Introduction

In the DeepSeek Harness (DSH) workflow, agent teams are typically used to scale execution capabilities: assign different roles, break down subtasks, and merge outputs. This pattern is well suited for tasks that are highly repetitive or require parallel execution.

dsh-consult addresses a different kind of problem: decision quality. It is not responsible for executing tasks, but instead provides a “second opinion” for decisions that require careful judgment. After a task is completed, the main agent is responsible for investigation and verification, bundles the collected evidence, and then initiates a consultation with a group of independent models. These models answer the same question based on the same evidence, after which the main agent compares and synthesizes their anonymous opinions to make the final decision.

Features and Purpose

This plugin is a standalone DeepSeek Harness plugin maintained by DK-Zhu and released under the MIT License. Its core purpose is to introduce a multi-model perspective at decision boundaries, used to review high-leverage issues such as architecture choices, security boundaries, and release readiness.

Key capabilities include:
* Evidence-first: Consultations are triggered only when the evidence is ready, so no automatic model invocation costs are incurred during execution.
* Parallel independent review: The configured models (consultants) work in parallel and cannot converse, debate, or view each other’s opinions.
* Anonymous synthesis: The main model sees only stable labels (such as “Consultant A”), not specific model provider names, ensuring judgments are based on content rather than brand.
* Permission isolation: During consultation, consultant models have no tool access, and no access to the workspace or history.

Installation and Configuration

Before use, the following environment requirements must be met:
* Node.js >= 22.19.0
* DeepSeek Harness 0.1.5-rc.2

1. Install the Plugin

Install the plugin through the DSH plugin manager:

dsh plugin --profile web add dsh-consult

2. Configure the Consultation Panel

After installation, the plugin remains dormant. You need to enable the consultation feature in the corresponding Profile configuration file. For example, add the following to ~/.dsh/profiles/web/cordis.patch.yml (the exact path depends on the DSH_HOME environment variable):

- id: consult
  config:
    consultants:
      - provider: "<provider-id-a>"
        model: "<model-id-a>"
      - provider: "<provider-id-b>"
        model: "<model-id-b>"

Configuration item descriptions:
* provider: The model service identifier.
* model: The specific model identifier under that service.
* You can configure 2 to 5 models as consultants. The same provider may use different models.
* Validate the configuration and start DSH:

dsh --profile web --dump-config
dsh web

Usage

Consultations are triggered by the /consult command. After completing implementation and testing, the main agent prepares the evidence and then asks the panel.

Typical Invocation Examples

/consult Finish the implementation and tests, then ask the panel whether this is ready to ship.

/consult Review the cache invalidation redesign. Inspect the implementation and tests first, then decide whether it is ready to ship.

Execution Flow

  1. The user issues the /consult command.
  2. The main agent investigates, implements, and verifies, then prepares an evidence package.
  3. The evidence package is sent to all configured consultant models.
  4. The consultant models generate independent opinions in parallel.
  5. The main agent aggregates these anonymous opinions and outputs the final recommendation.

Use Cases and Considerations

Use Cases

The plugin is best suited for decision boundaries, especially when there is uncertainty or high-leverage impact:
* Choosing among multiple reasonable architecture options.
* Reviewing security, privacy, or data retention boundaries.
* Checking readiness before release.
* Evaluating migration or rollback plans.
* Balancing performance, maintainability, compatibility, and user experience.

Considerations

  • Manual triggering: Consultation is explicit and is not automatically triggered during execution, avoiding unnecessary token consumption.
  • Data privacy: The complete evidence package is sent to all configured consultants and persisted to DSH’s session event log. Before requesting a consultation, confirm that the material is appropriate to send to external providers and to persist locally.
  • Cost control: Model invocations occur only at consultation time, and the user explicitly controls the panel size, so there are no background automatic costs.

Summary

dsh-consult keeps model diversity at decision time rather than during execution. By providing a controlled, evidence-first review mechanism, it helps developers validate critical decisions.