Preface

The DeepSeek Harness (DSH) ecosystem emphasizes “everything is a plugin.” When building complex Agent tasks, a single inference may be limited by the context window or randomness, leading to unstable results. Multi-candidate mode improves the final success rate of tasks by using additional compute power through parallel generation and a verifier mechanism.

Plugin Positioning

dsh-multi-candidate is a test-time scaling plugin. It is maintained by WintryGrass, and its core logic is to have the model generate N independent candidate solutions for the same task in parallel, and then select the best one through a verifier mechanism. This is essentially a “run it N times and submit the best one” strategy.

Core Features

  • Whale Panel: Provides a floating ball interaction interface, supporting free dragging and position memory.
  • Purely configuration-driven: No special instruction prefix is required; after checking the configuration, you can directly input the task.
  • Session isolation: The configuration only takes effect in the current session. New sessions are disabled by default to avoid contamination.
  • Configuration options: Supports candidate count (2–5), second-opinion review (multi-voice deliberation), and optional validation criteria.
  • Hardwired tool: Provides the multi_candidate_run tool for explicit invocation by the model.

Installation and Enablement

  1. Ensure the host environment provides the peer dependencies (such as @deepseek-ai/dsh-tools, etc., all optional).
  2. Run the installation command:
dsh plugin --profile web add dsh-multi-candidate
  1. After installation, a 🐋 button will appear in the bottom-right corner of the Web GUI.

Typical Usage

  1. Click the 🐋 icon in the bottom-right corner to open the floating ball panel.
  2. Check “Enable multi-candidate mode (current session only).”
  3. Adjust the candidate count (default 3), second opinion, and validation criteria as needed.
  4. Enter the task normally in the main input box (for example, “Help me write a Python number-guessing game”).
  5. The model will automatically execute the pipeline of parallel candidate generation, validation-based selection, and reporting.

Notes

  • Installation mode: Installation uses copy mode. If you modify the source code, you need to run remove + add to reinstall, or manually sync it to the profile’s node_modules (pnpm cache does not refresh automatically).
  • Environment dependencies: Server-side changes require restarting DSH; client-side changes only require refreshing the page.
  • Execution boundary: The final execution of multi-candidate mode relies on the workflow tool at the agent layer; server-side tools cannot directly orchestrate sub-agents.

Conclusion

This plugin improves the stability of DSH tasks through simple configuration. For more details, please refer to the source code: GitHub