oh-my-deepseekharness

Introduction¶

DeepSeek Harness (DSH) uses a plugin-based architecture. Implementing multi-agent orchestration in DSH typically requires manually handling complex task decomposition, role assignment, and state management. oh-my-deepseekharness adapts the OMC/OMX orchestration philosophy, encapsulating it as a single DSH plugin and providing a standardized multi-agent interaction protocol.

Core Positioning¶

The multi-agent orchestration layer for DeepSeek Harness — a DSH host adapter implementation of the OMC / OMX orchestration philosophy.

Installation and Enablement¶

This plugin is distributed as an npm package. Ensure the environment meets the requirements before installation: Node.js >= 22.19.0, and dsh host version >= 0.1.7-rc.1.

dsh plugin --profile <name> add @sakura12/oh-my-deepseekharness

After installation, you can enter /omd-doctor in the session to run an installation diagnostic.

Three Execution Modes¶

The plugin provides three mutually exclusive execution modes, triggered by natural language keywords or slash commands:

1. autopilot
   Natural language triggered, driven by the dsh native goal tool in a foreground loop. Supports end-to-end automatic orchestration (interview → planning → execution → dual review gates → checkpoint persistence).

   • Example: autopilot help me build a CLI word-counting tool

2. ralph
   Explicitly triggered (with the ralph: prefix), driven by the dsh native ralph tool. Uses a PRD + evidence contract iteration loop, blocking in the foreground.

   • Example: ralph: sync the test badge number in the README to the current actual value

3. team
   Explicitly invoked, driven by the lead runtime in lib/team.js. Uses a five-stage pipeline (stage state machine + teammate UUID lifecycle + mailbox + heartbeat + merge plan).

   • Example: use team to research the plugin market ecosystem and produce a report

To cancel the current mode, enter cancelomd (or /omd-cancel).

Capability Overview¶

The plugin ships with a rich set of skills and tools:

• Roles and Skills: Includes 33 role cards (omd-agent-*, such as api-reviewer, product-manager, etc.) and 42 Skills (bilingual Chinese-English assets).
• MCP Tools: Mounts the built-in omd-state MCP server, providing 55 tools, covering state management, memory storage, coordination, code intelligence, graph runtime, etc.
• Runtime: Supports graph DAG runtime, Team pipeline, HUD panel, cross-session memory, and adaptive routing.
• Code Intelligence: Supports ast_grep structured search/rewrite and the LSP toolset (diagnostics, symbol references, etc.).

Configuration¶

The plugin is configured using cordis.patch.yml, supporting language, tier mapping, role overrides, iteration round control, heartbeat detection, etc.

# cordis.patch.yml
oh-my-dsh:
  language: zh
  tiers:
    low: inherit
    high: deepseek-reasoner
  autopilot:
    maxIterations: 15
  team:
    heartbeatIntervalMs: 60000
  codeIntel:
    lspServers:
      typescript:
        command: typescript-language-server
        args: ["--stdio"]
        languages: [typescript]

Dependencies and Notes¶

• Dependency Boundary: All @deepseek-ai/* dependencies are required by the host; if any are missing, automatic degradation occurs. @ast-grep/napi, vscode-rpc, etc., are optional dependencies; if absent, an explicit error is raised but execution is not blocked.
• Environment Requirements: Node.js >= 22.19.0, dsh host 0.1.7-rc.1+.
• Deprecated Keywords: ultrawork, ulw, uw, ccg no longer trigger; recommend using autopilot or team.
• Runtime Artifacts: All state and logs are stored under the .omd/ directory; it is recommended to add it to .gitignore.

Ecosystem Links¶

• GitHub: https://github.com/1225Sakura/oh-my-deepseekharness
• Catalog Page: https://www.skillhub.cn/plugins/1225Sakura/oh-my-deepseekharness