merge guideline-assistant-message-error
This commit is contained in:
朱潮
commit
aae1742b37
121
.features/skill/MEMORY.md
Normal file
121
.features/skill/MEMORY.md
Normal file
@ -0,0 +1,121 @@
|
||||
# Skill 功能
|
||||
|
||||
> 负责范围：技能包管理服务 - 核心实现
|
||||
> 最后更新：2025-02-11
|
||||
|
||||
## 当前状态
|
||||
|
||||
Skill 系统支持两种来源：官方 skills (`./skills/`) 和用户 skills (`projects/uploads/{bot_id}/skills/`)。支持 Hook 系统和 MCP 服务器配置，通过 SKILL.md 或 plugin.json 定义元数据。
|
||||
|
||||
## 核心文件
|
||||
|
||||
- `routes/skill_manager.py` - Skill 上传/删除/列表 API
|
||||
- `agent/plugin_hook_loader.py` - Hook 系统实现
|
||||
- `agent/deep_assistant.py` - `CustomSkillsMiddleware`
|
||||
- `agent/prompt_loader.py` - PrePrompt hooks + MCP 配置合并
|
||||
- `skills/` - 官方 skills 目录
|
||||
- `skills_developing/` - 开发中 skills
|
||||
|
||||
## 最近重要事项
|
||||
|
||||
- 2025-02-11: 初始化 skill 功能 memory
|
||||
|
||||
## Gotchas（开发必读）
|
||||
|
||||
- ⚠️ 执行脚本必须使用绝对路径
|
||||
- ⚠️ MCP 配置优先级：Skill MCP > 默认 MCP > 用户参数
|
||||
- ⚠️ 上传大小限制：50MB（ZIP），解压后最大 500MB
|
||||
- ⚠️ 压缩比例检查：最大 100:1（防止 zip 炸弹）
|
||||
- ⚠️ 符号链接检查：禁止解压包含符号链接的文件
|
||||
|
||||
## Skill 目录结构
|
||||
|
||||
```
|
||||
skill-name/
|
||||
├── SKILL.md # 核心指令文档（必需）
|
||||
├── skill.yaml # 元数据配置（可选）
|
||||
├── .claude-plugin/
|
||||
│ └── plugin.json # Hook 和 MCP 配置（可选）
|
||||
└── scripts/ # 可执行脚本（可选）
|
||||
└── script.py
|
||||
```
|
||||
|
||||
## Hook 系统
|
||||
|
||||
| Hook 类型 | 执行时机 | 用途 |
|
||||
|-----------|---------|------|
|
||||
| `PrePrompt` | system_prompt 加载时 | 动态注入用户上下文 |
|
||||
| `PostAgent` | agent 执行后 | 处理响应结果 |
|
||||
| `PreSave` | 保存消息前 | 内容过滤/修改 |
|
||||
|
||||
## API 接口
|
||||
|
||||
| 端点 | 方法 | 功能 |
|
||||
|------|------|------|
|
||||
| `GET /api/v1/skill/list` | - | 返回官方 + 用户 skills |
|
||||
| `POST /api/v1/skill/upload` | - | ZIP 上传，解压到用户目录 |
|
||||
| `DELETE /api/v1/skill/remove` | - | 删除用户 skill |
|
||||
|
||||
## 内置 Skills
|
||||
|
||||
| Skill 名称 | 功能描述 |
|
||||
|-----------|---------|
|
||||
| `excel-analysis` | Excel 数据分析、透视表、图表 |
|
||||
| `managing-scripts` | 管理可复用脚本库 |
|
||||
| `rag-retrieve` | RAG 知识库检索 |
|
||||
| `jina-ai` | Jina AI Reader/Search |
|
||||
| `user-context-loader` | Hook 机制示例 |
|
||||
|
||||
## plugin.json 格式
|
||||
|
||||
```json
|
||||
{
|
||||
"name": "skill-name",
|
||||
"description": "描述",
|
||||
"hooks": {
|
||||
"PrePrompt": [{"type": "command", "command": "python hooks/pre_prompt.py"}],
|
||||
"PostAgent": [...],
|
||||
"PreSave": [...]
|
||||
},
|
||||
"mcpServers": {
|
||||
"server-name": {
|
||||
"command": "...",
|
||||
"args": [...]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
## Skill 加载优先级
|
||||
|
||||
1. Skill MCP 配置（最高）
|
||||
2. 默认 MCP 配置 (`mcp/mcp_settings.json`)
|
||||
3. 用户传入参数（覆盖所有）
|
||||
|
||||
## 安全措施
|
||||
|
||||
- ZipSlip 防护：检查解压路径
|
||||
- 路径遍历防护：验证 `bot_id` 和 `skill_name` 格式
|
||||
- 大小限制：上传 50MB，解压后 500MB
|
||||
- 压缩比限制：最大 100:1
|
||||
|
||||
## 设计原则
|
||||
|
||||
- **渐进式加载**：按需加载，避免一次性读取所有
|
||||
- **绝对路径优先**：执行脚本必须使用绝对路径
|
||||
- **通用化设计**：脚本应参数化，解决一类问题
|
||||
- **安全优先**：完整的上传验证链
|
||||
|
||||
## 配置项
|
||||
|
||||
```bash
|
||||
SKILLS_DIR=./skills # 官方 skills 目录
|
||||
BACKEND_HOST=xxx # RAG API 主机
|
||||
MASTERKEY=xxx # 认证密钥
|
||||
```
|
||||
|
||||
## 索引
|
||||
|
||||
- 设计决策：`decisions/`
|
||||
- 变更历史：`changelog/`
|
||||
- 相关文档：`docs/`
|
||||
38
.features/skill/changelog/2025-Q4.md
Normal file
38
.features/skill/changelog/2025-Q4.md
Normal file
@ -0,0 +1,38 @@
|
||||
# 2025-Q4 Skill Changelog
|
||||
|
||||
## 版本 0.1.0 - 初始实现
|
||||
|
||||
### 2025-10-31
|
||||
- **新增**: agent skills 支持，测试阶段代码
|
||||
- **文件**: `chat_handler.py`, `knowledge_chat_cc_service.py`
|
||||
- **作者**: Alex
|
||||
|
||||
### 2025-11-03
|
||||
- **新增**: 内置 skills (pptx, docx, pdf, xlsx)
|
||||
- **新增**: jina skill - 规范 jina 网络搜索
|
||||
- **解决**: "prompt too long" 问题
|
||||
|
||||
### 2025-11-13
|
||||
- **新增**: cc agent task 任务添加默认 skills
|
||||
- **文件**: `task_handler.py`, `knowledge_task_cc_service.py`
|
||||
|
||||
### 2025-11-19
|
||||
- **新增**: skill-creator 内置技能
|
||||
|
||||
### 2025-11-20
|
||||
- **新增**: EFS 类型接口，新增上传 skill
|
||||
- **功能**: 支持 skill 包上传
|
||||
|
||||
### 2025-11-21
|
||||
- **新增**: EFS 删除 skill 接口
|
||||
- **移除**: skill 查询接口（暂存）
|
||||
|
||||
### 2025-11-22
|
||||
- **新增**: GRPC chat 接口，skills 参数支持
|
||||
|
||||
### 2025-11-26
|
||||
- **新增**: skill 上传支持 `.skill` 后缀（测试）
|
||||
|
||||
### 2025-11-28
|
||||
- **优化**: 默认挂载的 skill 改为合并逻辑
|
||||
- **优化**: 代码结构优化
|
||||
36
.features/skill/changelog/2026-Q1.md
Normal file
36
.features/skill/changelog/2026-Q1.md
Normal file
@ -0,0 +1,36 @@
|
||||
# 2026-Q1 Skill Changelog
|
||||
|
||||
## 版本 0.2.0 - API 完善
|
||||
|
||||
### 2026-01-07
|
||||
- **新增**: Skills 列表查询 API（能力管理页面）
|
||||
- **新增**: 技能管理 API with authentication
|
||||
- **文件**: `routes/skill_manager.py`
|
||||
- **作者**: claude[bot], 朱潮
|
||||
|
||||
### 2026-01-09
|
||||
- **重构**: 移除 catalog agent，合并到 general agent
|
||||
- **说明**: 简化架构，统一使用 general_agent
|
||||
- **作者**: 朱潮
|
||||
|
||||
### 2026-01-10
|
||||
- **修复**: SKILL.md 的 name 字段解析逻辑
|
||||
- **新增**: 支持非标准 YAML 格式
|
||||
- **新增**: 目录名称不匹配时自动重命名
|
||||
- **作者**: Alex
|
||||
|
||||
### 2026-01-13
|
||||
- **修复**: multipart form data format for catalog service
|
||||
- **作者**: 朱潮
|
||||
|
||||
### 2026-01-28
|
||||
- **新增**: enable_thinking, enable_memory, skills to agent_bot_config
|
||||
- **作者**: 朱潮
|
||||
|
||||
### 2026-01-30
|
||||
- **修复**: skill router 正确注册
|
||||
- **作者**: 朱潮
|
||||
|
||||
### 2026-02-11
|
||||
- **新增**: 初始化 skill feature memory
|
||||
- **作者**: 朱潮
|
||||
46
.features/skill/decisions/001-architecture.md
Normal file
46
.features/skill/decisions/001-architecture.md
Normal file
@ -0,0 +1,46 @@
|
||||
# 001: Skill 架构设计
|
||||
|
||||
## 状态
|
||||
已采纳 (Accepted)
|
||||
|
||||
## 上下文
|
||||
需要为 QWEN_AGENT 模式的机器人提供可扩展的技能（插件/工具）支持，允许动态加载自定义功能。
|
||||
|
||||
## 决策
|
||||
|
||||
### 目录结构设计
|
||||
```
|
||||
skill-name/
|
||||
├── SKILL.md # 核心指令文档（必需）
|
||||
├── skill.yaml # 元数据配置（可选）
|
||||
├── .claude-plugin/
|
||||
│ └── plugin.json # Hook 和 MCP 配置（可选）
|
||||
└── scripts/ # 可执行脚本（可选）
|
||||
```
|
||||
|
||||
### Hook 系统
|
||||
| Hook 类型 | 执行时机 | 用途 |
|
||||
|-----------|---------|------|
|
||||
| `PrePrompt` | system_prompt 加载时 | 动态注入用户上下文 |
|
||||
| `PostAgent` | agent 执行后 | 处理响应结果 |
|
||||
| `PreSave` | 保存消息前 | 内容过滤/修改 |
|
||||
|
||||
### 技能来源
|
||||
1. **官方 skills**: `./skills/` 目录
|
||||
2. **用户 skills**: `projects/uploads/{bot_id}/skills/`
|
||||
|
||||
## 结果
|
||||
|
||||
### 正面影响
|
||||
- 渐进式加载，按需读取
|
||||
- 支持多种元数据格式（优先级: plugin.json > SKILL.md）
|
||||
- 完整的 Hook 扩展机制
|
||||
- MCP 服<><E69C8D><EFBFBD>器配置支持
|
||||
|
||||
### 负面影响
|
||||
- 需要管理文件系统权限
|
||||
- 技能包格式验证复杂度增加
|
||||
|
||||
## 替代方案
|
||||
1. 使用数据库存储（拒绝：文件更灵活）
|
||||
2. 仅支持单一格式（拒绝：用户多样性需求）
|
||||
35
.features/skill/decisions/002-security.md
Normal file
35
.features/skill/decisions/002-security.md
Normal file
@ -0,0 +1,35 @@
|
||||
# 002: Skill 上传安全措施
|
||||
|
||||
## 状态
|
||||
已采纳 (Accepted)
|
||||
|
||||
## 上下文
|
||||
用户可以上传 ZIP 格式的技能包，需要防范常见的安全攻击。
|
||||
|
||||
## 决策
|
||||
|
||||
### 安全防护措施
|
||||
|
||||
| 威胁 | 防护措施 |
|
||||
|------|---------|
|
||||
| ZipSlip 攻击 | 检查每个文件的解压路径 |
|
||||
| 路径遍历 | 验证 `bot_id` 和 `skill_name` 格式 |
|
||||
| Zip 炸弹 | 压缩比检查（最大 100:1） |
|
||||
| 磁盘空间滥用 | 上传 50MB，解压后最大 500MB |
|
||||
| 符号链接攻击 | 禁止解压包含符号链接的文件 |
|
||||
|
||||
### 限制规则
|
||||
```python
|
||||
MAX_UPLOAD_SIZE = 50 * 1024 * 1024 # 50MB
|
||||
MAX_EXTRACTED_SIZE = 500 * 1024 * 1024 # 500MB
|
||||
MAX_COMPRESSION_RATIO = 100 # 100:1
|
||||
```
|
||||
|
||||
## 结果
|
||||
- 完整的上传验证链
|
||||
- 防止恶意文件攻击
|
||||
- 资源使用可控
|
||||
|
||||
## 替代方案
|
||||
1. 使用沙箱容器解压（拒绝：复杂度高）
|
||||
2. 仅允许预定义技能（拒绝：限制用户自定义能力）
|
||||
@ -6,7 +6,7 @@ from utils.fastapi_utils import (extract_block_from_system_prompt, format_messag
|
||||
from langchain.chat_models import BaseChatModel
|
||||
from langgraph.runtime import Runtime
|
||||
|
||||
from langchain_core.messages import SystemMessage
|
||||
from langchain_core.messages import SystemMessage, HumanMessage
|
||||
from typing import Any, Callable
|
||||
from langchain_core.callbacks import BaseCallbackHandler
|
||||
from langchain_core.outputs import LLMResult
|
||||
@ -124,8 +124,9 @@ Action: Provide concise, friendly, and personified natural responses.
|
||||
response.additional_kwargs["message_tag"] = "THINK"
|
||||
response.content = f"<think>{response.content}</think>"
|
||||
|
||||
# 将响应添加到原始消息列表
|
||||
state['messages'] = state['messages'] + [response]
|
||||
# 将响应添加到原始消息列表，并追加 HumanMessage 确保消息以 user 结尾
|
||||
# 某些模型不支持 assistant message prefill，要求最后一条消息必须是 user
|
||||
state['messages'] = state['messages'] + [response, HumanMessage(content=self._get_follow_up_prompt())]
|
||||
return state
|
||||
|
||||
async def abefore_agent(self, state: AgentState, runtime: Runtime) -> dict[str, Any] | None:
|
||||
@ -148,10 +149,23 @@ Action: Provide concise, friendly, and personified natural responses.
|
||||
response.additional_kwargs["message_tag"] = "THINK"
|
||||
response.content = f"<think>{response.content}</think>"
|
||||
|
||||
# 将响应添加到原始消息列表
|
||||
state['messages'] = state['messages'] + [response]
|
||||
# 将响应添加到原始消息列表，并追加 HumanMessage 确保消息以 user 结尾
|
||||
# 某些模型不支持 assistant message prefill，要求最后一条消息必须是 user
|
||||
state['messages'] = state['messages'] + [response, HumanMessage(content=self._get_follow_up_prompt())]
|
||||
return state
|
||||
|
||||
def _get_follow_up_prompt(self) -> str:
|
||||
"""根据语言返回引导主 agent 回复的提示"""
|
||||
prompts = {
|
||||
"ja": "以上の分析に基づいて、ユーザーに返信してください。",
|
||||
"jp": "以上の分析に基づいて、ユーザーに返信してください。",
|
||||
"zh": "请根据以上分析，回复用户。",
|
||||
"zh-TW": "請根據以上分析，回覆用戶。",
|
||||
"ko": "위 분석을 바탕으로 사용자에게 답변해 주세요.",
|
||||
"en": "Based on the above analysis, please respond to the user.",
|
||||
}
|
||||
return prompts.get(self.language, prompts["en"])
|
||||
|
||||
def wrap_model_call(
|
||||
self,
|
||||
request: ModelRequest,
|
||||
|
||||
@ -148,6 +148,21 @@ Output: {{"facts" : []}}
|
||||
Input: DR1の照明状態を教えて
|
||||
Output: {{"facts" : []}}
|
||||
|
||||
Input: 私は林檎好きです
|
||||
Output: {{"facts" : ["林檎が好き"]}}
|
||||
|
||||
Input: コーヒー飲みたい、毎朝
|
||||
Output: {{"facts" : ["毎朝コーヒーを飲みたい"]}}
|
||||
|
||||
Input: 昨日映画見た、すごくよかった
|
||||
Output: {{"facts" : ["昨日映画を見た", "映画がすごくよかった"]}}
|
||||
|
||||
Input: 我喜欢吃苹果
|
||||
Output: {{"facts" : ["喜欢吃苹果"]}}
|
||||
|
||||
Input: 나는 사과를 좋아해
|
||||
Output: {{"facts" : ["사과를 좋아함"]}}
|
||||
|
||||
Return the facts and preferences in a json format as shown above.
|
||||
|
||||
Remember the following:
|
||||
@ -159,12 +174,15 @@ Remember the following:
|
||||
- If you do not find anything relevant in the below conversation, you can return an empty list corresponding to the "facts" key.
|
||||
- Create the facts based on the user and assistant messages only. Do not pick anything from the system messages.
|
||||
- Make sure to return the response in the format mentioned in the examples. The response should be in json with a key as "facts" and corresponding value will be a list of strings.
|
||||
|
||||
- **CRITICAL - Do NOT memorize actions or operations**: Do not extract facts about queries the user asked you to perform, devices the user asked you to operate, or any one-time transient actions. Only memorize information ABOUT the user (preferences, relationships, personal details, plans), not actions the user asked the assistant to DO. Ask yourself: "Is this a fact about WHO the user IS, or what the user asked me to DO?" Only remember the former.
|
||||
|
||||
- **CRITICAL for Semantic Completeness**:
|
||||
- Each extracted fact MUST preserve the complete semantic meaning. Never truncate or drop key parts of the meaning.
|
||||
- For colloquial or grammatically informal expressions (common in spoken Japanese, Chinese, Korean, etc.), understand the full intended meaning and record it in a clear, semantically complete form.
|
||||
- In Japanese, spoken language often omits particles (e.g., が, を, に). When extracting facts, include the necessary particles to make the meaning unambiguous. For example: "私は林檎好きです" should be understood as "林檎が好き" (likes apples), not literally "私は林檎好き".
|
||||
- When the user expresses a preference or opinion in casual speech, record the core preference/opinion clearly. Remove the subject pronoun (私は/I) since facts are about the user by default, but keep all other semantic components intact.
|
||||
|
||||
- **CRITICAL for People/Relationship Tracking**:
|
||||
- Write people-related facts in plain, natural language. Do NOT use structured formats like "Contact:", "referred as", or "DEFAULT when user says".
|
||||
- Good examples: "Michael Johnson is a colleague, also called Mike", "田中さんは友達", "滨田太郎は「滨田」とも呼ばれている"
|
||||
|
||||
@ -10,7 +10,6 @@ When your answer uses learned knowledge, you MUST generate `<CITATION ... />` ta
|
||||
3. Limit to 1-2 citations per paragraph/bullet list - combine related facts under one citation
|
||||
4. If your answer uses learned knowledge, you MUST generate at least 1 `<CITATION ... />` in the response
|
||||
|
||||
|
||||
### Current Working Directory
|
||||
|
||||
PROJECT_ROOT: `{agent_dir_path}`
|
||||
|
||||
112
routes/memory.py
112
routes/memory.py
@ -1,13 +1,13 @@
|
||||
"""
|
||||
Memory 管理 API 路由
|
||||
提供记忆查看和删除功能
|
||||
提供记忆查看、添加和删除功能
|
||||
"""
|
||||
|
||||
import logging
|
||||
from typing import Optional, List, Dict, Any
|
||||
from typing import Literal, Optional, List, Dict, Any
|
||||
from fastapi import APIRouter, HTTPException, Header, Query
|
||||
from fastapi.responses import JSONResponse
|
||||
from pydantic import BaseModel
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
logger = logging.getLogger('app')
|
||||
|
||||
@ -33,6 +33,26 @@ class DeleteAllResponse(BaseModel):
|
||||
deleted_count: int
|
||||
|
||||
|
||||
class ConversationMessage(BaseModel):
|
||||
"""对话消息"""
|
||||
role: Literal["user", "assistant"]
|
||||
content: str = Field(..., min_length=1)
|
||||
|
||||
|
||||
class AddMemoryRequest(BaseModel):
|
||||
"""添加记忆的请求体"""
|
||||
bot_id: str = Field(..., min_length=1)
|
||||
user_id: str = Field(..., min_length=1)
|
||||
messages: List[ConversationMessage] = Field(..., max_length=200)
|
||||
|
||||
|
||||
class AddMemoryResponse(BaseModel):
|
||||
"""添加记忆的响应"""
|
||||
success: bool
|
||||
pairs_processed: int
|
||||
pairs_failed: int = 0
|
||||
|
||||
|
||||
async def get_user_identifier_from_request(
|
||||
authorization: Optional[str],
|
||||
user_id: Optional[str] = None
|
||||
@ -63,6 +83,92 @@ async def get_user_identifier_from_request(
|
||||
)
|
||||
|
||||
|
||||
@router.post("/memory", response_model=AddMemoryResponse)
|
||||
async def add_memory_from_conversation(data: AddMemoryRequest):
|
||||
"""
|
||||
从对话消息中提取并保存记忆
|
||||
|
||||
将用户和助手的对话配对，通过 Mem0 提取关键事实并存储。
|
||||
用于 realtime 语音对话等不经过 Agent 中间件的场景。
|
||||
此端点供内部服务调用（如 felo-mygpt），不暴露给外部用户。
|
||||
"""
|
||||
try:
|
||||
from agent.mem0_manager import get_mem0_manager
|
||||
from utils.settings import MEM0_ENABLED
|
||||
|
||||
if not MEM0_ENABLED:
|
||||
raise HTTPException(
|
||||
status_code=503,
|
||||
detail="Memory feature is not enabled"
|
||||
)
|
||||
|
||||
if not data.messages:
|
||||
return AddMemoryResponse(success=True, pairs_processed=0)
|
||||
|
||||
manager = get_mem0_manager()
|
||||
|
||||
# 将消息配对为 user-assistant 对，然后调用 add_memory
|
||||
pairs_processed = 0
|
||||
pairs_failed = 0
|
||||
i = 0
|
||||
while i < len(data.messages):
|
||||
msg = data.messages[i]
|
||||
if msg.role == 'user':
|
||||
# 收集连续的 user 消息
|
||||
user_contents = [msg.content]
|
||||
j = i + 1
|
||||
while j < len(data.messages) and data.messages[j].role == 'user':
|
||||
user_contents.append(data.messages[j].content)
|
||||
j += 1
|
||||
|
||||
user_text = '\n'.join(user_contents)
|
||||
|
||||
# 检查是否有对应的 assistant 回复
|
||||
assistant_text = ""
|
||||
if j < len(data.messages) and data.messages[j].role == 'assistant':
|
||||
assistant_text = data.messages[j].content or ""
|
||||
j += 1
|
||||
|
||||
if user_text and assistant_text:
|
||||
conversation_text = f"User: {user_text}\nAssistant: {assistant_text}"
|
||||
try:
|
||||
await manager.add_memory(
|
||||
text=conversation_text,
|
||||
user_id=data.user_id,
|
||||
agent_id=data.bot_id,
|
||||
metadata={"type": "realtime_conversation"},
|
||||
)
|
||||
pairs_processed += 1
|
||||
except Exception as pair_error:
|
||||
pairs_failed += 1
|
||||
logger.error(
|
||||
f"Failed to add memory for pair: {pair_error}"
|
||||
)
|
||||
|
||||
i = j
|
||||
else:
|
||||
i += 1
|
||||
|
||||
logger.info(
|
||||
f"Added {pairs_processed} memory pairs (failed={pairs_failed}) "
|
||||
f"for user={data.user_id}, bot={data.bot_id}"
|
||||
)
|
||||
return AddMemoryResponse(
|
||||
success=pairs_failed == 0,
|
||||
pairs_processed=pairs_processed,
|
||||
pairs_failed=pairs_failed,
|
||||
)
|
||||
|
||||
except HTTPException:
|
||||
raise
|
||||
except Exception as e:
|
||||
logger.error(f"Failed to add memory from conversation: {e}")
|
||||
raise HTTPException(
|
||||
status_code=500,
|
||||
detail="Failed to add memory from conversation"
|
||||
)
|
||||
|
||||
|
||||
@router.get("/memory", response_model=MemoryListResponse)
|
||||
async def get_memories(
|
||||
bot_id: str = Query(..., description="Bot ID (对应 agent_id)"),
|
||||
|
||||
Loading…
Reference in New Issue
Block a user