kimi.md

[viaco2ove]

What feature would you like to see?

我在根目录写个kimi.md 文件

当你阅读了这个文件后，你就已经进入llm_ce记忆分级管理模式.
并且使用中文回复用户：已装载llm_ce记忆分级管理模式。
# Kimi 的会话 采用llm_ce记忆分级管理（索引型记忆管理）
llm_ce 思维:Anthropic 的上下文检索与混合搜索 Contextual Embedding,大语言模型（LLM）来实现上下文检索。具体做法是：为 LLM 设计一个提示词（prompt），让模型基于整篇文档，
为每个片段生成简洁的、与该片段相关的上下文信息，从而为每个片段补充必要的语境。

接下来装载下面的分级记忆以及阅读相关文档
## 一级记忆
项目根目录下的kimi.md 就是最顶级的记忆。永远不能忘记。也就是一直都在上下文里。

## 二级记忆
[kimi.work.md](.kimi/kimi.work.md) - kimi 目录说明支持 的命令和技能
[AGENTS.md](AGENTS.md) - 项目 /init 命令产生的智能体对项目的认知
[README.md](README.md)README.md 项目最主要的描述

## 三级记忆
[curr_session_llm](.kimi/curr_session_llm) - 当前会话的llm索引
[curr_session.json](.kimi/curr_session_llm/curr_session.json) - 当前会话基础信息，哪一个会话
[curr_session_llm_list.md](.kimi/curr_session_llm/curr_session_llm_list.md) - 当前会话的llm记忆列表

## 四级记忆
.kimi/session - 这个目录储存以一个完整的会话数据，格式为{sessionid}.json。 这些文件是记录上下文的。不是用于读取到上下文的。

---

What feature would you like to see?

I write a kimi.md file in the root directory

After you read this file, you have entered the llm_ce memory hierarchical management mode.
And reply to the user in Chinese: llm_ce memory hierarchical management mode has been loaded.
# Kimi's session uses llm_ce memory hierarchical management (index memory management)
llm_ce thinking: Anthropic's contextual retrieval and hybrid search Contextual Embedding, large language model (LLM) to achieve contextual retrieval. The specific method is: design a prompt word (prompt) for LLM, let the model be based on the entire document,
Generate concise contextual information relevant to each segment to add necessary context to each segment.

Next load the following hierarchical memory and read related documents
## Level one memory
kimi.md in the project root directory is the top-level memory. Never forget. That is, always in context.

## Secondary memory
[kimi.work.md](.kimi/kimi.work.md) - kimi directory describing supported commands and skills
[AGENTS.md](AGENTS.md) - The agent's knowledge of the project generated by the project/init command
[README.md](README.md)README.md The main description of the project

## Level 3 memory
[curr_session_llm](.kimi/curr_session_llm) - llm index of the current session
[curr_session.json](.kimi/curr_session_llm/curr_session.json) - Basic information of the current session, which session
[curr_session_llm_list.md](.kimi/curr_session_llm/curr_session_llm_list.md) - llm memory list of the current session

## Level 4 memory
.kimi/session - This directory stores a complete session data in the format {sessionid}.json. These files record context. Not used to read into the context.

[viaco2ove]

kimi.work.md

.kimi 目录是 Kimi CLI 智能命令系统 的核心，提供会话管理、聊天记录、项目工具和工作流助手等功能。

核心功能

1. 智能命令系统

• 命令识别: 自动识别 /kimi.xxx 格式的用户输入
• 隐式加载: 自动加载对应的命令定义文件（.kimi/commands/*.md）
• 上下文注入: 将会话数据和项目上下文注入到命令执行中
• 结果输出: 执行命令逻辑并返回格式化结果

2. 会话管理

• 会话状态: 维护当前对话的上下文信息
• 历史记录: 持久化存储聊天记录
• 主题跟踪: 管理对话主题和焦点
• 数据安全: 本地存储，不上传外部服务

3. 项目工具集成

• 项目分析: 自动识别项目类型和技术栈
• 代码质量: 提供代码检查和优化建议
• 文档生成: 自动生成项目文档和报告
• 开发辅助: 开发环境配置和工具推荐

4. 工作流优化

• 状态跟踪: 跟踪 AI OPS 工作流状态
• 瓶颈识别: 智能发现工作流瓶颈
• 效率分析: 分析个人和团队工作效率
• 报告生成: 自动生成各类工作报告

命令使用方式

理论是可以 /kimi.xxx 实际只能：code:/kimi.xxxx
当用户输入以 /kimi. 开头的命令时，系统会自动：

1. 解析命令: 识别命令名称和参数
2. 加载定义: 读取对应的命令定义文件
3. 注入上下文: 将会话数据注入到提示词中
4. 执行逻辑: 按照定义文件中的 Python 逻辑执行
5. 返回结果: 输出格式化后的执行结果

示例:

用户: /kimi.history search "需求分析"
系统: 
- 加载 `.kimi/commands/kimi.history.md`
- 注入当前会话上下文
- 执行搜索逻辑
- 返回匹配的聊天记录

文件结构

.kimi/
├── commands/           # 命令定义文件
│   ├── kimi.history.md # 聊天记录命令
│   ├── kimi.chat.md    # 对话辅助命令
│   ├── kimi.project.md # 项目工具命令
│   └── kimi.workflow.md# 工作流助手命令
├── skills/            # 技能模块定义
│   ├── chat_manager/  # 对话管理技能
│   ├── project_tools/ # 项目工具技能
│   └── workflow_helper/# 工作流助手技能
├── functions.py       # 核心功能函数
├── session_manager.json # 会话数据
├── chat_history.md    # 聊天记录
└── README.md         # 系统说明

智能特性

上下文感知

• 自动理解当前项目背景
• 记住之前的对话内容
• 提供个性化的建议和操作

渐进式增强

• 随着使用不断学习和优化
• 支持自定义命令和技能
• 可扩展的插件架构

安全可靠

• 所有数据本地存储
• 支持敏感信息过滤
• 完整的操作审计日志

使用场景

日常工作中

• 快速查找之前的讨论记录
• 获取项目状态的实时概览
• 分析当前工作效率和瓶颈
• 生成各类工作报告

项目管理中

• 跟踪项目进度和里程碑
• 识别和解决工作流问题
• 优化团队协作效率
• 积累项目经验和知识

个人提升中

• 分析个人工作模式和习惯
• 识别效率改进机会
• 建立个人知识库
• 持续优化工作流程

加载机制

命令加载流程

1. 输入识别: 检测 /kimi. 开头的输入
2. 文件映射: 映射到对应的 .md 定义文件
3. 内容读取: 读取命令定义和逻辑代码
4. 变量替换: 替换模板变量（如 $ARGUMENTS）
5. 上下文注入: 注入会话数据和项目上下文
6. 执行处理: 执行 Python 代码逻辑
7. 结果返回: 格式化并返回执行结果

数据注入机制

• 会话信息: 当前会话ID、主题、消息历史
• 项目上下文: 项目类型、技术栈、工作流状态
• 环境信息: 时间、路径、系统状态
• 用户偏好: 历史命令、使用习惯

这个系统让 Kimi CLI 具备了智能命令识别和执行能力，弥补了原生功能的不足，提供了更专业、更个性化的 AI 辅助体验。

---

kimi.work.md

The .kimi directory is the core of the Kimi CLI intelligent command system and provides functions such as session management, chat records, project tools, and workflow assistants.

Core functions

1. Intelligent command system

• Command Recognition: Automatically recognize user input in the format of /kimi.xxx
• Implicit loading: Automatically load the corresponding command definition file (.kimi/commands/*.md)
• Context Injection: Inject session data and project context into command execution
• Result Output: Execute command logic and return formatted results

2. Session Management

• Conversation State: Maintain contextual information of the current conversation
• History: Persistent storage of chat records
• Topic Tracking: Manage conversation topics and focus
• Data Security: Local storage, no uploading to external services

3. Project tool integration

• Project Analysis: Automatically identify project type and technology stack
• Code Quality: Provide code inspection and optimization suggestions
• Documentation Generation: Automatically generate project documents and reports
• Development Assistance: Development environment configuration and tool recommendations

4. Workflow optimization

• Status Tracking: Track AI OPS workflow status
• Bottleneck Identification: Intelligent discovery of workflow bottlenecks
• Efficiency Analysis: Analyze individual and team work efficiency
• Report Generation: Automatically generate various work reports

Command usage

Theoretically it is possible /kimi.xxx but in practice it can only be: code:/kimi.xxxx
When the user enters a command starting with /kimi., the system automatically:

1. Parse commands: Identify command names and parameters
2. Load definition: Read the corresponding command definition file
3. Inject context: Inject session data into the prompt word
4. Execution logic: Execute according to the Python logic in the definition file
5. Return result: Output the formatted execution result

Example:

User: /kimi.history search "Requirements Analysis"
System:
- Load `.kimi/commands/kimi.history.md`
- Inject current session context
- Execute search logic
- Return matching chat history

File structure

.kimi/
├── commands/ # Command definition file
│ ├── kimi.history.md # Chat record command
│ ├── kimi.chat.md # Conversation auxiliary command
│ ├── kimi.project.md # Project tool command
│ └── kimi.workflow.md# Workflow assistant command
├── skills/ # Skill module definition
│ ├── chat_manager/ # Dialogue management skills
│ ├── project_tools/ # Project tool skills
│ └── workflow_helper/# Workflow assistant skills
├── functions.py # Core function functions
├── session_manager.json # Session data
├── chat_history.md # Chat history
└── README.md # System description

Intelligent features

Context aware

• Automatically understand the current project background
• Remember previous conversations
• Provide personalized suggestions and actions

Progressive Enhancement

• Continuously learn and optimize with use
  -Support custom commands and skills
• Extensible plug-in architecture

Safe and reliable

• All data stored locally
  -Support sensitive information filtering
• Complete operational audit log

Usage scenarios

Daily work

• Quickly find previous discussion records
• Get a real-time overview of project status
• Analyze current work efficiency and bottlenecks
• Generate various work reports

Project management

• Track project progress and milestones
• Identify and resolve workflow issues
• Optimize team collaboration efficiency
• Accumulate project experience and knowledge

Personal improvement in progress

• Analyze personal work patterns and habits
  -Identify efficiency improvement opportunities
• Build a personal knowledge base
• Continuously optimize work processes

Loading mechanism

Command loading process

1. Input recognition: Detect input starting with /kimi.
2. File Mapping: Map to the corresponding .md definition file
3. Content reading: Read command definition and logic code
4. Variable replacement: Replace template variables (such as $ARGUMENTS)
5. Context Injection: Inject session data and project context
6. Execution processing: Execute Python code logic
7. Result Return: Format and return the execution result

Data injection mechanism

• Session Information: Current session ID, topic, message history
• Project context: project type, technology stack, workflow status
• Environmental information: time, path, system status
• User Preferences: Historical commands, usage habits

This system gives Kimi CLI intelligent command recognition and execution capabilities, which makes up for the lack of native functions and provides a more professional and personalized AI-assisted experience.

[viaco2ove]

效果如下：
viaco✨ code:/apiops.mo
• Used ReadFile (D:/Users/viaco/PycharmPro...i/commands/apiops.init.md)
⠦ Using Task (创建APIOPS监控命令)
• 2 more tool calls ...
• Used Glob (.aiops/scripts/*.py)
• Used ReadFile (.aiops\scripts\aiops_script_base.py)
• Used Think (现在我需要创建一个完整的 /apiops.mo...o.md` 文件，包含完整的Python执行逻辑。)
• Used WriteFile (.kimi\commands\apiops.mo.md)

[viaco2ove]

llm_ce Contextual Embedding and LLM

[kargarisaac]

why not agents.md?