Introduction

In the DeepSeek Harness plugin ecosystem, approving tool-call requests often requires manual intervention. The dsh-approval-ai-review plugin aims to address this pain point. Before a human responder is invoked, it uses a helper model to review each pending approval request. For safe operation requests, they can be approved automatically; for risky requests, it attaches the analysis, including the operation content and risks, to the request escalated to a human, thereby reducing repetitive manual work.

Plugin Positioning

This is a plugin for DeepSeek Harness, maintained by developer kiiiiile. Its core value lies in using AI to pre-check the approval process, automatically releasing safe operations, and ensuring that risky requests are clearly carried into the human decision process.

Core Features

  • Pre-check and automatic approval: Uses a helper model call to review each pending approval request before a human responder is invoked. If determined to be safe, the operation request is automatically approved (must be enabled).
  • Attached risk analysis: Attaches the review analysis (including the operation content and risks) to escalated or failed requests for human reference.
  • Enable/disable control: Can be enabled by the command /auto-approve [on|off] or by setting enabled: true in the configuration file.
  • Configurable parameters: Supports configuring the review provider (reviewProvider), model (reviewModel), reasoning effort (reasoningEffort), maximum output token count (maxTokens), and timeout (reviewTimeoutMs).

Installation and Enablement

Installation
Install the plugin in a DeepSeek Harness environment:

dsh plugin --profile web add dsh-approval-ai-review

Enablement
The plugin is inert by default; when it is not enabled, all requests remain unchanged. You can enable it as follows:

  1. Enable via command line:
    Use the following command in a DSH session:
    /auto-approve on
  2. Enable via configuration file:
    Add the following to the deployment configuration:
    - name: 'dsh-approval-ai-review'
      config:
        enabled: true

Implementation Details and Notes

  • Escalation only: The reviewer never rejects a request; dangerous requests are escalated to a human rather than being rejected.
  • Subsession limit: Subsessions do not inherit the approval/review/enabled toggle.
  • Latency impact: The review consumes wall-clock time. reviewTimeoutMs limits the maximum review duration, but it does not shorten the caller’s own tool timeout budget.
  • Dependency environment: The plugin depends on DeepSeek Harness core modules, including @deepseek-ai/cordis, @deepseek-ai/dsh-agent, @deepseek-ai/dsh-llm, and @deepseek-ai/dsh-user-approval, among others.

Conclusion

This plugin provides an automated approval intermediate layer for DeepSeek Harness, suitable for scenarios where you want to reduce manual operations and improve response speed. You can learn more through the following links:
* Directory page
* GitHub repository