Introduction

When developing or debugging complex tasks with DeepSeek Harness (DSH), the output of a single model often cannot guarantee accuracy and completeness. The dsh-council plugin introduces a multi-model council mechanism: multiple models answer independently, peer review anonymously, and an arbiter makes the final decision. This helps identify blind spots, uncover inconsistencies, and improve the quality of the final answer.

What Is It

This is a multi-model council plugin designed for DeepSeek Harness, maintained by user a1exsun under the MIT License. It allows simulating a decision-making council in DSH and generates the final result through three steps: independent answers, anonymous peer review, and arbiter decision.

Core Features

  1. Parallel answering: 2 to 8 models independently answer the same question simultaneously.
  2. Anonymous peer review: 1 to 8 reviewers anonymously compare the answers, point out differences and blind spots, and provide rankings.
  3. Arbiter decision: An arbiter synthesizes all answers and review comments to provide a final recommendation.
  4. Inspectable results: The output includes information such as confidence, consensus, disagreements, and blind spots. Original answers and reviews are retained in DSH sub-sessions.
  5. Language support: Supports English and Chinese, depending on the DSH language setting.
  6. Model reuse: Uses providers and credentials already configured in DSH. Different roles can reuse the same model.

Installation and Enablement

Before use, ensure your environment meets the following requirements:
* Node.js version: ^22.19.0 or >=24.0.0
* pnpm version: 11.7.0
* At least two authenticated models are configured in DeepSeek Harness Web

The installation commands are as follows:

npx --yes @deepseek-ai/dsh@latest plugin --profile web add @a1exsun/dsh-council@0.1.0
npx --yes @deepseek-ai/dsh@latest web

If DSH Web is already running, restart it to load the new plugin.

Typical Usage

Enter the following command in the DSH Web chat window:

/council

Then complete the following two steps:
1. Select roles: Choose the answerers, reviewers, and arbiter in the panel. These selections apply only to the current run.
2. Ask a question: Provide current context information for each participant, ensuring they can think independently based on the same background.

For example, you can ask a technical architecture question that requires multi-angle analysis.

Notes

  • Cost: Each run involves 4 to 17 model participants. Some models may make multiple requests or use Web tools.
  • Accuracy: Model consistency does not imply answer correctness. It is recommended to carefully review the confidence explanations and review details.
  • Privacy: Your question and intermediate answers will be sent to the selected model providers.
  • Configuration: You can adjust limits by configuring answerMaxTokens, reviewMaxTokens, and runTimeoutMs. Each model request has independent Token limits.

Conclusion

By introducing multi-model competition and arbitration mechanisms, dsh-council can effectively improve the quality of handling complex problems. For more details, visit the community directory or view the source code.