DeepSeek Harness Open-Source Review: Surprising Insights Beyond Its Plugin System

DeepSeek Harness (dsh) has been open‑sourced and quickly attracted over 55 K stars; its slogan "Everything is a Plugin" reflects a design where model visibility, transparency, and the entire agent lifecycle—including thinking, prompts, context, tool calls, skills, sub‑agents, token consumption, and cache usage—are fully recorded and exposed.

The Trajectory view records the complete workflow of an agent, using color coding (e.g., orange for tool calls, green for context) and displays these elements at the top of the UI, making the agent’s execution path instantly observable.

In a test run the author observed 1 round with 57 steps, LLM processing time of 9 m 53 s, tool‑call time of 9 m 44 s, average first‑token latency of 1.8 s, throughput of 53 tok/s, cache hit rate of 96 %, 3.1 M input tokens and 26.1 K output tokens. This mirrors DeepSeek R1’s chain‑of‑thought transparency but extends it to the whole execution pipeline, turning model reasoning and agent actions into an auditable event stream.

Installation is straightforward via npm: npx @deepseek-ai/dsh web. The Web UI offers four modes: Standard, PTC, Minimal, and Creation.

Minimal mode provides only persistent bash and a str_replace_editor tool, with a simple system prompt "You are a helpful software engineer assistant"; the response quality is modest and the model does not recognize itself as DeepSeek Harness.

Standard mode delivers a full‑featured coding agent supporting file editing, shell execution, web and file retrieval, skills, planning, goals, sub‑agents, and workflow orchestration.

Under the hood, dsh is built on the Cordis plugin framework (vendored into the repository). Cordis defines five core concepts: Service implementation, context containers ( ctx.tools, ctx.llm, ctx.sessions), dependency declaration via inject, typed events for communication, and reversible registration (installable and removable).

All parts of dsh—model adapters, tool registries, session logs, and the agent loop—are treated as interchangeable plugins, configurable through patches. The runtime plugin tree stacks from the base dsh-base (model, tools, persistence, sandbox, settings, credentials, telemetry) to either dsh-web-app or dsh-headless, then to user‑provided cordis.patch.yml and --patch overlays. The command dsh --profile web --dump-config prints the effective configuration, which can be fully overridden by custom patches.

The "capability seam" introduces a three‑role structure: Service Definition (e.g., dsh-shell), Service Provider (e.g., dsh-bash-local or dsh-bash-sandbox), and Consumer (e.g., dsh-tool-bash). Swapping a Provider replaces the entire product behavior; for example, moving Bash, PTY, and LSP capabilities to a remote sandbox requires no code changes, only a new Provider implementation.

Community plugins are now available, such as dsh-web-ui (a collection of UI components and skins for the DSH Web UI) and dsh-genui (adds an interactive UI layer to model responses). Repository links are provided for these extensions.

Further reading links cover agent orchestration, graph‑based engineering, and recent surveys on Loop + Harness agents.