openclaw-plugin-dsh

What it does

openclaw-plugin-dsh replaces the turn executor inside OpenClaw with an embedded DeepSeek Harness (DSH) agent. When a session routes to a dsh/... model ref, your message no longer goes through OpenClaw's provider transport — it is handed to a live DSH Agent running in the same process, with the full DeepSeek Harness stack underneath:

• its own model route and reasoning effort (from your DSH settings.yaml),
• its own tools — bash, file search, editing, subagents, workflows, goals, web search, and more,
• its own session persistence and compaction,
• streaming deltas bridged back into OpenClaw's reply pipeline.

OpenClaw keeps what it is good at — channels, routing, sessions, the visible transcript — while DSH owns the agent loop. It is the same split OpenClaw uses for its own Codex harness, now available for the DeepSeek Harness.

Why you want it

Quick start

# 1. Install (ClawHub, recommended):
openclaw plugins install clawhub:openclaw-deepseek-harness
# or from a source checkout: npm install && npm run build first
openclaw plugins install <path-to-this-package>

# 2. Point an agent at the harness
openclaw config set agents.defaults.model "dsh/deepseek-v4-pro"

# 3. Talk to it
openclaw agent --local --agent main --message "Hello from DeepSeek Harness"

That's it. dsh/<any-model-id> refs are accepted — the actual model is chosen by the DeepSeek Harness itself (its agent-default-model route), so you never maintain a second model catalog.

How it works

OpenClaw gateway process
┌───────────────────────────────────────────────────────────────┐
│  channel / session / routing / transcript        (OpenClaw)   │
│                        │                                       │
│                        ▼  prepared attempt                     │
│   ┌─────────────────────────────────────────────┐             │
│   │        dsh agent harness (this plugin)      │             │
│   │  supports() · runAttempt() · reset()        │             │
│   └───────────────────────┬─────────────────────┘             │
│                           │                                    │
│   ┌───────────────────────▼─────────────────────┐             │
│   │        embedded DeepSeek Harness host       │             │
│   │  dsh-base bundle · model route · sandbox    │             │
│   │   ┌─────────┐  ┌─────────┐  ┌────────────┐  │             │
│   │   │ Agent A │  │ Agent B │  │ Agent …    │  │             │
│   │   │ session │  │ session │  │            │  │             │
│   │   └─────────┘  └─────────┘  └────────────┘  │             │
│   └───────────────────────┬─────────────────────┘             │
│                           │ session/event firehose             │
│   streaming deltas · reasoning · tool results ───────────────▶ │
│                          reply pipeline                        │
└───────────────────────────────────────────────────────────────┘

1. Boot once. On the first claimed attempt, the plugin boots one DSH core tree (the dsh-base profile bundle plus a small patch layer: HMR off, persona, sandbox root) through @deepseek-ai/dsh-app-boot.
2. One Agent per OpenClaw session. Each OpenClaw session maps to a stable DSH session id (oc-<session-id>). First use creates the Agent; later gateway processes resume the persisted DSH session from disk.
3. Bridge the turn. The user prompt enters via agent.followup(...); the scoped session/event firehose maps assistant/chunk deltas to onPartialReply, reasoning to onReasoningStream, tool results to onToolResult, and assembled messages to the attempt result.
4. Stay reversible. Abort signals and the turn timeout cancel the DSH agent; sessions.flush persists before the attempt resolves; /new disposes the session's Agent; plugin unload disposes the whole host.

Configuration

All fields are optional, under plugins.entries.dsh.config in openclaw.json:

{
  plugins: {
    entries: {
      dsh: {
        config: {
          modelProvider: "deepseek-official",
          model: "deepseek-v4-pro",
          workspaceRoot: "/Users/me/projects",
          agentPreset: "cordis", // self-modifying agent
        },
      },
    },
  },
}

Authoring your own preset: copy a shipped composition (~/.dsh/.agent-presets/<id>/), edit the rows, and set agentPreset to your id — your OpenClaw agent gets exactly the tools, persona, and skills you choose. The cordis preset ships with the editing-cordis-compositions skill that documents this.

Activation modes

// A. dsh model ref (recommended) — any dsh/<model> id resolves
{ agents: { defaults: { model: "dsh/deepseek-v4-pro" } } }

// B. Pin the harness on an existing model entry
{
  agents: {
    defaults: {
      model: "openai/gpt-5.5",
      models: { "openai/gpt-5.5": { agentRuntime: { id: "dsh" } } },
    },
  },
}

Verified end to end

Validated on openclaw 2026.7.1-2 + @deepseek-ai/dsh-* 0.1.0-rc.6:

Troubleshooting

Known limitations

• OpenClaw's agent-harness surface for third-party plugins is still experimental; this plugin pins the validated version pair above.
• OpenClaw's own tools and custom instructions do not flow into the DSH agent — DSH runs its own tool ecosystem and persona (extraSystemPrompt is bridged as injected context).
• One embedded host per gateway process means one global workspaceRoot; per-session working directories still follow the OpenClaw session.
• DSH approval prompts have no interactive answerer in this embedding, so sandbox escalations fail closed unless you set DSH_PERMISSION_MODE=danger-full-access for unattended gateways.
• Dynamic plugins that include a Client (browser) half need an approval UI; in this embedding host-only packages run directly, while Client packages fail closed without an interactive reviewer.

Roadmap

• ClawHub distribution (clawhub:openclaw-deepseek-harness)
• DSH agent presets + dynamic Cordis plugin runtime
• Map OpenClaw workspace files into the DSH agent bootstrap context
• Optional bridge of OpenClaw tools into DSH via a scoped tool preset
• Approvals surfaced through OpenClaw's review UI
• npm distribution

License

MIT — see LICENSE. DeepSeek Harness packages are MIT-licensed (deepseek-ai/deepseek-harness).

Chinese docs: 中文说明