Introduction

In DeepSeek Harness (DSH) usage scenarios, agents are often responsible for proposing solutions and optimization suggestions. However, verifying whether these proposals are truly feasible and free of defects often requires additional testing or manual review. The FormalSwarm plugin addresses this pain point by structuring the validation process: multiple agents generate independent arguments, adversarial critics raise challenges, and the Seal finally issues a verdict based on actual command execution results rather than relying on AI-generated text confidence.

Plugin Overview

FormalSwarm is a multi-agent validation plugin that supports DeepSeek Harness, Claude Code, and ZCode as runtimes. It builds a debate workflow using independent “argument writers” and “critics,” with the “Seal” responsible for validation. The verdict is computed from actual command exit codes and outputs CONFIRM, REVISE, or INCONCLUSIVE. The plugin is not limited to code repository review; it also supports document-driven non-code problem solving.

Core Features

  1. Multi-Agent Validation Architecture
    The plugin includes independent argument writers and adversarial critics. Argument writers propose solutions, while critics challenge assumptions, verify constraints, and identify gaps. They operate independently to ensure a comprehensive evaluation.

  2. Verdict Based on Command Exit Codes
    The verdict is not based on AI linguistic evaluation or confidence scores, but on actual command execution results. The Seal runs predefined checks, reads exit codes and test-case counts, and aggregates a deterministic verdict.

  3. Document-Driven Non-Code Workflows
    The plugin does not require a Git repository and can process document folders directly. Users can provide documents describing goals, product facts, constraints, and other context, allowing agents to review and validate proposals for non-code tasks such as marketing funnel design and release planning.

  4. Deterministic Summaries
    The plugin provides three deterministic outcomes: CONFIRM, REVISE, and INCONCLUSIVE. The output clearly distinguishes verified parts from unvalidated recommendations, ensuring that the verdict is grounded in evidence.

Typical Usage

For document-driven tasks, you can use the following command flow:

  1. Initialize the Repository
    Even without a Git repository, you can point --repo to a document folder to initialize it:
   $FS init --repo ./funnel-project

  1. Define Validation Questions
    In the brief stage, specify the validation questions that need to be answered and the executable checks. For example, validate whether an allocation plan satisfies budget constraints, or whether provided results are internally consistent.

  2. Run Validation
    The plugin launches a multi-agent debate, including 2 argument writers, 2 critics, and 2 seal validators. The final output includes the verdict and the specific scope that was validated.

Installation and Activation

The plugin is provided under the MIT License. Before use, ensure your environment meets the following requirement:
* Node.js version >= 18.17

Official resource links:
* GitHub repository: https://github.com/fashionmascherine-svg/formalswarm
* Community catalog: https://www.skillhub.cn/plugins/fashionmascherine-svg/formalswarm

Applicable Scenarios and Notes

  • Applicable scenarios: code change review, designing testable processes (such as funnel design), and proposal review for non-Git projects.
  • Notes:
    • A Git repository is not required; document folders can be processed directly.
    • Missing evidence should be treated as relevant information and described honestly in the Brief rather than fabricated.
    • Unverified recommendations must not be treated as facts.
    • The plugin runs with the permissions of the current DSH process; review the source code and license before installation.