/
免费使用 Get Started
📖 Technical Docs

技术文档 Technical Documentation

深入解析双核引擎、自进化元认知层与系统架构设计 Deep dive into dual-core engine, metacognitive self-evolution, and system architecture

快速开始 (Getting Started) Getting Started

ONE 是基于强大的 OpenHarness 双核底座，专为复杂软件研发与自主系统协同打造的高阶桌面级智能体工作站。它超越了传统的单智能体交互，演进为一套具备自我反思、主动净化与跨会话成长的自进化元认知智能体系统。 ONE is a high-level desktop-grade agent workstation built on the robust OpenHarness dual-core foundation, tailored for complex software development and autonomous collaboration. It transcends traditional single-agent interactions, evolving into a self-evolving metacognitive agent system featuring self-reflection, active purification, and cross-session growth.

安装与环境初始化 Installation & Initialization

通过命令行，可以一键初始化运行环境并自动安装必要的系统依赖： Run the setup script via CLI to initialize the environment and install system dependencies: 

# 克隆项目并运行依赖安装脚本 / Clone and run setup
./scripts/setup.sh

# 启动开发工作站，同时加载后端 API、WebSocket 网关与 React/Electron 前端 / Launch workspace
./scripts/dev.sh

双核智能体架构 (Dual-Core Engine) Dual-Core Agent Architecture

ONE 系统内核通过**双引擎**深度耦合大语言模型的生成逻辑与物理系统的安全操作： The core of ONE deeply couples the generative logic of LLMs with the secure operations of physical systems via a **Dual-Engine** design:

  • ClaudeCode Core: 作为核心编译与深度逻辑审计引擎。能深入理解复杂代码图谱，自动编写语法树分析、运行高级 Git 版本流控制及多文件增量重构。 Serves as the core compiler and logical audit engine. It gains a deep understanding of complex codebase maps, runs AST parsing, executes Git workflows, and builds incremental multi-file refactorings.
  • OpenHarness Engine: 挂载 43 个高阶物理层研发工具（包括全栈 CDP 浏览器控制、加密沙箱终端和文件流读写），保障智能体的动作均在安全防线以内。 Mounts 43 high-level toolsets (including full CDP browser control, encrypted sandboxed terminals, and file streams), guaranteeing agent actions stay within security boundaries.

在此之上，ONE 实现了**多智能体蜂群协同 (Swarm Architecture)**。系统在一个包容性框架中挂载多智能体协同运行，并在前端提供独创的**分屏协同面板 (Split-Screen Panel)**。用户可以在同一个工作区内开启多列视窗，实时横向比对并监控不同 Worker 智能体的思考链、执行日志与状态，解决暗盒痛点。 Above this, ONE implements **Multi-Agent Swarm Collaboration (Swarm Architecture)**. Multiple concurrent agents run side-by-side in a single window using a premium **Split-Screen Panel**, allowing users to monitor thoughts, logs, and tools across agents in real time, resolving black-box uncertainties.

在浏览器自动化侧，系统深度集成了 **OpenCLI 人类行为模拟模块 (Human Simulation)**。它通过仿真鼠标悬停延迟（300-800ms）、随机按键间隔（50-150ms）和平滑多次滚动，跨 100 多个主流内容和商业平台进行安全的数据采集和分发，完美穿透防护拦截。 For web tasking, the system integrates **OpenCLI Human Simulation Module**. By mimicking human mouse delays (300-800ms), keystroke pauses (50-150ms), and smooth scrolls, it securely scrapes and posts across 100+ content and business platforms without triggering bot blocks.

执行 - 评估 - 进化（EEE）自进化环路 The Execution - Evaluation - Evolution Loop

系统核心运行着由“执行-评估-进化”组成的三层元认知闭环： The system core runs a 3-layer metacognitive loop of Execution, Evaluation, and Evolution:

  1. 执行层 (Execution) Execution Layer : Workers 智能体根据任务需求在隔离沙箱中调度工具进行输出。 Workers schedule tools in isolated sandboxes to complete tasks based on requests.
  2. 评估层 (Evaluation - Metacognitive 4D Scorecard) Evaluation Layer (Metacognitive 4D Scorecard) : 严苛审计每次任务。在**正确性 (40%)、优雅度 (20%)、高效性 (20%)、经济性 (20%)** 四个维度运行雷达评估，并在控制台无缝渲染多维评分与反思日志。 Audits tasks on four physical dimensions: **Correctness (40%), Elegance (20%), Efficiency (20%), and Runway Economy (20%)**, rendering multi-dimensional scores and reflections in the console.
  3. 进化层 (Evolution - Anti-Patterns & Dreaming) Evolution Layer (Anti-Patterns & Dreaming) : 一旦评估得分低于 70 分，系统哨兵将自动提炼并记录“负向约束记忆（Anti-Patterns）”，以防智能体再次踩坑。在空闲时，后台触发 Auto-Dreaming，提炼并整合零散记忆，蒸馏出高维商业认知。 If a session scores below 70, the system automatically distills and records "Negative Constraint Memory (Anti-Patterns)" to prevent future failures. Offline Auto-Dreaming runs during idle times to merge memory fragments and distill core insights.

AI 生态 12 层产业链图谱 (12-Layer Ecology) The 12-Layer AI Ecosystem Map

AI 的本质是“能源向 Token 的高效转化”。整个 AI 生态链自底向上可划分为 12 层： The essence of AI is "efficient conversion of energy into tokens." The complete AI ecosystem consists of 12 distinct layers:

  • 第 1-3 层：能源、芯片制程及集群数据中心基建。Layers 1-3: Energy, silicon fabrication, and data center infrastructures.
  • 第 4-6 层：大模型大脑（LLM）、工作记忆上下文（Context）及长期硬盘存储。Layers 4-6: Foundational brain (LLM), work memory (Context), and long-term storage.
  • 第 7-9 层：手脚集成（CDP 浏览器、安全终端）、调度控制框架（OpenHarness）以及最小执行单元（Agent）。Layers 7-9: Hand-foot integration, scheduling harness (OpenHarness), and functional agents.
  • 第 10-12 层：多个 Agent 并行协作的 **AI 原生部门**、极薄人力重度硅基的 **AI 原生公司**，以及智能体之间直接交易自发组网的 **AI 原生经济生态**。Layers 10-12: **AI-Native Departments**, leverage-intensive **AI-Native Companies**, and **AI-Native Economic Networks** where agents contract directly.

ONE 的设计精准落脚于**第 10-11 层**，致力于将独立创业者打造为“硅基商业帝国的编排者”。 ONE is engineered precisely for **Layers 10-11**, aiming to empower solo founders as orchestrators of silicon-leveraged business flows.

真正的“AI 原生思维” (AI Mindset) The True AI-Native Mindset

在 AI 原生时代，传统的“线性努力”已被打破。优秀的联合创始人必须完成认知的飞跃： In the AI-native era, traditional linear effort is obsolete. Co-founders must achieve cognitive leaps:

  1. 代码是耗材，逻辑是资产 Code is Consumable, Logic is Asset : 不要珍惜一段具体的代码。它只是实现目标的瞬时排泄物。如果运行不好，让 AI 在几秒内重写。您唯一的资产是您的“Prompt 轨迹”与“业务闭环逻辑”。 Do not value static code. It is merely a transient byproduct to reach an output. Let AI rewrite it on failure. Your only real asset is the prompt lineage and business logic.
  2. 反馈驱动优于完美设计 Feedback-Driven over Rigid Design : 不要试图在前期防御所有 Bug。直接撞上去运行（Run it），获取报错日志回灌给 AI 修复。AI 在处理脏乱报错中的进化速度远超人类的纯脑力预防。 Avoid rigid up-front defensive designs. Run the task immediately, retrieve error logs, and feed them back to the AI. AI adapts to real failures faster than any prevention.
  3. 杠杆率就是一切 Leverage is Everything : 个人的精力不再是瓶颈。人类的价值在于做那 1% 的“方向决策与审美确认”，其余 99% 的重度执行完全交给无限并发、不眠不休的 Agent 蜂群。 Human time is no longer the bottleneck. Human value lies in the 1% direction choices and aesthetic audits, while 99% execution is delegated to parallel agent swarms.

欲深入了解此设计范式，欢迎下载阅读 Anthropic 官方发布的《创始人手册原文》： To learn more about this design paradigm, download the original playbook published by Anthropic:

下载《Anthropic 创始人手册原文》(PDF) Download "Anthropic Founder Playbook" (PDF)

Obsidian 知识库挂载与沙箱隔离 Obsidian Vault Integration & Sandbox

ONE 支持将您本地的 Obsidian 笔记库挂载为 Agent 的外挂脑。通过将记忆路径重定向至隔离目录，在提取有效背景的同时完美保护您的个人隐私笔记免遭扫描，防止 Token 爆炸。 ONE mounts local Obsidian vaults to act as the agent's external brain. By redirecting memory paths to isolated folders, it leverages contexts without exposing private journals or expanding token usage.

智能路径写重定向与路由逻辑 Smart Write Redirection & Routing Logic

为防止智能体写入的产物污染您的 Obsidian 笔记根目录，系统后端会在 `resolve_note_path` 中运行以下解耦逻辑： To prevent agent outputs from cluttering your root Obsidian vault directories, the backend runs the following routing loop in `resolve_note_path`: 

def resolve_note_path(vault_path: Path, note_path_str: str, check_exists: bool = False) -> Path:
 note_path_str = note_path_str.strip().strip("/")
 direct_path = (vault_path / note_path_str).resolve()
 try:
 direct_path.relative_to(vault_path)
 if check_exists and direct_path.exists() and direct_path.is_file():
 return direct_path
 except (ValueError, RuntimeError):
 pass
 
 parts = Path(note_path_str).parts
 # 如果是未分类的相对路径，自动重定向至“智能体产出”区
 # Redirect loose relative paths to designated Agent Output folder
 if len(parts) == 1 or not any(note_path_str.startswith(prefix) for prefix in ["2.项目", "0.经营决策", "Agent_Memory", "1.Agent_Memory"]):
 note_path_str = f"2.项目/One Platform/2.智能体产出/{note_path_str}"
 return (vault_path / note_path_str).resolve()

长期记忆宫殿与记忆商城 (Memory Palace & Mall) Long-Term Memory Palace & Mall

ONE 打破了传统 AI 的 7 秒短期记忆宿命。您的商业决策、核心技术文档以及特定运营偏好都将在系统底座中得到永久保存。 ONE breaks the short-term memory constraints of traditional AI. Every commercial decision, technical layout, and custom operational rule is permanently retained in the system database.

  • 认知胶囊 (Cognitive Capsule) Cognitive Capsule : 系统将经营策略、工具调用心流、以及“负面防犯错约束（Anti-Patterns）”打包为高维独立认知包。支持一键导入与导出。 Packages operations manuals, tool-use intuition, and safety policies into clean, self-contained files, ready for export.
  • 记忆商城 (Memory Marketplace) Memory Marketplace : 未来 AI 原生生态的交易节点。共享和交易各行业顶尖创始人提炼的“商业心流胶囊”与“敏捷研发胶囊”。将非对称的经验直接泵入系统底座，实现即插即用。 The trading hub for AI native nodes. Import and deploy battle-tested strategy capsules built by top industry leaders to pump instant expertise into your agent swarm.