Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture, allowing developers to extend behavior. When the same model both writes code and validates code, it tends to “prove itself correct,” which can lead to three common failure modes in long-chain Coding Agents: self-validation blind spots, long-horizon drift, and premature completion claims. The core role of dsh-advisor-plugin is to introduce an “independent second opinion” and use a review model to mitigate these issues.

Plugin Positioning

The plugin is named leeyoung1/dsh-advisor-plugin and is positioned as the DeepSeek Harness advisor strategy pattern plugin. It pairs a fast execution model with a stronger review model. At key checkpoints, it passes the entire session to the review model, receives plan / correction / stop signals, and then continues execution.

Core Features

The plugin provides two complementary channels and a set of supporting capabilities:

  1. Explicit consultation channel
    The execution model can call advisor() with zero parameters, and the full session together with the tool inventory is automatically forwarded to the review model. The model can consult without organizing the context itself.

  2. Automatic patrol channel
    Every N steps, a session snapshot is sent to the review model. The review model determines whether the current state is ON-TRACK, CORRECTION, or STOP. The earliest intervention is applied to the next model request and does not interrupt the turn currently in flight.

  3. Supporting capabilities

    • Read-only investigation tools: The review model can first use search_files / read_file in the workspace to verify facts before giving advice.
    • Failure isolation and dual fallback: If the review call fails, times out, returns an empty body, or overflows the context, the execution model’s turn is not interrupted. If the context overflows, it is safely truncated and retried.
    • Execution model blacklist: Supports three syntaxes: model, provider/model, and provider/model@effort, for disabling the review feature for specific models.
    • Settings page visualization: All review configuration is completed on the DSH settings page, including configuration for the two blocks covering model and reasoning effort, and patrol mode.

Installation and Activation

  1. Install the plugin:
    dsh plugin --profile web add dsh-advisor-plugin
  1. Restart dsh web.

After installation, open Settings → Plugins and locate the Advisor Review Model card. Select a review model that is stronger than the execution model, and configure the reasoning effort as needed. Decide whether to enable patrol mode based on the task type.

Typical Usage

  • Session command: Run /advisor in the session, or let the execution model call the advisor() tool.
  • Configuration file: You can also directly edit the advisor section in ~/.dsh/settings.yaml:
    advisor:
      enabled: true
      provider: example-provider
      model: reviewer-model
      effort: high
      patrolEnabled: true
      patrolEverySteps: 6
      patrolImmuneTurns: 3
      investigate: true
      disabledForModels:
        - deepseek-v4-pro
        - gpt-5.2@high
  • Zero cost when not armed: If no review model is configured, neither the tools nor prompt snippets are registered.

Dependencies and Compatibility

  • DSH version: The target versions are 0.1.5-rc.1 or 0.1.5-rc.2.
  • Node.js: The required version is >=22.19 or >=24.
  • Installation method: dsh plugin --profile web add dsh-advisor-plugin is recommended. Bare npm install is not recommended.
  • DSH Desktop 0.2.1: A legacy package name compatibility issue exists. In that version, cordis.patch.yml still references the old package name dsh-advisor. If this issue occurs, remove the old package before installing.