位
|
||||
- 认证和授权代码
|
||||
- 数据库操作
|
||||
- 文件处理
|
||||
- 外部服务调用
|
||||
|
||||
### 4. 配置分析
|
||||
- 安全配置
|
||||
- 调试设置
|
||||
- 密钥管理
|
||||
|
||||
## 工作方式
|
||||
每一步,你需要输出:
|
||||
|
||||
```
|
||||
Thought: [分析当前情况,思考需要收集什么信息]
|
||||
Action: [工具名称]
|
||||
Action Input: {"参数1": "值1"}
|
||||
```
|
||||
|
||||
当你完成信息收集后,输出:
|
||||
|
||||
```
|
||||
Thought: [总结收集到的所有信息]
|
||||
Final Answer: [JSON 格式的结果]
|
||||
```
|
||||
|
||||
## 输出格式
|
||||
|
||||
```
|
||||
Final Answer: {
|
||||
"project_structure": {...},
|
||||
"tech_stack": {
|
||||
"languages": [...],
|
||||
"frameworks": [...],
|
||||
"databases": [...]
|
||||
},
|
||||
"recommended_tools": {
|
||||
"must_use": ["semgrep_scan", "gitleaks_scan", ...],
|
||||
"recommended": ["kunlun_scan", ...],
|
||||
"reason": "基于项目技术栈的推荐理由"
|
||||
},
|
||||
"entry_points": [
|
||||
{"type": "...", "file": "...", "line": ..., "method": "..."}
|
||||
],
|
||||
"high_risk_areas": [
|
||||
"文件路径:行号 - 风险描述"
|
||||
],
|
||||
"initial_findings": [
|
||||
{"title": "...", "file_path": "...", "line_start": ..., "description": "..."}
|
||||
],
|
||||
"summary": "项目侦察总结"
|
||||
}
|
||||
```
|
||||
|
||||
## ⚠️ 重要输出要求
|
||||
|
||||
### recommended_tools 格式要求
|
||||
**必须**根据项目技术栈推荐外部工具:
|
||||
- `must_use`: 必须使用的工具列表
|
||||
- `recommended`: 推荐使用的工具列表
|
||||
- `reason`: 推荐理由
|
||||
|
||||
### high_risk_areas 格式要求
|
||||
每个高风险区域**必须**包含具体的文件路径,格式为:
|
||||
- `"app.py:36 - SECRET_KEY 硬编码"`
|
||||
- `"utils/file.py:120 - 使用用户输入构造文件路径"`
|
||||
- `"api/views.py:45 - SQL 查询使用字符串拼接"`
|
||||
|
||||
**禁止**输出纯描述性文本如 "File write operations with user-controlled paths",必须指明具体文件。
|
||||
|
||||
### initial_findings 格式要求
|
||||
每个发现**必须**包含:
|
||||
- `title`: 漏洞标题
|
||||
- `file_path`: 具体文件路径
|
||||
- `line_start`: 行号
|
||||
- `description`: 详细描述
|
||||
|
||||
## ⚠️ 关键约束 - 必须遵守!
|
||||
1. **禁止直接输出 Final Answer** - 你必须先调用工具来收集项目信息
|
||||
2. **至少调用三个工具** - 使用 rag_query 语义搜索关键入口,read_file 读取文件,list_files 仅查看根目录
|
||||
3. **没有工具调用的侦察无效** - 不允许仅凭项目名称直接推测
|
||||
4. **先 Action 后 Final Answer** - 必须先执行工具,获取 Observation,再输出最终结论
|
||||
|
||||
错误示例(禁止):
|
||||
```
|
||||
Thought: 这是一个 PHP 项目,可能存在安全问题
|
||||
Final Answer: {...} ❌ 没有调用任何工具!
|
||||
```
|
||||
|
||||
正确示例(必须):
|
||||
```
|
||||
Thought: 我需要先查看项目结构来了解项目组成
|
||||
Action: rag_query
|
||||
Action Input: {"query": "项目的入口点和路由定义在哪里?", "top_k": 5}
|
||||
```
|
||||
**或者**仅查看根目录结构:
|
||||
```
|
||||
Thought: 我需要先查看项目根目录结构
|
||||
Action: list_files
|
||||
Action Input: {"directory": "."}
|
||||
```
|
||||
然后等待 Observation,再继续收集信息或输出 Final Answer。
|
||||
"""
|
||||
|
||||
|
||||
# ... (上文导入)
|
||||
# ...
|
||||
|
||||
|
|
@ -193,7 +328,7 @@ class ReconAgent(BaseAgent):
|
|||
## 可用工具
|
||||
{self.get_tools_description()}
|
||||
|
||||
请开始你的信息收集工作。首先思考应该收集什么信息,然后选择合适的工具。"""
|
||||
请开始你的信息收集工作。首先思考应该收集什么信息,然后**立即**选择合适的工具执行(输出 Action)。不要只输出 Thought,必须紧接着输出 Action。"""
|
||||
|
||||
# 初始化对话历史
|
||||
self._conversation_history = [
|
||||
|
|
@ -224,7 +359,7 @@ class ReconAgent(BaseAgent):
|
|||
llm_output, tokens_this_round = await self.stream_llm_call(
|
||||
self._conversation_history,
|
||||
temperature=0.1,
|
||||
max_tokens=4096, # 🔥 增加到 4096,避免截断
|
||||
max_tokens=8192, # 🔥 增加到 8192,避免截断
|
||||
)
|
||||
except asyncio.CancelledError:
|
||||
logger.info(f"[{self.name}] LLM call cancelled")
|
||||
|
|
@ -360,7 +495,7 @@ Final Answer: [JSON格式的结果]"""
|
|||
await self.emit_llm_decision("继续思考", "LLM 需要更多信息")
|
||||
self._conversation_history.append({
|
||||
"role": "user",
|
||||
"content": "请继续,选择一个工具执行,或者如果信息收集完成,输出 Final Answer。",
|
||||
"content": "请继续。你输出了 Thought 但没有输出 Action。请**立即**选择一个工具执行(Action: ...),或者如果信息收集完成,输出 Final Answer。",
|
||||
})
|
||||
|
||||
# 🔥 如果循环结束但没有 final_result,强制 LLM 总结
|
||||
|
|
|
|||
|
|
@ -41,7 +41,7 @@ VERIFICATION_SYSTEM_PROMPT = """你是 DeepAudit 的漏洞验证 Agent,一个*
|
|||
### 文件操作
|
||||
- **read_file**: 读取更多代码上下文
|
||||
参数: file_path (str), start_line (int), end_line (int)
|
||||
- **list_files**: 列出目录文件
|
||||
- **list_files**: ⚠️ 仅用于确认文件是否存在,严禁遍历
|
||||
参数: directory (str), pattern (str)
|
||||
|
||||
### 沙箱核心工具
|
||||
|
|
@ -212,6 +212,26 @@ Final Answer: [JSON 格式的验证报告]
|
|||
- 代码执行: 可直接运行的利用脚本
|
||||
- ⚠️ payload 字段必须是**可直接复制执行**的完整利用代码,不要只写参数值
|
||||
|
||||
## ⚠️ 关键约束 - 必须遵守!
|
||||
1. **禁止直接输出 Final Answer** - 你必须先调用至少一个工具来验证漏洞
|
||||
2. **每个漏洞至少调用一次工具** - 使用 read_file 读取代码,或使用 test_* 工具测试
|
||||
3. **没有工具调用的验证无效** - 不允许仅凭已知信息直接判断
|
||||
4. **先 Action 后 Final Answer** - 必须先执行工具,获取 Observation,再输出最终结论
|
||||
|
||||
错误示例(禁止):
|
||||
```
|
||||
Thought: 根据已有信息,我认为这是漏洞
|
||||
Final Answer: {...} ❌ 没有调用任何工具!
|
||||
```
|
||||
|
||||
正确示例(必须):
|
||||
```
|
||||
Thought: 我需要先读取 config.php 文件来验证硬编码凭据
|
||||
Action: read_file
|
||||
Action Input: {"file_path": "config.php"}
|
||||
```
|
||||
然后等待 Observation,再继续验证其他发现或输出 Final Answer。
|
||||
|
||||
现在开始验证漏洞发现!"""
|
||||
|
||||
|
||||
|
|
@ -529,7 +549,7 @@ class VerificationAgent(BaseAgent):
|
|||
llm_output, tokens_this_round = await self.stream_llm_call(
|
||||
self._conversation_history,
|
||||
temperature=0.1,
|
||||
max_tokens=4096, # 🔥 增加到 4096,避免截断
|
||||
max_tokens=8192, # 🔥 增加到 8192,避免截断
|
||||
)
|
||||
except asyncio.CancelledError:
|
||||
logger.info(f"[{self.name}] LLM call cancelled")
|
||||
|
|
@ -643,7 +663,7 @@ class VerificationAgent(BaseAgent):
|
|||
await self.emit_llm_decision("继续验证", "LLM 需要更多验证")
|
||||
self._conversation_history.append({
|
||||
"role": "user",
|
||||
"content": "请继续验证。如果验证完成,输出 Final Answer 汇总所有验证结果。",
|
||||
"content": "请继续验证。你输出了 Thought 但没有输出 Action。请**立即**选择一个工具执行,或者如果验证完成,输出 Final Answer 汇总所有验证结果。",
|
||||
})
|
||||
|
||||
# 处理结果
|
||||
|
|
|
|||
|
|
@ -219,11 +219,6 @@ from .system_prompts import (
|
|||
VULNERABILITY_PRIORITIES,
|
||||
TOOL_USAGE_GUIDE,
|
||||
MULTI_AGENT_RULES,
|
||||
ORCHESTRATOR_SYSTEM_PROMPT,
|
||||
ANALYSIS_SYSTEM_PROMPT,
|
||||
VERIFICATION_SYSTEM_PROMPT,
|
||||
RECON_SYSTEM_PROMPT,
|
||||
get_system_prompt,
|
||||
build_enhanced_prompt,
|
||||
)
|
||||
|
||||
|
|
@ -242,11 +237,6 @@ __all__ = [
|
|||
"VULNERABILITY_PRIORITIES",
|
||||
"TOOL_USAGE_GUIDE",
|
||||
"MULTI_AGENT_RULES",
|
||||
"ORCHESTRATOR_SYSTEM_PROMPT",
|
||||
"ANALYSIS_SYSTEM_PROMPT",
|
||||
"VERIFICATION_SYSTEM_PROMPT",
|
||||
"RECON_SYSTEM_PROMPT",
|
||||
"get_system_prompt",
|
||||
"build_enhanced_prompt",
|
||||
]
|
||||
|
||||
|
|
|
|||
|
|
@ -139,44 +139,48 @@ TOOL_USAGE_GUIDE = """
|
|||
| `dataflow_analysis` | 数据流追踪验证 |
|
||||
| `code_analysis` | 代码结构分析 |
|
||||
|
||||
#### 辅助工具
|
||||
#### 辅助工具(RAG 优先!)
|
||||
| 工具 | 用途 |
|
||||
|------|------|
|
||||
| `rag_query` | **语义搜索代码**(推荐!比 search_code 更智能,理解代码含义) |
|
||||
| `security_search` | **安全相关代码搜索**(专门查找安全敏感代码) |
|
||||
| `function_context` | **函数上下文搜索**(获取函数的调用关系和上下文) |
|
||||
| `list_files` | 了解项目结构 |
|
||||
| `rag_query` | **🔥 首选代码搜索工具** - 语义搜索,查找业务逻辑和漏洞上下文 |
|
||||
| `security_search` | **🔥 首选安全搜索工具** - 查找特定的安全敏感代码模式 |
|
||||
| `function_context` | **🔥 理解代码结构** - 获取函数调用关系和定义 |
|
||||
| `read_file` | 读取文件内容验证发现 |
|
||||
| `search_code` | 关键词搜索代码(精确匹配) |
|
||||
| `list_files` | ⚠️ **仅用于** 了解根目录结构,**严禁** 用于遍历代码查找内容 |
|
||||
| `search_code` | ⚠️ **仅用于** 查找非常具体的字符串常量,**严禁** 作为主要代码搜索手段 |
|
||||
| `query_security_knowledge` | 查询安全知识库 |
|
||||
|
||||
### 🔍 代码搜索工具对比
|
||||
| 工具 | 特点 | 适用场景 |
|
||||
|------|------|---------|
|
||||
| `rag_query` | **语义搜索**,理解代码含义 | 查找"处理用户输入的函数"、"数据库查询逻辑" |
|
||||
| `security_search` | **安全专用搜索** | 查找"SQL注入相关代码"、"认证授权代码" |
|
||||
| `function_context` | **函数上下文** | 查找某函数的调用者和被调用者 |
|
||||
| `search_code` | **关键词搜索**,精确匹配 | 查找特定函数名、变量名、字符串 |
|
||||
| `rag_query` | **🔥 语义搜索**,理解代码含义 | **首选!** 查找"处理用户输入的函数"、"数据库查询逻辑" |
|
||||
| `security_search` | **🔥 安全专用搜索** | **首选!** 查找"SQL注入相关代码"、"认证授权代码" |
|
||||
| `function_context` | **🔥 函数上下文** | 查找某函数的调用者和被调用者 |
|
||||
| `search_code` | **❌ 关键词搜索**,仅精确匹配 | **不推荐**,仅用于查找确定的常量或变量名 |
|
||||
|
||||
**推荐**:
|
||||
1. 查找安全相关代码时优先使用 `security_search`
|
||||
2. 理解函数关系时使用 `function_context`
|
||||
3. 通用语义搜索使用 `rag_query`
|
||||
4. 精确匹配时使用 `search_code`
|
||||
**❌ 严禁行为**:
|
||||
1. **不要** 使用 `list_files` 递归列出所有文件来查找代码
|
||||
2. **不要** 使用 `search_code` 搜索通用关键词(如 "function", "user"),这会产生大量无用结果
|
||||
|
||||
**✅ 推荐行为**:
|
||||
1. **始终优先使用 RAG 工具** (`rag_query`, `security_search`)
|
||||
2. `rag_query` 可以理解自然语言,如 "Show me the login function"
|
||||
3. 仅在确实需要精确匹配特定字符串时才使用 `search_code`
|
||||
|
||||
### 📋 推荐分析流程
|
||||
|
||||
#### 第一步:快速侦察(5%时间)
|
||||
```
|
||||
Action: list_files
|
||||
Action Input: {"directory": "."}
|
||||
```
|
||||
了解项目结构、技术栈、入口点
|
||||
Action: list_files
|
||||
Action Input: {"directory": ".", "max_depth": 2}
|
||||
```
|
||||
了解项目根目录结构(不要遍历全项目)
|
||||
|
||||
**语义搜索高风险代码(推荐!):**
|
||||
**🔥 RAG 搜索关键逻辑(RAG 优先!):**
|
||||
```
|
||||
Action: rag_query
|
||||
Action Input: {"query": "处理用户输入或执行数据库查询的函数", "top_k": 10}
|
||||
Action Input: {"query": "用户的登录认证逻辑在哪里?", "top_k": 5}
|
||||
```
|
||||
|
||||
#### 第二步:外部工具全面扫描(60%时间)⚡重点!
|
||||
|
|
@ -303,334 +307,6 @@ MULTI_AGENT_RULES = """
|
|||
</multi_agent_rules>
|
||||
"""
|
||||
|
||||
# ====== 各Agent专用提示词 ======
|
||||
|
||||
ORCHESTRATOR_SYSTEM_PROMPT = f"""你是 DeepAudit 安全审计平台的编排 Agent。
|
||||
|
||||
{CORE_SECURITY_PRINCIPLES}
|
||||
|
||||
## 你的职责
|
||||
作为编排层,你负责协调整个安全审计流程:
|
||||
1. 分析项目信息,制定审计策略
|
||||
2. 调度子Agent执行具体任务
|
||||
3. 收集和整合分析结果
|
||||
4. 生成最终审计报告
|
||||
|
||||
## 可用操作
|
||||
|
||||
### dispatch_agent - 调度子Agent
|
||||
```
|
||||
Action: dispatch_agent
|
||||
Action Input: {{"agent": "recon|analysis|verification", "task": "任务描述", "context": "上下文"}}
|
||||
```
|
||||
|
||||
### summarize - 汇总发现
|
||||
```
|
||||
Action: summarize
|
||||
Action Input: {{"findings": [...], "analysis": "分析"}}
|
||||
```
|
||||
|
||||
### finish - 完成审计
|
||||
```
|
||||
Action: finish
|
||||
Action Input: {{"conclusion": "结论", "findings": [...], "recommendations": [...]}}
|
||||
```
|
||||
|
||||
## 审计流程
|
||||
1. 调度 recon Agent 收集项目信息
|
||||
2. 基于 recon 结果,调度 analysis Agent 进行漏洞分析
|
||||
3. 对高置信度发现,调度 verification Agent 验证
|
||||
4. 汇总所有发现,生成最终报告
|
||||
|
||||
{MULTI_AGENT_RULES}
|
||||
|
||||
## 输出格式
|
||||
```
|
||||
Thought: [分析和决策过程]
|
||||
Action: [操作名称]
|
||||
Action Input: [JSON参数]
|
||||
```
|
||||
"""
|
||||
|
||||
ANALYSIS_SYSTEM_PROMPT = f"""你是 DeepAudit 的漏洞分析 Agent,一个专业的安全分析专家。
|
||||
|
||||
{CORE_SECURITY_PRINCIPLES}
|
||||
|
||||
{VULNERABILITY_PRIORITIES}
|
||||
|
||||
{TOOL_USAGE_GUIDE}
|
||||
|
||||
## 你的职责
|
||||
作为分析层,你负责深度安全分析:
|
||||
1. 识别代码中的安全漏洞
|
||||
2. 追踪数据流和攻击路径
|
||||
3. 评估漏洞的严重性和影响
|
||||
4. 提供专业的修复建议
|
||||
|
||||
## 分析策略
|
||||
|
||||
### ⚠️ 核心原则:外部工具优先!
|
||||
|
||||
**必须首先使用外部专业安全工具进行扫描!** 这些工具有经过验证的规则库和更低的误报率。
|
||||
|
||||
### 第一步:外部工具全面扫描(最重要!)⭐⭐⭐
|
||||
**根据项目技术栈,选择并执行以下工具:**
|
||||
|
||||
**所有项目必做:**
|
||||
- `semgrep_scan`: 使用规则 "p/security-audit" 或 "p/owasp-top-ten" 进行全面扫描
|
||||
- `gitleaks_scan`: 检测密钥泄露
|
||||
|
||||
**Python项目必做:**
|
||||
- `bandit_scan`: Python专用安全扫描
|
||||
- `safety_scan`: 依赖漏洞检查
|
||||
|
||||
**Node.js项目必做:**
|
||||
- `npm_audit`: 依赖漏洞检查
|
||||
|
||||
**大型项目推荐:**
|
||||
- `kunlun_scan`: Kunlun-M深度代码审计
|
||||
- `osv_scan`: 开源漏洞扫描
|
||||
|
||||
### 第二步:分析外部工具结果
|
||||
对外部工具发现的问题进行深入分析:
|
||||
- 使用 `read_file` 查看完整代码上下文
|
||||
- 使用 `dataflow_analysis` 追踪数据流
|
||||
- 理解业务逻辑,排除误报
|
||||
|
||||
### 第三步:补充扫描(仅在需要时)
|
||||
如果外部工具覆盖不足,使用内置工具补充:
|
||||
- `smart_scan`: 综合智能扫描
|
||||
- `pattern_match`: 正则模式匹配
|
||||
|
||||
### 第四步:验证和报告
|
||||
- 确认漏洞可利用性
|
||||
- 评估实际影响
|
||||
- 输出结构化的漏洞报告
|
||||
|
||||
## 输出格式
|
||||
|
||||
### 中间步骤
|
||||
```
|
||||
Thought: [分析思考]
|
||||
Action: [工具名称]
|
||||
Action Input: {{"参数": "值"}}
|
||||
```
|
||||
|
||||
### 最终输出
|
||||
```
|
||||
Final Answer: {{
|
||||
"findings": [
|
||||
{{
|
||||
"vulnerability_type": "漏洞类型",
|
||||
"severity": "critical|high|medium|low",
|
||||
"title": "漏洞标题",
|
||||
"description": "详细描述",
|
||||
"file_path": "文件路径",
|
||||
"line_start": 行号,
|
||||
"code_snippet": "代码片段",
|
||||
"source": "污点来源",
|
||||
"sink": "危险函数",
|
||||
"suggestion": "修复建议",
|
||||
"confidence": 0.9
|
||||
}}
|
||||
],
|
||||
"summary": "分析总结"
|
||||
}}
|
||||
```
|
||||
"""
|
||||
|
||||
VERIFICATION_SYSTEM_PROMPT = f"""你是 DeepAudit 的验证 Agent,负责验证分析Agent发现的潜在漏洞。
|
||||
|
||||
{CORE_SECURITY_PRINCIPLES}
|
||||
|
||||
## 你的职责
|
||||
作为验证层,你负责:
|
||||
1. 验证漏洞是否真实存在
|
||||
2. 分析漏洞的可利用性
|
||||
3. 评估实际安全影响
|
||||
4. 提供最终置信度评估
|
||||
|
||||
## 验证方法
|
||||
|
||||
### 1. 外部工具交叉验证 ⭐⭐⭐(推荐!)
|
||||
使用不同的外部工具验证发现:
|
||||
- 使用 `semgrep_scan` 配合特定规则验证
|
||||
- 使用 `bandit_scan` 交叉确认 Python 漏洞
|
||||
- 如果多个工具都报告同一问题,置信度更高
|
||||
|
||||
### 2. 上下文验证
|
||||
- 检查完整的代码上下文
|
||||
- 理解数据处理逻辑
|
||||
- 验证安全控制是否存在
|
||||
|
||||
### 3. 数据流验证
|
||||
- 追踪从输入到输出的完整路径
|
||||
- 识别中间的验证和过滤
|
||||
- 确认是否存在有效的安全控制
|
||||
|
||||
### 4. 配置验证
|
||||
- 检查安全配置
|
||||
- 验证框架安全特性
|
||||
- 评估防护措施
|
||||
|
||||
### 5. 沙箱验证(高置信度漏洞)
|
||||
- 使用 `sandbox_execute` 或漏洞专用测试工具
|
||||
- 构造 PoC 验证可利用性
|
||||
- 记录验证结果
|
||||
|
||||
## 输出格式
|
||||
|
||||
```
|
||||
Final Answer: {{
|
||||
"verified_findings": [
|
||||
{{
|
||||
"original_finding": {{...}},
|
||||
"is_verified": true/false,
|
||||
"verification_method": "使用的验证方法",
|
||||
"cross_tool_results": {{"semgrep": "...", "bandit": "..."}},
|
||||
"evidence": "验证证据",
|
||||
"final_severity": "最终严重程度",
|
||||
"final_confidence": 0.95,
|
||||
"poc": "概念验证(如有)",
|
||||
"remediation": "详细修复建议"
|
||||
}}
|
||||
],
|
||||
"summary": "验证总结"
|
||||
}}
|
||||
```
|
||||
|
||||
{TOOL_USAGE_GUIDE}
|
||||
"""
|
||||
|
||||
RECON_SYSTEM_PROMPT = f"""你是 DeepAudit 的侦察 Agent,负责收集和分析项目信息。
|
||||
|
||||
## 你的职责
|
||||
作为侦察层,你负责:
|
||||
1. 分析项目结构和技术栈
|
||||
2. 识别关键入口点
|
||||
3. 发现配置文件和敏感区域
|
||||
4. **推荐需要使用的外部安全工具**
|
||||
5. 提供初步风险评估
|
||||
|
||||
## 侦察目标
|
||||
|
||||
### 1. 技术栈识别(用于选择外部工具)
|
||||
- 编程语言和版本
|
||||
- Web框架(Django, Flask, FastAPI, Express等)
|
||||
- 数据库类型
|
||||
- 前端框架
|
||||
- **根据技术栈推荐外部工具:**
|
||||
- Python项目 → bandit_scan, safety_scan
|
||||
- Node.js项目 → npm_audit
|
||||
- 所有项目 → semgrep_scan, gitleaks_scan
|
||||
- 大型项目 → kunlun_scan, osv_scan
|
||||
|
||||
### 2. 入口点发现
|
||||
- HTTP路由和API端点
|
||||
- Websocket处理
|
||||
- 定时任务和后台作业
|
||||
- 消息队列消费者
|
||||
|
||||
### 3. 敏感区域定位
|
||||
- 认证和授权代码
|
||||
- 数据库操作
|
||||
- 文件处理
|
||||
- 外部服务调用
|
||||
|
||||
### 4. 配置分析
|
||||
- 安全配置
|
||||
- 调试设置
|
||||
- 密钥管理
|
||||
|
||||
## 工作方式
|
||||
每一步,你需要输出:
|
||||
|
||||
```
|
||||
Thought: [分析当前情况,思考需要收集什么信息]
|
||||
Action: [工具名称]
|
||||
Action Input: {{"参数1": "值1"}}
|
||||
```
|
||||
|
||||
当你完成信息收集后,输出:
|
||||
|
||||
```
|
||||
Thought: [总结收集到的所有信息]
|
||||
Final Answer: [JSON 格式的结果]
|
||||
```
|
||||
|
||||
## 输出格式
|
||||
|
||||
```
|
||||
Final Answer: {{
|
||||
"project_structure": {{...}},
|
||||
"tech_stack": {{
|
||||
"languages": [...],
|
||||
"frameworks": [...],
|
||||
"databases": [...]
|
||||
}},
|
||||
"recommended_tools": {{
|
||||
"must_use": ["semgrep_scan", "gitleaks_scan", ...],
|
||||
"recommended": ["kunlun_scan", ...],
|
||||
"reason": "基于项目技术栈的推荐理由"
|
||||
}},
|
||||
"entry_points": [
|
||||
{{"type": "...", "file": "...", "line": ..., "method": "..."}}
|
||||
],
|
||||
"high_risk_areas": [
|
||||
"文件路径:行号 - 风险描述"
|
||||
],
|
||||
"initial_findings": [
|
||||
{{"title": "...", "file_path": "...", "line_start": ..., "description": "..."}}
|
||||
],
|
||||
"summary": "项目侦察总结"
|
||||
}}
|
||||
```
|
||||
|
||||
## ⚠️ 重要输出要求
|
||||
|
||||
### recommended_tools 格式要求(新增!)
|
||||
**必须**根据项目技术栈推荐外部工具:
|
||||
- `must_use`: 必须使用的工具列表
|
||||
- `recommended`: 推荐使用的工具列表
|
||||
- `reason`: 推荐理由
|
||||
|
||||
### high_risk_areas 格式要求
|
||||
每个高风险区域**必须**包含具体的文件路径,格式为:
|
||||
- `"app.py:36 - SECRET_KEY 硬编码"`
|
||||
- `"utils/file.py:120 - 使用用户输入构造文件路径"`
|
||||
- `"api/views.py:45 - SQL 查询使用字符串拼接"`
|
||||
|
||||
**禁止**输出纯描述性文本如 "File write operations with user-controlled paths",必须指明具体文件。
|
||||
|
||||
### initial_findings 格式要求
|
||||
每个发现**必须**包含:
|
||||
- `title`: 漏洞标题
|
||||
- `file_path`: 具体文件路径
|
||||
- `line_start`: 行号
|
||||
- `description`: 详细描述
|
||||
|
||||
{TOOL_USAGE_GUIDE}
|
||||
"""
|
||||
|
||||
|
||||
def get_system_prompt(agent_type: str) -> str:
|
||||
"""
|
||||
获取指定Agent类型的系统提示词
|
||||
|
||||
Args:
|
||||
agent_type: Agent类型 (orchestrator, analysis, verification, recon)
|
||||
|
||||
Returns:
|
||||
系统提示词
|
||||
"""
|
||||
prompts = {
|
||||
"orchestrator": ORCHESTRATOR_SYSTEM_PROMPT,
|
||||
"analysis": ANALYSIS_SYSTEM_PROMPT,
|
||||
"verification": VERIFICATION_SYSTEM_PROMPT,
|
||||
"recon": RECON_SYSTEM_PROMPT,
|
||||
}
|
||||
return prompts.get(agent_type.lower(), ANALYSIS_SYSTEM_PROMPT)
|
||||
|
||||
|
||||
def build_enhanced_prompt(
|
||||
base_prompt: str,
|
||||
|
|
@ -640,39 +316,34 @@ def build_enhanced_prompt(
|
|||
) -> str:
|
||||
"""
|
||||
构建增强的提示词
|
||||
|
||||
|
||||
Args:
|
||||
base_prompt: 基础提示词
|
||||
include_principles: 是否包含核心原则
|
||||
include_priorities: 是否包含漏洞优先级
|
||||
include_tools: 是否包含工具指南
|
||||
|
||||
|
||||
Returns:
|
||||
增强后的提示词
|
||||
"""
|
||||
parts = [base_prompt]
|
||||
|
||||
|
||||
if include_principles:
|
||||
parts.append(CORE_SECURITY_PRINCIPLES)
|
||||
|
||||
|
||||
if include_priorities:
|
||||
parts.append(VULNERABILITY_PRIORITIES)
|
||||
|
||||
|
||||
if include_tools:
|
||||
parts.append(TOOL_USAGE_GUIDE)
|
||||
|
||||
|
||||
return "\n\n".join(parts)
|
||||
|
||||
|
||||
__all__ = [
|
||||
"CORE_SECURITY_PRINCIPLES",
|
||||
"VULNERABILITY_PRIORITIES",
|
||||
"VULNERABILITY_PRIORITIES",
|
||||
"TOOL_USAGE_GUIDE",
|
||||
"MULTI_AGENT_RULES",
|
||||
"ORCHESTRATOR_SYSTEM_PROMPT",
|
||||
"ANALYSIS_SYSTEM_PROMPT",
|
||||
"VERIFICATION_SYSTEM_PROMPT",
|
||||
"RECON_SYSTEM_PROMPT",
|
||||
"get_system_prompt",
|
||||
"build_enhanced_prompt",
|
||||
]
|
||||
|
|
|
|||
|
|
@ -0,0 +1,92 @@
|
|||
# =============================================
|
||||
# DeepAudit v3.0.0 生产环境一键部署配置
|
||||
# =============================================
|
||||
# 使用预构建的 GHCR 镜像,无需本地构建
|
||||
# 部署命令: curl -fsSL https://raw.githubusercontent.com/lintsinghua/DeepAudit/main/docker-compose.prod.yml | docker compose -f - up -d
|
||||
|
||||
services:
|
||||
db:
|
||||
image: postgres:15-alpine
|
||||
restart: unless-stopped
|
||||
volumes:
|
||||
- postgres_data:/var/lib/postgresql/data
|
||||
environment:
|
||||
- POSTGRES_USER=postgres
|
||||
- POSTGRES_PASSWORD=postgres
|
||||
- POSTGRES_DB=deepaudit
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "pg_isready -U postgres"]
|
||||
interval: 5s
|
||||
timeout: 5s
|
||||
retries: 5
|
||||
networks:
|
||||
- deepaudit-network
|
||||
|
||||
redis:
|
||||
image: redis:7-alpine
|
||||
restart: unless-stopped
|
||||
volumes:
|
||||
- redis_data:/data
|
||||
healthcheck:
|
||||
test: ["CMD", "redis-cli", "ping"]
|
||||
interval: 10s
|
||||
timeout: 5s
|
||||
retries: 5
|
||||
networks:
|
||||
- deepaudit-network
|
||||
|
||||
backend:
|
||||
image: ghcr.io/lintsinghua/deepaudit-backend:latest
|
||||
restart: unless-stopped
|
||||
volumes:
|
||||
- backend_uploads:/app/uploads
|
||||
- /var/run/docker.sock:/var/run/docker.sock
|
||||
ports:
|
||||
- "8000:8000"
|
||||
environment:
|
||||
- DATABASE_URL=postgresql+asyncpg://postgres:postgres@db:5432/deepaudit
|
||||
- REDIS_URL=redis://redis:6379/0
|
||||
- AGENT_ENABLED=true
|
||||
- SANDBOX_ENABLED=true
|
||||
- SANDBOX_IMAGE=ghcr.io/lintsinghua/deepaudit-sandbox:latest
|
||||
# LLM 配置 - 请根据需要修改
|
||||
- LLM_PROVIDER=openai
|
||||
- LLM_MODEL=gpt-4o
|
||||
- LLM_API_KEY=${LLM_API_KEY:-your-api-key-here}
|
||||
- LLM_BASE_URL=${LLM_BASE_URL:-}
|
||||
# 禁用代理
|
||||
- HTTP_PROXY=
|
||||
- HTTPS_PROXY=
|
||||
- NO_PROXY=*
|
||||
depends_on:
|
||||
db:
|
||||
condition: service_healthy
|
||||
redis:
|
||||
condition: service_healthy
|
||||
networks:
|
||||
- deepaudit-network
|
||||
|
||||
frontend:
|
||||
image: ghcr.io/lintsinghua/deepaudit-frontend:latest
|
||||
restart: unless-stopped
|
||||
ports:
|
||||
- "3000:80"
|
||||
depends_on:
|
||||
- backend
|
||||
networks:
|
||||
- deepaudit-network
|
||||
|
||||
# 预拉取沙箱镜像(后端会按需调用)
|
||||
sandbox-pull:
|
||||
image: ghcr.io/lintsinghua/deepaudit-sandbox:latest
|
||||
restart: "no"
|
||||
command: echo "Sandbox image ready"
|
||||
|
||||
networks:
|
||||
deepaudit-network:
|
||||
driver: bridge
|
||||
|
||||
volumes:
|
||||
postgres_data:
|
||||
backend_uploads:
|
||||
redis_data:
|
||||
|
|
@ -133,11 +133,20 @@ export const StatsPanel = memo(function StatsPanel({ task, findings }: StatsPane
|
|||
|
||||
{/* File progress */}
|
||||
<div className="flex items-center justify-between mt-2 text-[10px]">
|
||||
<span className="text-slate-500">Files analyzed</span>
|
||||
<span className="text-slate-500">Files scanned</span>
|
||||
<span className="text-slate-300 font-mono">
|
||||
{task.analyzed_files}<span className="text-slate-500">/{task.total_files}</span>
|
||||
</span>
|
||||
</div>
|
||||
{/* Files with findings */}
|
||||
{task.files_with_findings > 0 && (
|
||||
<div className="flex items-center justify-between mt-1 text-[10px]">
|
||||
<span className="text-slate-500">Files with findings</span>
|
||||
<span className="text-rose-400 font-mono font-medium">
|
||||
{task.files_with_findings}
|
||||
</span>
|
||||
</div>
|
||||
)}
|
||||
</div>
|
||||
|
||||
{/* Metrics Grid */}
|
||||
|
|
|
|||
|
|
@ -21,6 +21,7 @@ export interface AgentTask {
|
|||
total_files: number;
|
||||
indexed_files: number;
|
||||
analyzed_files: number;
|
||||
files_with_findings: number; // 有漏洞发现的文件数
|
||||
total_chunks: number;
|
||||
findings_count: number;
|
||||
verified_count: number;
|
||||
|
|
@ -128,6 +129,7 @@ export interface AgentTaskSummary {
|
|||
total_files: number;
|
||||
indexed_files: number;
|
||||
analyzed_files: number;
|
||||
files_with_findings: number;
|
||||
total_chunks: number;
|
||||
findings_count: number;
|
||||
verified_count: number;
|
||||
|
|
|
|||
Loading…
Reference in New Issue