DeepSeek Harness

DeepSeek Harness (dsh) is a plugin-based agent harness where everything is a plugin — models, tools, skills, sessions, storage, the agent loop, and even the UI are all replaceable Cordis plugins, freely recomposed per profile. It ships both as a CLI/Web harness and as a native desktop app (deepseek-harness-desktop).

Community fork: this repository is a community derivative of deepseek-ai/deepseek-harness, maintained by PineSound, adding a desktop app, plugin marketplace, image recognition, and a release channel on top of the unchanged core harness. This is not the official DeepSeek repository. Full list of changes: MODIFICATIONS.md · MODIFICATIONS.zh.md.

---

Proof

---

What it is

An agent harness assembled from plugins. Instead of a monolithic binary, every capability lives behind a documented seam — Service Definition / Provider / Consumer — so you swap tools, skills, storage, terminal, web, or sandbox without touching the loop. Runs are traceable: everything a model sees is reconstructable from the session log.

Why it is different

• Everything is a plugin — models, tools, skills, sessions, storage, loop, and UI are replaceable, recomposable units.
• Traceable by default — model-visible input is reconstructable from the session log, not a black box.
• A native desktop, not just a browser tab — zero-dependency installers for macOS and Windows.
• Open & extensible — a documented seam for every capability; the plugin ecosystem is growing.

Features

Desktop app (native)

• Double-click to launch — no browser tab or URL needed. The Electron shell spawns the real dsh on loopback and opens a native window.
• Zero external dependencies — the installer bundles the full harness runtime, vendored pnpm, and a platform Node.
• Cross-platform — macOS (.dmg) and Windows (.nsis) installers, Chinese menus, branded icons.
• Update mechanism — the About page checks an online releases.json; desktop and publish site share one manifest.

Plugin marketplace

• Visual plugin list with search, status, and enable/disable; required system plugins are guarded.
• Install from npm / tarball / GitHub via the built-in pnpm, with multi-mirror fallback and no-restart hot reload.

Image recognition

• An OpenAI-compatible vision provider (e.g. Alibaba DashScope) fills the gap for text-only models; the vision key / endpoint / model are configured independently of the chat model.

Skill manager

• Manage local skills from the settings UI: list, install, uninstall, enable/disable, and edit descriptions.

---

How it works

The desktop is a thin launcher: it spawns the real dsh (web profile) as a separate system-Node child, parses the readiness URL, and opens a window at that address. Native addons keep their system Node ABI and are never loaded under Electron.

---

Get started

1. Desktop (recommended)

Download the installer for your platform from the publish site — https://deepseek.pinesound.cn/. No install steps; on first open, unsigned builds may need right-click → Open.

2. CLI / Web (npm)

Install Node.js, then run:

npx @deepseek-ai/dsh web

The Web UI is served at http://127.0.0.1:3080 by default.

3. From source

git clone https://github.com/PineKings/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Desktop development

pnpm install
pnpm run build          # builds lib + the web frontend dist
pnpm desktop:dev        # opens the desktop window

Packaging a release

pnpm --filter @deepseek-ai/dsh-desktop run build:harness   # assemble the self-contained runtime
pnpm desktop:pack                                          # electron-builder: .dmg / .nsis
pnpm --filter @deepseek-ai/dsh-desktop run stage-release   # copy installers + manifest into the publish site

Quality gates

pnpm run typecheck
pnpm run lint
pnpm run test
pnpm run test:coverage
pnpm run test:snapshot     # keyless replay of assembled transcripts

---

Repository layout

apps/        cli, web, desktop (Electron shell)
packages/    @deepseek-ai/dsh-<pkg> plugin workspaces
vendor/      vendored Cordis source
docs/        architecture, user guide, cookbook
assets/      readme visuals

---

Community and support

• Publish site: https://deepseek.pinesound.cn/
• Desktop repo: https://github.com/PineKings/deepseek-harness-desktop
• Upstream: https://github.com/deepseek-ai/deepseek-harness
• Tag plugins with the dsh-plugin topic for discoverability.

---

License

MIT. This project is a derivative of deepseek-harness (MIT); the original copyright is preserved. Third-party dependencies and their licenses are disclosed in THIRD_PARTY_NOTICES.md.