DeepSeek Harness Studio

官方网站 · 中文 · English

赋范空间出品 · DeepSeek Harness 的零代码桌面增强

视觉增强 + 本地模型 + 插件市场 + Preset 广场 · 0 代码一键部署和使用

自动发现并推送生态新插件，AI 智能推荐值得安装的能力；无需命令行即可完成搜索、校验、安装、启停与卸载。

下载 macOS arm64 开发预览版 · 下载 Windows x64 开发预览版

点击快速查看功能演示

https://github.com/user-attachments/assets/0717f7c7-a872-4d2b-acc2-3a1c4874c732

Core Features

Lists only in-code, desktop, and user-flow capabilities; future ideas excluded.

Capability	What You Can Do
Desktop workspace and session management	Open local projects via native directory selection, manage and search sessions by workspace, and rename, archive, fork, and resume sessions with history.
Long-session outline and full-text jump	A complete history outline generated by user, assistant, and tool messages appears on the right side of the chat; clicking a summary automatically loads older, unrendered messages, collapses the outline, and jumps to and highlights the full original content.
Plugin discovery and Agent recommendations	Browse curated, recently updated, and ecosystem-trending plugins, filter by scenario or search, or describe your need in natural language and let the Agent filter candidates from the public dsh-plugin catalog.
Trustworthy plugin installation and lifecycle recovery	Verify the pinned version, permissions, compatibility, and risks before installation; uniformly enable, disable, update, and uninstall after installation; incomplete transactions are rolled back automatically, and if the running manifest is still inconsistent, the plugin enters a safe mode that allows disabling/uninstalling.
Preset plaza and seven built-in workflows	Browse Funman official and community Agent Presets, view Skills, tools, and environment requirements before installing, and use them directly in new sessions from the "Installed" list.
Application center and FF–LLM Wiki	Launch full AI applications with their own interfaces, data, and runtime flows from a dedicated top-level entry, and show app sidebar shortcuts on demand.
Multi-model and local inference	Configure DeepSeek and other compatible providers, or connect Ollama, vLLM, SGLang, and custom OpenAI-compatible services from a top-level entry.
Native vision, compatible vision, and image attachments	Process images directly with the DeepSeek multimodal model, or call a verified cloud/self-hosted vision pipeline; image attachments are persisted and sent via a single path.
Plan, Goal, Todo, Jobs, and Workflow	Enter planning mode, manage goals and to-dos, view background jobs in the current process, and review multi-stage Workflow member statuses within the conversation.
SubAgents and multi-Agent collaboration	Create one-shot or continuable sub-Agents, view parent–child session lineage, running status, and duration, and continue conversations or stop the current turn in supported sub-sessions.
Project rules, context references, and output files	Read repository instructions, use @file／@session to reference context, and view, open, or locate the files the Agent actually produced at the end of a response.
Permissions, sandboxing, and human confirmation	Choose read-only, workspace-write, or full access for the current or future sessions; dangerous permissions, tool approvals, and proactive Agent questions are all confirmed explicitly in the UI.
Theme skins and cross-platform desktop delivery	Switch built-in or local backgrounds with automatic interface color adaptation; download macOS arm64 and Windows x64 preview builds from GitHub Releases.

Recent Roadmap

These lack full first-class entry points; not counted in current features.

Direction	Product capabilities planned to fill the gap
Standalone capability center	Provide a separate discovery, connection, and project-level composition management path for MCP servers, Skills, and tools that are not wrapped in a Bundle.
Visual Agent orchestration	On top of the existing Preset and SubAgent runtime capabilities, offer a custom Agent, role assignment, and team workflow editor.
Remote control and automation	After establishing clear permission and audit boundaries, complete browser/desktop operations, mobile continuation, and message notification entry points.

Project Overview

DeepSeek Harness Studio uses Electron to host the DeepSeek Harness web workspace, while the desktop main process launches and manages a local dsh web service. This repository provides a full source-code development environment: users can clone or download the code from GitHub, install dependencies locally, edit the source, launch the desktop app, and continue development.

Desktop installers are released exclusively through this repository's GitHub Releases, not via third-party download sites. Verified macOS arm64 preview ZIPs and Windows x64 preview installers have already been provided through real Electron acceptance tests; when continued development is needed, the full source code can still be obtained and launched locally.

Workspace and Agent Execution Capabilities

• Workspace and sessions: native local directory selection; register, group, search, and delete by workspace; sessions support renaming, archiving, and forking at the last completed turn.
• Long-session outline: a lightweight outline on the right covers the full history of the current session while the main text still loads page by page; outline summaries are capped at 80 characters, clicking an unloaded item auto-paginates continuously, and after locating it the panel collapses and the full message is highlighted.
• Planning and work management: organize the current task via Plan, Goal, and Todo; the Jobs panel shows background jobs within the current process, and after a process restart these running jobs are no longer treated as live tasks.
• Workflow and SubAgent: the conversation log displays Workflow phases, members, and outcomes; the SubAgent outline supports parent–child lineage navigation, continued conversation, and stopping the current turn of a running, continuable sub-session.
• References and deliverables: @file and @session attach files or sessions as context; successfully produced files appear at the end of the response and can be opened via the local Host or located in the folder.
• Human–machine collaboration: the Agent can initiate structured single-choice, multi-choice, or custom questions; tool approvals, full access, and plan reviews all require explicit user input in the UI.
• Security boundaries: permission presets pin the sandbox mode and approval policy to a session; credentials are saved through a write-only interface, and the page never reads back or displays stored key values.

DeepSeek Harness v0.1.1-rc.2 Compatible Capabilities

Studio 0.1.0-rc.19 integrates the core and web capabilities of DeepSeek Harness 0.1.1-rc.2, while retaining Funman's plugin center, plugin discovery, Preset plaza, application center, theme skins, and desktop recovery pipeline. Studio version numbers and upstream Harness version numbers are managed separately; the download links at the top of the page match this version.

• Multimodal capabilities: retains Pro/Flash text models, integrates DeepSeek-V4-Flash-Vision-Exp, persistent image attachments, and Files API image reuse; expired references undergo bounded retransmission, and on parse failure the entire request falls back to a restricted inline image.
• Agent runtime capabilities: integrates @ file/session references, Plan, Goal, background Jobs, Workflow, SubAgent, concurrent web search, and persistent PowerShell PTY on Windows.
• Desktop adaptations: the Host uses --no-open, preserves native directory selection, plugin transaction recovery, and existing user data directories; when historical plugin lock files are incompatible, it automatically enters a compatibility recovery that does not rewrite the lock file.

Plugin Ecosystem: Discover, Install, and Manage

Plugin Discovery: Not Sure What to Install? Start Here

Not sure which plugins to find or what's trending? Enter Plugin Discovery from the left; the app reads the online catalog and organizes plugins into a browsable, actionable page.

_{真实 Desktop 界面：目录精选、最近更新、生态热门、场景分类、搜索以及安装与管理入口。}

• New finds daily: see catalog picks, recent updates, and trending plugins right away—no need to search repos one by one.
• Quick scenario filters: covers Agent & workflows, Web UI, browser & search, vision & media, memory & context, models & services, dev tools, integrations & notifications.
• Search directly: look up by plugin name, feature keyword, or author, and view avatar, description, version, and update time.
• Use it right away: uninstalled plugins go straight into the secure install flow; installed plugins can be managed in one click from the Plugin Center.

Not Sure of the Exact Package Name? Let the Agent Filter for You

When you only know something like "I want a desktop pet," there's no need to guess the npm package name first. Enter a natural-language description in Plugin Discovery, and the app passes it to the current Agent as a /find-plugins request; the Agent loads its built-in skill, queries the public dsh-plugin catalog read-only, and returns the most relevant candidates with versions, authors, update times, and matching reasons to the current conversation.

_{真实 Desktop 验收：对话发出“找一个桌面宠物插件”，Agent 加载 find-plugins、执行公开目录搜索，并从 8 个结果中列出 5 个相关候选。}

• No keyword memorization: just state your goal, use case, or the problem to solve.
• Verifiable rationale: results include exact package name, version, publisher, update time, and per-item match reasons.
• Separate search and install: recommendations reflect only public catalog metadata; after picking a package, compatibility checks and install confirmation still go through the Plugin Center.

Plugin Center: Online Install, Toggle, and Remove

_{真实 Desktop 界面：插件头像、公开目录、已安装区域、“安装”按钮与三点管理入口。}

Once you've chosen a plugin, head to the Plugin Center, where you can look up plugins and Skill Packs published to the npm public registry by short package name, full npm package name, or explicit GitHub repository. dsh-plugin serves only as a discovery signal; GitHub is also used solely to map already-published npm packages—Studio never installs repository source code directly. A pinned version must still pass Bundle, integrity, and runtime compatibility validation.

• Online Discovery: Search public plugins; view version, capabilities, permissions, compatibility, and risk notes.
• One-Click Install: Download a pinned version, verify package identity, integrity, and Bundle manifest; after confirmation, auto-install and restart Harness Host to verify runtime.
• Installed Management: Centrally view system, public catalog, and local sources; use the three-dot menu to enable, disable, update, or uninstall plugins.
• Safe Removal: Uninstall retains config and plugin data by default; when data cleanup is needed, the user confirms it separately.

Preset Plaza Is Live: One-Click Install of Complete Working Styles

Plugins typically solve "give the Agent one more tool," and Skills solve "teach the Agent how to do things a certain way," while an Agent Preset solves a more complete problem: it assembles a role, working rules, Skills, Plugin/MCP, and standard Harness tools into a reusable working style. Users don't need to understand and configure each piece by hand—after installing a Preset, they can immediately create new sessions with the corresponding role.

Capability Layer	What It Is	Main Problem Solved
Skill	Reusable methods, steps, and constraints	Tells the Agent "how to" handle a class of tasks
Plugin / MCP	Executable tools or external service connections	Enables the Agent to actually read/write systems, call services, and complete actions
Agent Preset	A combination of role, Skills, tools, and runtime rules	Assembles scattered capabilities into an out-of-the-box digital employee or workflow

Source provides Preset Plaza at the same level as "Plugin Center" and "Plugin Discovery," completing the desktop loop: discovery, details, install, management, new-session use, delete, reinstall.

_{真实 Desktop 界面：Preset 广场、赋范官方内置目录、搜索与排序，以及安装、查看详情和用于新会话入口。}

Usage path: Discover Preset → view capability composition and prerequisites → one-click install → select "Use in New Session" from "Installed" → complete tasks via the workflow → delete or reinstall at any time.

1. Navigate to Preset Plaza from the left menu; search by use case, name, or author, and sort by download count or update time.
2. Open the details to review the role, Skills, tools, external dependencies, and source notes it contains.
3. After clicking install, Desktop validates the source, size, digest, and archive path, then writes to the local user Preset directory.
4. No Host restart is needed after installation; click "Use in New Session" in Installed to start a task with the corresponding Preset.
5. User Presets can be deleted and reinstalled; system Presets remain protected. After installing or deleting, you stay on the current page without interrupting browsing.

Install, delete, and "Use in New Session" involve local files and the Host, and are executed only in Desktop; browser dev mode is for quick UI inspection and acceptance and does not modify local Presets.

Funman Official Built-in Presets: Seven Real-World Scenario Workflows

The built-in content is not a "plugin bundle" but seven Agent Preset + Skills + tool integrations organized around real deliverables. In total it includes 7 Agent Presets, 9 Skills, and 1 image generation plugin; the Feishu digital employee additionally integrates Feishu MCP and time-parsing MCP, and PPT Office uses a built-in motion runtime adapter.

Naming note: "Funman Official" means built in and maintained by the Funman desktop development team; it does not represent the official DeepSeek Harness. After installation they remain regular user Presets and can be deleted and reinstalled.

Category	Built-in Preset	Capability Composition	Direct Deliverable
Product & app development	AI WebApp	1 Preset + 3 Skills	A runnable web product from requirement clarification and spec organization to TDD acceptance
Office & presentations	PPT Office	1 Preset + 1 Skill + motion runtime adapter	An 8-page, four-theme, interactive single-file HTML presentation
Vision & media	Video Generation	1 Preset + 1 Skill	A 16:9 MP4 from one-line research and storyboarding to rendered output
Content production	Content Factory	1 Preset + 1 Skill + 1 image generation plugin	1–10 style-consistent image-text cards from long-form analysis
Data analysis	AI Report	1 Preset + 1 Skill	A verifiable offline interactive report generated from local Excel
Enterprise collaboration	Feishu Digital Employee	1 Preset + 1 Skill + Feishu MCP + time-parsing MCP	Real Feishu tasks and dual-end receipts from natural-language instructions
Knowledge base R&D	LLM Wiki Producer	1 Preset + 1 Skill	Use the "LLM Wiki Full-Stack Engineer" to develop, validate, and deliver an enterprise knowledge base project phase by phase

The six cases below all use 16:10 preview images to avoid banners and tall images breaking the layout; click any image to view the full original.

Products & Presentations: From Idea to Deliverable

01 · AI WebApp 解决什么：需求讨论结束后，仍不知道如何稳定落成一个可运行、可测试的 Web 产品。 功能逻辑：需求澄清 → 规格整理 → TDD 纵向开发 → 类型检查、构建与浏览器验收。 能力组成：grill-me、to-spec、tdd 3 个 Skills + 专属产品开发 Preset。 示例结果：可启动、可拖拽、可更新数据的“创作指挥舱”WebApp。	02 · PPT Office 解决什么：只有主题和内容大纲时，快速形成结构完整、视觉统一、可离线播放的演示成果。 功能逻辑：保存原始内容 → 组织 8 页叙事大纲 → 选择主题并生成 → 确定性检查页数、动画与离线性。 能力组成：1 个动效演示 Skill + 专属导演 Preset + 内置动效运行适配器。 示例结果：四套主题、逐页语义动画的单文件 HTML；当前明确不生成 .pptx。

Content & Media: Turn One Line or a Doc into a Finished Product

03 · 视频生成 解决什么：只有一句产品想法时，仍能建立从事实调研到正式成片的完整生产链。 功能逻辑：四维 Web 检索 → 研究与简报 → 设计规范、分镜和视频源码 → HyperFrames 检查 → FFmpeg 渲染与媒体验收。 能力组成：1 个产品发布片 Skill + 专属视频导演 Preset。 示例结果：可编辑视频工程、联系表和 16:9 MP4；运行环境需要 FFmpeg 与 ffprobe。	04 · 内容工厂 解决什么：把一篇长文稳定拆解成风格统一、可以连续发布的系列视觉内容。 功能逻辑：分析受众与观点 → 确认风格、布局和配色 → 编写系列大纲与逐图提示 → 生成 1–10 张图片 → 校验尺寸、页数与一致性。 能力组成：1 个图文 Skill + 专属内容视觉 Preset + Codex ImageGen Bridge Plugin。 示例结果：9 张连续图文卡片；真实生图前需要本机 Codex CLI 已登录并具备 ImageGen 能力。

Data & Enterprise: From Local Data to Real Business Actions

05 · AI 报表 解决什么：把本地 Excel 从“原始表格”转成结构清晰、可追溯、可直接交付的数据叙事。 功能逻辑：只读核对工作簿与数据基线 → 清洗聚合 → 选择图表叙事 → 内联 ECharts 生成 → 桌面与移动双端验收。 能力组成：1 个 Excel 报表 Skill + 专属数据分析 Preset。 示例结果：不依赖 CDN 的单文件交互 HTML，原始 Excel 保持只读，报告只嵌入聚合数据。	06 · 飞书数字员工 解决什么：让对话不止返回文字，而是把一句自然语言真正变成企业协作系统中的任务。 功能逻辑：Harness Desktop 或飞书接收指令 → Skill 判断动作 → 时间 MCP 解析“明天下午 6 点” → 飞书官方 MCP 创建任务 → Harness 与飞书双端返回回执。 能力组成：1 个飞书 Skill + 专属数字员工 Preset + 飞书 MCP + 时间解析 MCP。 示例结果：真实飞书任务、负责人和截止时间均可查看；首次使用需要配置飞书应用凭据与默认负责人。

All images above are real outputs from six workflow case studies. Desktop ships only the slimmed Presets, Skills, and tool adapters needed to run, not case source code, input data, screenshots, or generated artifacts.

App Center Is Live: From Capability Extensions to Full AI Apps

Plugins add tools or UI capabilities to the Harness, Presets combine a role, Skills, and tools into a working style, and apps target more complete product scenarios: each comes with its own interface, dedicated data, and runtime flow. The App Center sits at the same level as Plugin Center, Plugin Discovery, and Preset Plaza—users don't have to wade through plugin listings to find technical packages; they discover and launch complete products from a single stable entry.

Form	Primary Role	Typical Usage
Plugin / MCP	Adds callable tools, service connections, or partial UI capabilities	Triggered by the Agent or Harness features after installation
Agent Preset	Combines role, Skills, tools, and working rules	Create a new session with the Preset and complete tasks according to the established workflow
AI Application	Provides a dedicated interface, data, and end-to-end business flows	Open directly from the Application Center and continuously manage your own business data

_{真实 Desktop 界面：应用中心作为一级入口展示赋范官方应用；可直接打开应用，也可选择是否把快捷入口显示在左侧导航。}

First Built-in Application: FF–LLM Wiki Enterprise Knowledge Base

FF–LLM Wiki tackles scattered enterprise documents, hard-to-organize knowledge relationships, and untraceable Q&A results, progressively turning raw material into searchable, linkable, and verifiable knowledge assets. It is not a demo card embedded in the chat page but a full application whose local runtime is managed by Desktop and opens in the system browser.

_{真实运行界面：从文档解析、Wiki 编译和知识图谱构建，到基于检索证据的 Agent 问答与质量评估。}

• Complete knowledge pipeline: document parsing → entity and entry extraction → Wiki and knowledge graph construction → vectorization and indexing → traceable Q&A.
• Verifiable sources: Q&A uses only the evidence snippets currently retrieved and retains source citations; when the model is unavailable it explicitly degrades to local search results.
• Local data isolation: documents, Wiki, graph, and SQLite data are written to an application-dedicated directory under DSH Home, not mixed into the current project workspace.
• Keys in the user's hands: the application does not carry the Funman team's API Key. Knowledge compilation, browsing, and local search require no model key; before using DeepSeek generative RAG, users must configure their own DEEPSEEK_API_KEY in the Harness credential center.
• Entry points shown on demand: the Application Center is always accessible; the FF–LLM Wiki left-side shortcut is off by default, and users can toggle it via "Show in Sidebar" at any time.

Usage path: Open the Application Center → select FF–LLM Wiki → click "Open Application" → import or manage enterprise material → build the Wiki and knowledge graph → verify sourced answers in smart Q&A.

Naming note: "Funman Official" means built in and maintained by the Funman desktop development team; it does not represent the official DeepSeek Harness.

Built-in Skins & Free Reskinning

Go to Settings → Background to switch built-in skins; when you choose a custom image, the app performs 1920×1080 WebP cropping and interface color adaptation locally without uploading the original.

大肥鱼拟人 · 默认 蓝白鲸灵助手与明亮宫殿，中央留白适配对话区。	云端猫咪 保留原有柔和蓝白猫咪主题，清爽、安静、低干扰。

Models, Permissions & Thinking Modes

• Permission Selection: The input area uses three Chinese permission tiers — 只读, 工作区写入, and 完全访问 — that apply to the current session; general settings only determine the default permission for new sessions, and full access requires confirming the risk before enabling.
• Model & Thinking Mode: The model and API key are still managed in the settings page; the model selector on the right shows the current main model, and clicking "Vision Enhanced" switches it directly to DeepSeek-V4-Flash-Vision-Exp.

Local Models & Self-Hosted Inference

Studio supports connecting to local or LAN model services via the standard OpenAI-compatible interface. Users must start the inference service and prepare the model first; Studio will not auto-download models, consume extra disk for weights, or manage GPU runtime parameters.

Framework	Default API Base	Connection Method
Ollama	http://127.0.0.1:11434/v1	Use Ollama's OpenAI-compatible interface and enter the actual model ID; the local API Key can be left blank.
vLLM	http://127.0.0.1:8000/v1	Connect to a vLLM OpenAI-compatible server and enter the loaded model ID and an optional API Key.
SGLang	http://127.0.0.1:30000/v1	Connect to a SGLang OpenAI-compatible endpoint and enter the started model ID and an optional API Key.
Custom service	Entered by user	Supports other HTTP(S) services that are compatible with /v1/chat/completions.

• General chat models: open Settings → Models → Add Local Model and directly choose Ollama, vLLM, SGLang or OpenAI-compatible; the API Base is prefilled, and the model ID can be read or entered manually, with the local API Key being optional.
• Local vision models: select Ollama, vLLM, SGLang or a custom service from the "Vision Enhancement" configuration; the configuration is only saved after the image passes validation.
• No silent cloud offload: when a self-hosted route fails, images are not automatically forwarded to Bailian, OpenRouter or other cloud providers.

Vision Enhancement: Let DeepSeek Understand Images

DeepSeek-V4-Flash-Vision-Exp is an image-text model separate from Pro/Flash; it accepts both text and images and reuses the existing DeepSeek API Key from the settings page. It is not a side channel that returns image descriptions for Pro/Flash; before sending an image, the current session must be switched to this model. Pure text models can still use the pre-configured and validated Bailian, OpenRouter, Ollama, vLLM, SGLang or custom OpenAI-compatible vision services.

• One-click: the input box keeps a single "Vision Enhancement" button on the left; clicking it switches the main model name on the right to DeepSeek-V4-Flash-Vision-Exp.
• Clear labels: the model list marks image-capable models with "Supports Images" only; the settings page shows it as a "Vision Enhancement Model", no longer mixed with the default text model.
• Native image pipeline: the DeepSeek Files API reuses uploaded images; expired references only trigger bounded retransmission, and if file parsing fails the whole request uses the same limited inline image, no duplicates.
• Compatible vision pipeline: self-hosted services accept an API Base, vision model ID and optional API Key; on failure it will not fall back to sending images to the cloud.
• Explicit off: when the switch is off, images are excluded from the model's visible context; attachment history remains in the UI.
• Dev scenarios: understands product screenshots, error screens, design drafts, data charts, photos and in-image text, and can read workspace images by path.

Download Desktop

GitHub Releases now provides a macOS Apple Silicon preview ZIP and a Windows x64 preview installer, both validated through a real Electron acceptance test. Running the desktop version does not require installing Node.js or pnpm separately. All of these are currently development preview assets; the official release will still provide a platform-signed macOS .dmg and Windows x64 .exe.

下载 macOS arm64 预览版 · 下载 Windows x64 安装程序

macOS arm64 build

After downloading and extracting the preview ZIP, drag DeepSeek Harness.app into the "Applications" folder. Since this preview is not yet Apple-notarized, run the following in "Terminal" before first launch:

xattr -dr com.apple.quarantine "/Applications/DeepSeek Harness.app"
open "/Applications/DeepSeek Harness.app"

If the app is not in "Applications", replace the path with the actual one. This command is for preview packages from this repo's GitHub Releases that have been SHA-256 verified only; do not use it on apps of unknown origin. After the first successful launch, start it from Finder or the Dock as usual.

Windows x64

Download DeepSeek-Harness-Desktop-Windows-x64-0.1.0-rc.19-Setup.exe and run the installer directly. The public download section of the Release only keeps the macOS ZIP and the Windows installer; checksum files, blockmap and platform acceptance records are retained in the corresponding GitHub Actions builds to prevent ordinary users from accidentally downloading development files.

The development preview uses a separate Pre-release tag and does not trigger a formal installer release. The formal process only accepts desktop-v* tags that exactly match the Desktop version; after macOS and Windows installers each complete platform signature verification, GitHub will publish the installer files and SHA256SUMS simultaneously.

Quick Start

获取源码

Clone the repository with Git:

git clone https://github.com/fufankeji/deepseek-harness-studio.git
cd deepseek-harness-studio

You can also choose Code → Download ZIP on the GitHub repository page, then download and extract the source code and enter the project directory.

环境要求

• Node.js ^22.19.0 || >=24.0.0
• pnpm 11.7.0

外部服务准备

下载源码、安装依赖和启动桌面开发环境不需要预先填写 API 密钥。需要在应用中实际调用模型时，再在设置中配置所选模型服务与凭证；凭证不要提交到 Git。

Installation and Launch

安装工作区依赖：

pnpm install

构建所需模块并启动桌面开发环境：

pnpm run dev:desktop

开发启动器会在相关源码或构建输入变化时重新构建；需要强制完整重建时运行：

pnpm run dev:desktop:rebuild

目录结构

deepseek-harness-studio/
├── apps/
│   ├── desktop/       # Electron 主进程、preload、Host 生命周期与桌面构建脚本
│   ├── web/           # DeepSeek Harness Web 界面入口与桌面端组合
│   └── cli/           # dsh CLI、运行配置与 Agent Preset
├── packages/          # Agent、模型、工具、会话、插件和客户端能力包
├── native/            # 原生沙箱辅助模块
├── python/            # Python SDK 与相关运行时
├── examples/          # 可运行示例与配置
├── scripts/           # 构建、检查、生成和发布脚本
├── website/           # 项目文档站源码
├── vendor/            # 固定版本的 Cordis 基础源码
└── assets/            # README 使用的项目图片

常用开发命令

Command	Purpose
pnpm run dev:desktop	Build the required modules and start the Electron desktop app
pnpm run dev:desktop:rebuild	Force a full rebuild, then start the desktop app
pnpm run build	Build the Host, client, Web and desktop versions
pnpm run package:desktop	Generate an unbundled desktop app for the current platform
pnpm run typecheck	Run TypeScript type checking
pnpm run test	Run Vitest unit tests

Suggested Reading Order

1. apps/desktop/src/main.ts: desktop app entry point, windows, tray and local Host composition.
2. apps/desktop/src/host-supervisor.ts: startup, readiness detection and exit management of dsh web.
3. apps/desktop/src/preload.ts: fixed desktop interfaces accessible to the Renderer.
4. apps/web/: the Web workspace loaded in the desktop window.
5. apps/cli/ and packages/: the CLI composition and implementations of each Harness capability.

Relationship with DeepSeek Harness

This project continues desktop development based on the Harness core, Cordis plugin system and Web interface of deepseek-ai/deepseek-harness. This repository maintains the Electron desktop entry point, local Host management, desktop interactions and accompanying development scripts.

License

This project is licensed under the MIT License. For license information of third-party components, see THIRD_PARTY_NOTICES.md.