DeepSeek Harness is a plugin-based agent runtime with sessions, model adapters, tools, subagents, presets and a browser workspace, and it expects to be composed from a terminal. This project keeps that runtime intact and builds the desktop product around it.

The shell is a real one. Electron starts and supervises a local dsh web Host on loopback with an access token, owns the tray, recovery and installer, and shuts the Host down with the window. Profiles, plugins, settings and session history stay under DSH_HOME, and the desktop application never rewrites user credentials or session data.

The product surfaces are ordinary client plugins mounted into the same slots as upstream ones, so an upstream upgrade does not require forking the shell. Plugin Center and Plugin Discovery show exact version, permissions, compatibility and risk before an install, roll back an unfinished transaction, and enter a safe mode that still lets a user disable or remove a plugin. Preset Square lists built-in and community presets with their skills and requirements; Application Center launches complete applications, starting with a bundled LLM Wiki; local providers cover Ollama, vLLM, SGLang and any OpenAI-compatible endpoint; and vision can route through a compatible provider when the model cannot read images.

Behind the chrome it is the real runtime: workspaces and sessions with a full-conversation outline and full-text jump, plan mode, goals and todos, background jobs, multi-stage workflows, one-shot and continuable subagents with lineage, file and session references, artifacts, and explicit permission and approval surfaces. It tracks upstream dsh 0.2.1-rc.5 and carries its own desktop version.

Read it for the packaging boundary — how to wrap a runtime you do not own in a product you do — and for a plugin install path that treats rollback and safe mode as part of the flow rather than an error page.