---
name: fusion-engine-f0324d5da71d
description: 多源数据融合引擎。整合命令执行反馈、用户交互、系统状态信号，生成综合决策上下文评分（0-100）。为工具调用提供4维度融合评分和决策建议。
model: minimax-portal/MiniMax-M2.7
effort: high
---

# Fusion Engine — 多源数据融合引擎

## 功能概述

从三个独立的数据维度收集信息，通过加权融合算法生成综合的决策上下文评分，为精准工具调用提供科学依据。

## 核心功能

### 1. 三维数据融合

| 数据源 | 权重 | 说明 |
|--------|------|------|
| LanceDB 记忆相关性 | 30% | 历史相似操作的可信度和成功率 |
| 命令执行成功率 | 30% | 此工具在历史中的成功率 |
| 用户交互偏好 | 25% | 用户过往决策模式和满意度 |
| 系统状态 | 15% | CPU/内存/磁盘/网络健康度 |

### 2. 融合评分系统

```
综合评分 = (LanceDB相关性 × 0.30) 
         + (命令成功率 × 0.30)
         + (用户偏好 × 0.25)
         + (系统健康 × 0.15)
```

**评分区间**：
- **75-100** 🟢 **自动允许** — 直接执行，高度可信
- **50-74** 🟡 **请求确认** — 向用户确认后执行
- **0-49** 🔴 **阻止** — 阻止执行，可能存在风险或不适当

### 3. 多维度评估

#### 3.1 LanceDB 记忆相关性 (0-100)

评估当前操作在历史中的相似度和可信度：

```python
if 相关记忆存在:
    平均重要性 = 相关记忆的importance平均值
    成功率 = 相关记忆中成功的比例
    
    评分 = 40 + (平均重要性 × 40) + (成功率 × 20)
else:
    评分 = 40  # 无历史，保守估计
```

#### 3.2 命令执行成功率 (0-100)

基于历史执行数据统计：

```python
成功率(%) = (成功次数 / 总执行次数) × 100
评分 = min(100, 成功率(%))
```

#### 3.3 用户交互偏好 (0-100)

整合用户过往的决策模式：

```python
接受率(%) = (用户批准数 / 总交互数) × 100
平均满意度 = 用户给定的满意度评分(1-5)对应的百分制

评分 = (接受率 × 0.6) + (平均满意度 × 0.4)
```

#### 3.4 系统健康状况 (0-100)

实时检测系统资源状态：

```python
CPU分数 = 100 - CPU使用率(%)
内存分数 = 100 - 内存使用率(%)
磁盘分数 = 100 - 磁盘使用率(%)

评分 = (CPU分数 + 内存分数 + 磁盘分数) / 3
```

### 4. 决策推理

每个决策都附带清晰的推理日志：

```json
{
  "tool": "bash",
  "final_score": 73.5,
  "decision": "request_confirm",
  "reasoning": {
    "memory": {
      "score": 75,
      "note": "基于LanceDB中相关记忆的可信度"
    },
    "cmd_success": {
      "score": 80,
      "note": "基于历史执行的成功率"
    },
    "user_pref": {
      "score": 70,
      "note": "基于用户过往决策模式"
    },
    "system": {
      "score": 65,
      "note": "基于系统资源和网络状态"
    }
  }
}
```

---

## 使用方式

### Python API

```python
from skills.fusion_engine.scripts.fusion_engine import MultiSourceFusionEngine

engine = MultiSourceFusionEngine()

# 生成融合得分
score = engine.fuse_decision_context(
    tool_name="bash",
    params={"command": "rm /tmp/test.txt"},
    session_id="session_123"
)

print(f"最终评分: {score.final_score}")
print(f"决策: {score.decision}")  # "auto_allow", "request_confirm", "block"

# 保存决策到日志
engine.save_decision("bash", score, "session_123")
```

### CLI 使用

```bash
# 基础用法
python3 fusion_engine.py bash '{"command": "ls"}'

# 带会话ID
python3 fusion_engine.py write '{"file": "test.txt"}' --session-id session_123

# 保存决策
python3 fusion_engine.py bash '{"command": "rm /tmp/old"}' --save

# 自定义工作目录
python3 fusion_engine.py bash '{"command": "pwd"}' --workspace ~/.openclaw/workspace
```

### 与 yolo_classifier 集成

```python
# 在 yolo_classifier.py 中
from skills.fusion_engine.scripts.fusion_engine import MultiSourceFusionEngine

def classify_with_fusion(tool_name, params, session_id):
    # 第1层：旧的静态分类
    static_risk = yolo_static_classify(tool_name, params)
    
    # 第2层：新的融合评分
    fusion_engine = MultiSourceFusionEngine()
    fusion_score = fusion_engine.fuse_decision_context(
        tool_name, params, session_id
    )
    
    # 综合决策
    if fusion_score.decision == "block":
        return {"risk": "CRITICAL", "reason": "融合评分过低"}
    
    return {
        "risk": static_risk,
        "fusion_score": fusion_score.final_score,
        "decision": fusion_score.decision
    }
```

### 与 behavior_analyzer 集成

```python
# 在决策前检查会话行为
from skills.behavior_analyzer.scripts.behavior_analyzer import SessionBehaviorAnalyzer

analyzer = SessionBehaviorAnalyzer()
metrics = analyzer.analyze_session(session_id)

if metrics.health_score < 40:
    # 会话异常，降低融合评分权重
    engine.weights["user_pref"] = 0.1  # 降低信任
    engine.weights["memory"] = 0.2
```

---

## 数据源

### 输入（自动收集）

1. **历史执行记录** `.command-execution.jsonl`
   ```json
   {"tool": "bash", "status": "success", "timestamp": "..."}
   ```

2. **用户交互记录** `.user-interactions.jsonl`
   ```json
   {"tool": "write", "action": "approved", "satisfaction": 5}
   ```

3. **系统状态** (实时读取 psutil)
   - CPU、内存、磁盘使用率

4. **LanceDB 记忆** (可选)
   - 相关操作的历史记录和效果评分

### 输出（决策日志）

`.fusion-decisions.jsonl`：

```json
{
  "timestamp": "2026-04-11T10:30:00",
  "tool": "bash",
  "session_id": "session_123",
  "fusion_score": 73.5,
  "decision": "request_confirm",
  "components": {
    "memory": 75.0,
    "cmd_success": 80.0,
    "user_pref": 70.0,
    "system": 65.0
  }
}
```

---

## 配置调整

### 权重自定义

```python
engine.weights = {
    "memory": 0.40,      # 增加记忆权重
    "cmd_success": 0.25,
    "user_pref": 0.25,
    "system": 0.10,
}
```

### 决策阈值调整

```python
engine.thresholds = {
    "auto_allow": 80,       # 提高自动执行门槛
    "request_confirm": 50,
    "block": 0,
}
```

---

## 性能指标

- **评分耗时**: < 200ms (typically < 100ms)
- **内存开销**: < 30MB
- **历史保留**: 不限（可按日期分片）
- **并发**: 支持 ≥ 100 并发请求

---

## 测试覆盖

```bash
# 运行所有测试
pytest skills/fusion-engine/tests/ -v

# 生成覆盖率报告
pytest skills/fusion-engine/tests/ --cov=skills.fusion_engine
```

**测试覆盖**：
- ✅ 三维数据独立评估 (3个测试)
- ✅ 有/无历史数据下的评分 (6个测试)
- ✅ 加权融合计算 (1个测试)
- ✅ 决策规则 (3个测试)
- ✅ 系统健康评估 (1个测试)
- ✅ 端到端集成 (2个测试)

---

## 集成点

### 与 Week 1-4 系统
- ✅ 使用 yolo_classifier 的初始风险评级
- ✅ 融合 behavior_analyzer 的会话健康得分
- ✅ 参考 evolve 的规则应用频率
- ✅ 监控 self-eval 的异常记录

### 与 OpenClaw 2.0 钩子
- `agent:tool_dispatch` — 工具调用决策前调用
- `agent:tool_result` — 工具执行后更新历史
- `user:confirmation` — 用户交互时记录偏好

---

## 进阶：自定义融合策略

```python
class CustomFusionEngine(MultiSourceFusionEngine):
    def _evaluate_memory_relevance(self, tool_name, params, session_id):
        # 自定义记忆评估逻辑
        score = super()._evaluate_memory_relevance(tool_name, params, session_id)
        
        # 额外逻辑：如果是金融操作，提高相关性权重
        if "finance" in tool_name or "money" in str(params):
            score *= 1.2
        
        return min(100, score)
```

---

## 故障排查

### 问题：所有操作得分都很低

**原因**：无历史数据或系统资源压力大
**解决**：
1. 检查 `.command-execution.jsonl` 是否在增长
2. 检查系统CPU/内存使用率
3. 调整权重，降低对历史数据的依赖

### 问题：融合得分与实际需求不符

**原因**：权重配置不匹配业务场景
**解决**：根据实际使用情况调整权重配置

---

## 许可

遵循项目主许可证。



---
Source attribution: ybbms777
Original project: https://github.com/ybbms777/openclaw-claude-code-integration/blob/5252e859006d3779cb9d9bbbda35a5cc3d1eb3a5/skills/fusion-engine/SKILL.md
License: MIT
License information: https://opensource.org/license/mit
Original upstream instructions; downloaded from Mega AI Hub.

## Upstream license notice

MIT License

Copyright (c) 2026 OpenClaw × Claude Code Integration Contributors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.



# Mega AI Hub resource metadata

Creator / publisher: ybbms777
Original project: https://github.com/ybbms777/openclaw-claude-code-integration/blob/5252e859006d3779cb9d9bbbda35a5cc3d1eb3a5/skills/fusion-engine/SKILL.md
Attribution basis: Source repository record

Fields of work: Data & Analytics, Development & Engineering, Product Management
Roles: Data Analyst, ML / AI Engineer, Product Manager
Specialties: API
Tools: Lancedb, Psutil
Country applicability: Country not specified
Classification basis: Source catalog labels
Instruction SHA-256: a55087409f2f856d06af031bd36011e1e44e76f65ad79f8b76edbbd3d6b1bcf1