DeepSeek Harness Review: Everything Is a Plugin (dsh)

Key highlights:

• Absurd launch velocity — ~50,000 GitHub stars in 12 hours, ~92,000 by hour 28, and 186,000+ stars with 20,600+ forks as of August 23, 2026.

• Loop-as-plugin — `packages/core/agent-loop` is an ordinary package you swap from config. Changing Codex CLI's loop means editing its Rust core; changing dsh's means mounting a different plugin.

• Built on Cordis, a plugin framework with four years of production use inside the Koishi chatbot project.

• Append-only session logs with a runtime-enforced invariant: *model-visible means logged*.

• Model-agnostic — ~40 providers, and subagents can be delegated to a competitor's agent.

Most coding agents share a shape: a core agent loop — take input, call the model, run tools, repeat — written as privileged code, with a ring of extension points around it (MCP servers, skills, hooks). The core is framework-private; extensions observe rather than participate.

dsh inverts that. Underneath sits Cordis, a plugin framework where plugins contribute services, typed events, and reversible effects to a shared context; registrations unwind when their plugin unloads. The architecture doc puts it bluntly:

> There is no privileged core to patch: you extend dsh by mounting a plugin beside the others.

OpenClaw, the previous velocity record holder, took 84 days to reach 200,000 stars. dsh did ~92,000 in 28 hours and is past 186,000 inside two weeks.

But stars are a hype metric, not adoption. What makes this significant is who shipped it. A frontier model lab open-sourcing its entire harness layer under MIT — decision records included — argues that the harness should not be a paid product at all. That is a shot at the business model, not at Claude Code's feature list.

It arrived alongside the GA release of DeepSeek-V4-Pro-0813 (1.6T parameters, 49B activated per token, 1M context), which added native OpenAI Responses API support and Codex integration. DeepSeek simultaneously moved its API from flat pricing to peak/off-peak rates — a genuine increase. Some negative sentiment around the launch is actually about that price change, not the code.

A running dsh is a plugin tree composed at boot from ordered layers. A profile lists the bundles it stacks and keeps your own cordis.patch.yml (web and headless ship as templates). A bundle distributes Cordis config rows plus the code they mount, so whatever it inserts stays patchable from above.

dsh-base is the first layer of every profile — model adapters, tools, persistence, sandbox and approval policy, credentials, telemetry. dsh-web-app adds the browser app; dsh-headless a one-shot runner with no server. To dump the tree your machine actually boots:

Every row it prints can be replaced by a patch of your own. The core packages contributing to that tree:

This is a condensed version. The full article includes architecture diagrams, code examples, comparison tables, and practical use cases.

👉 Read the complete analysis on andrew.ooo

Topics: deepseek-harness, dsh, deepseek, coding-agent, agent-harness