Introduction

DeepSeek Harness (DSH) is a plugin-centric ecosystem. When using local models, developers often need to verify whether a model is an official release or a gray-release/debug version. Official models typically use the standard Let me construction, while some gray-release versions may use the I'm/I'll + verb-ing construction. The dsh-gray-test plugin addresses this verification need through concurrent testing and streaming monitoring of the chain of thought.

Overview

dsh-gray-test is a plugin maintained by LiLiLi-Zi and positioned as a gray-release model detection tool. It supports launching N concurrent DSH sessions with one click and performs judgment by streaming monitoring the chain of thought in real time.

Installation and Activation

  1. Place this directory into DSH’s plugins directory (e.g., ~/.dsh/plugins/dsh-gray-test or the plugins directory configured for DSH).
  2. Install dependencies:
   npm install
  1. Build the frontend:
   npm run build
  1. Restart DSH. A “gray test” entry will appear on the Settings page.

Configure API Key

This plugin does not include any built-in API keys. It uses the DSH host’s deepseek-official provider (official routing).

Method 1: dsh settings page (recommended)
1. Open DSH’s Settings → Models / Credentials.
2. Find the deepseek-official provider.
3. Enter your DeepSeek API key (format: sk-...).
4. Restart DSH after saving.

Method 2: Configuration file
1. Edit DSH’s configuration file (credentials or config under ~/.dsh/).
2. Add under the deepseek-official provider:

   { "apiKey": "sk-your-key" }
  1. Restart DSH after saving.

Core Features

  • Concurrent testing: Configure concurrency (1-100), determining how many test sessions are opened simultaneously.
  • Streaming monitoring: Monitor chain-of-thought content in real time.
  • Judgment logic:
  • Gray-release model: Count of I'm/I'll + verb-ing > count of Let me × 5.
  • Normal model: Let me appears ≥ 3 times and does not reach the gray-release ratio.
  • Abnormal model: No features appear (e.g., Chinese chain of thought).
  • Token speed detection: < 40 token/s is marked as slow.
  • Preset support: Supports standard/PTC/minimal/creative presets.

Usage

  1. Go to Settings → gray test.
  2. Configure concurrency, model (deepseek-v4-flash / deepseek-v4-pro / deepseek-v4-flash-vision-exp), thinking strength (off ~ max), and preset.
  3. Click “Start Test”.
  4. Review the gray-release occurrence rate.

Judgment Details

The plugin refreshes judgment every 2 seconds:
1. Gray-release judgment (priority, does not interrupt): Count of I'm/I'll + verb-ing (excluding going to) > count of Let me × 5. Mark as gray-release, run to completion, and count toward the gray-release total.
2. Normal judgment: Let me appears ≥ 3 times and does not reach the gray-release ratio. Judge as normal and immediately terminate the session.
3. Abnormal judgment: No features appear throughout (neither Let me nor I'm+verb-ing). Run to completion and treat as abnormal.

Output Location

Test results are written to ~/gray-test-runs/. Each session has its own subdirectory, and the complete output can be viewed in the workspace.

Presets

Test sessions can use the following Agent presets:
- Standard mode: A fully functional coding Agent (default).
- PTC mode: All standard capabilities plus Code Mode SDK presentation tools.
- Minimal mode: A dual-tool coding Agent with only persistent bash and str_replace_editor.
- Creative mode: For creating custom Agent presets.

Notes

  • This plugin runs with the current DSH process permissions.
  • License: MIT.
  • Ecosystem context: The DSH ecosystem is independent of DeepSeek / High-Flyer, and the plugin directory is a community site.

Conclusion

This tool helps developers quickly identify model versions by streaming monitoring chain-of-thought features. For more details, visit GitHub or the community directory.