DeepSeek Harness is built on the philosophy of “Everything is a plugin.” For developers, turning a one-sentence natural-language requirement into runnable code involves multiple stages, including requirement decomposition, architecture planning, parallel development, QA testing, and acceptance delivery. Managing these stages manually can easily lead to omissions or chaos, and Agents are also prone to “fake delivery.”

dsh-plugin-teamflow is designed to solve these problems. It is a distributable DeepSeek Harness plugin maintained by MichaelShii under the MIT license. It turns “a one-sentence user requirement into a multi-Agent pipeline of a real R&D team” into a host-level capability.

Core Features

  1. One-sentence requirement → acceptable delivery
    The plugin automatically orchestrates the full chain from requirement to delivery: requirement → PRD (document archiving) → Design (UI/UX) → architecture planning (scaffold + AGENTS.md) → technical proposal → parallel development → QA functional testing → product acceptance. Each stage has task cards, artifacts, and conclusions. Mechanical minor changes can use the patch / lite tiers to trim the stage set.
  2. Multi-Agent team + parallel development
    The plugin decomposes tasks according to the architecture blueprint and supports concurrency. The default is 3-way concurrency, with a maximum of 8. Product, architecture, development, and QA each operate in independent contexts without cross-contamination.
  3. Prevent fake delivery
    Delivery is judged not by wording but by evidence. Each stage must provide verification evidence in the form of “command + exit code + assertion count.” QA has independent adversarial probes, and acceptance only recognizes explicit conclusion lines.
  4. QA rejection closed loop
    P0–P2 defects are automatically sent back for fixes, with re-verification limited to 2 rounds or fewer. Each defect includes a “detection command + pass criterion.” If the limit is exceeded, the workflow escalates to a human and will not disguise itself as “completed.”
  5. Breakpoint resume + completion report
    After a process crash or restart, execution continues from the first uncompleted stage, while completed stages reuse existing artifacts. When the pipeline finishes, a summary (status / stages / token / follow-up guidance) is automatically returned to the originating session.
  6. Keep the repository clean
    Pipeline documentation is consolidated in the docs/teamflow/ task folder, and plugin run logs are archived out of the project at the end of a run. Commits contain only code + task folder. One run produces one commit, and no preconfigured .gitignore is required.
  7. TeamFlow Workbench (Dual Web Entry Points)
    Provides two Web entry points: an in-session tab (pipeline graph + drag-and-drop board) and an application-level main panel (product-line view, available across sessions). The interface is bilingual in English and Chinese and follows the host language in real time.
  8. Token accounting with traceable records
    Records the input (hit/miss), write cache, output, and call count for each stage according to official metrics, and provides cache hit rates. Reports and boards use the same accounting standard.

Installation and Enablement

Plugin installation requires the DeepSeek Harness (dsh) host environment, and the web profile is mandatory.

dsh plugin --profile web add dsh-plugin-teamflow

After installation, restart dsh --profile web. After the restart, 12 teamflow_* tools will appear on the model side (such as start / triage / status), and a “TeamFlow Workbench” entry will appear on the browser side.

Typical Usage

  1. Choose a team: Click the “TeamFlow Workbench” button next to the session input box, then select a team (or select “No Team” to skip the pipeline).
  2. Submit a requirement: Describe the requirement directly. The model will automatically call teamflow_start in automatic triage mode.
  3. View progress: In the session header, switch to the “TeamFlow Workbench” tab to view the pipeline graph and Backlog board. Click “⇥ Open in right panel” to move details into the right sidebar.
  4. Collect results: After the pipeline completes, the results are automatically reported back to the current session.

Environment Requirements and Dependencies

  • Host environment: DeepSeek Harness (dsh), and the web profile is required.
  • Node.js: ≥ 22.18.
  • Dependencies: Relies on @deepseek-ai/dsh-* and react provided by the host (peerDependencies).
  • Version compatibility: Host version must be >= 0.1.7-alpha.1 < 0.2.0.
  • Configuration injection: AGENTS.md is unconditionally injected by the harness into every session and is treated as a team asset.

Ecosystem Background

The DSH philosophy is “Everything is a plugin.” The community catalog is an independent site and has no official affiliation with DeepSeek / 0x. It should not be presented as an official app store. For more details, see the GitHub repository.