English | 中文
DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.
It is built on an everything-is-a-plugin architecture and powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability.
Documentation: https://deepseek-harness.github.io/deepseek-harness/
This repository is a customized fork of the official DeepSeek Harness: it tracks the official release as its base and layers the star-domain multi-agent routing, the A2A gateways and their local stack, the Web GUI polish, the model-facing whiteboard tools, and the Pi backends described in the next section on top. Everything stays a Cordis plugin, so the additions mount by composition and the official core remains intact.
| Package | What this fork adds |
|---|---|
@deepseek-ai/dsh-xingchen |
启明 (Qiming) routes a session to three specialist seats — 天权 (architecture and code review), 瑶光 (bug reproduction), 天梁 (delivery planning) — through /review, /bug and /planning, as delegated child agents or over A2A peers; the seats' answers enter the transcript as assistant-role messages. See star-domain routing. |
@deepseek-ai/dsh-a2a |
The A2A client: protocol types, direct HTTP dispatch, and the Kafka bus channel that carries tasks and replays events. |
@deepseek-ai/dsh-a2a-host |
The A2A host endpoint this deployment publishes for inbound peers. |
@deepseek-ai/dsh-tool-a2a |
Model-facing peer tools (a2a_peers, a2a_send) for conversational dispatch. |
deploy/a2a/ |
The local stack: a 3-broker Kafka cluster, the three A2A gateways vendored into this repository, and the Web app, started by one command with skip-if-running semantics. |
| Package | What this fork adds |
|---|---|
@deepseek-ai/dsh-pi-agent-loop |
The Pi coding agent as a second session backend beside the native loop, selected per session. |
@deepseek-ai/dsh-subagent-pi |
A subagent provider that delegates a task to the Pi coding agent over its RPC mode. |
@deepseek-ai/dsh-llm-pi-ai |
The provider routing and model-catalog resolution the Pi backends use. |
| Package | What this fork adds |
|---|---|
@deepseek-ai/dsh-client-ui-polish |
Web GUI polish: a whole-app background image (10MB upload cap, structural surfaces turned transparent behind it), a session stats float pricing each settled message at its model's rate against an editable rate card, in-place Git and LaTeX panels in the conversation view, an embedded Excalidraw whiteboard tab sharing one scene with the model's tools, and a configurable automatic-compaction threshold. |
@deepseek-ai/dsh-tool-excalidraw |
Model-facing whiteboard tools — excalidraw_read, excalidraw_write, excalidraw_draw, excalidraw_export — over the workspace scene file the canvas tab renders. |
@deepseek-ai/dsh-client-ui-ssh plus @deepseek-ai/dsh-remote-ssh, @deepseek-ai/dsh-remote-fs-sftp and @deepseek-ai/dsh-remote-subprocess-sftp |
SSH/SFTP workspaces: an interactive terminal panel in the Web UI and remote filesystem and subprocess providers over one SFTP seam. |
deploy/brand/ |
Background artwork for the Web app, with the light-theme readability constraints the layout imposes and the prompts behind each candidate. |
- Plugin-first architecture. Every capability — the agent loop, tools, sandboxing, storage, the Web UI — is a Cordis plugin composed from
cordis.ymlpatch layers. A deployment picks its stack by composition, not by forking. - Multiple runtime surfaces. The same composed tree powers the Web UI, the CLI, one-shot headless tasks, an ACP automation server, a JSON-RPC SDK, and the Python SDK.
- Full agent stack. Sessions with durable JSONL persistence, system prompts, a tool registry, the agent loop, subagents, background jobs, workflow definitions, and automatic context compaction.
- Sandboxed execution. Bash/pwsh shells, a code-execution runtime with Code Mode, and a process-confinement seam backed by bwrap, Landlock, and Seatbelt — each guarded by per-session approval and sandbox policy.
- Workspace-native tools. Filesystem read/write/edit with observation policy, git workflows, Excalidraw whiteboards shared between the model and the UI, web search and fetch, LSP, and a skill registry.
- Session intelligence. Projection read models (session stats, token usage), SQLite full-text search over history, lineage and relationship queries, and session log export.
- Extensible surfaces. Agent presets for per-session composition, installable
dsh --profilebundle patch layers, hooks bridges for Claude Code and Codex, MCP server integration, and self-modifying extensions. - User plane. User settings with a file backend, credential references, human feedback, goals, plan mode, and interactive approval flows.
See packages/README.md for the package inventory and docs/architecture.md for how the pieces compose.
DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.
Review the safety notice before running the project.
Install Node.js, then run:
npx @deepseek-ai/dsh webThe command starts the Web UI at http://127.0.0.1:3080 by default and opens it in the default browser for a local launch. An SSH launch only prints the host URL because the SSH client or editor owns the local forwarded address. Pass --no-open to run the server without opening a browser. See Web UI guide.
This path installs the official build; the customizations in this fork are available only from this repository's source.
To run from a repository checkout:
git clone https://github.com/ReachForStar/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh webpnpm run build prepares the repository artifacts. pnpm dsh web runs the built entry and uses those artifacts without rebuilding; pnpm run dsh:source runs apps/cli/src/bin.ts through tsx instead, for launcher-only work.
To bring up the whole local review/bug/planning stack — the Kafka cluster, the three A2A gateways vendored under deploy/a2a/gateway, and the Web app — use the stack launcher. It starts only what is not already running:
cd deploy/a2a/gateway && npm ci && npm run build
cd ../../.. && pnpm run stack:uppnpm run stack:status reports each component and pnpm run stack:down stops what the launcher started (add --kafka to stop the cluster). See deploy/a2a/README.md.
- Submit feedback or bug reports through GitHub Discussions.
- Add the
dsh-plugintopic to your plugin repository for discoverability. - Join DeepSeek Harness Discord community.
See CONTRIBUTING.md.
Start with the development guide and architecture documentation.
pnpm run dev:web builds, serves, and rebuilds client bundles on source edits in one terminal, and make help lists the matching Make targets for Web and Desktop; the guide's application commands section owns the full table.
For agents, follow AGENTS.md.
@misc{deepseek-harness2026,
title={DeepSeek Harness: Everything is a Plugin},
author={DeepSeek-AI},
year={2026},
publisher={GitHub},
howpublished={\url{https://github.com/deepseek-ai/deepseek-harness}},
}Third-party dependencies and their licenses are disclosed in THIRD_PARTY_NOTICES.md.