Introduction¶
In the plugin ecosystem of DeepSeek Harness (DSH), different types of tasks often require different levels of model reasoning capability. Manually switching models is both cumbersome and error-prone. The dsh-model-controller plugin automatically selects the appropriate model and reasoning intensity in each user interaction through keyword and semantic classification, addressing this need.
Plugin Overview¶
This plugin is maintained by user jadehare and belongs to the model reasoning classification. It provides DSH with dynamic model routing based on keywords and semantic classification, and supports automatic selection of models and reasoning intensity according to tasks.
Core Features¶
- Dynamic model routing based on keywords and semantic classification.
- Automatic selection of models and reasoning intensity.
- Supports hybrid / semantic-only / regex-only matching modes.
- Dedicated model routing for Plan Mode.
- Routing metadata display.
Installation and Activation¶
Before installing, please ensure that the current environment meets the following conditions:
* Node.js version: ^22.19.0 || >=24.0.0
* DeepSeek Harness version: 0.1.1-rc.2 or a compatible version
Run the following command to install the plugin:
dsh plugin --profile web add dsh-model-controller@next
After installation, go to Settings → Plugins → Plugin Configuration and save the configuration. It takes effect from the next user turn.
Configuration and Default Rules¶
In the settings interface, you can adjust the following configuration items:
* Global Matching Mode: Supports hybrid mode (regex first, then semantic), Semantic-only, and Regex-only.
* Route Rules: Add, delete, modify, and order routes; the built-in default table can be restored.
* Semantic Classifier: Used only in hybrid / Semantic-only mode.
* Plan Mode: When in Plan Mode, force use of the dedicated Pro/max model.
* Routing Metadata: Displays provider, model, reasoning intensity, and Route id at the beginning of the reasoning output.
The plugin’s default routing rules are as follows:
| Scenario | Model | Reasoning |
|---|---|---|
Complex structure (v4pm, refactoring, design) |
deepseek-v4-pro |
max |
Review (review) |
deepseek-v4-pro |
high |
Small script (v4fm) |
deepseek-v4-flash |
max |
Small method (v4fh) |
deepseek-v4-flash |
high |
| In Plan Mode | deepseek-v4-pro |
max |
| Semantic classifier (when no match is hit) | deepseek-v4-flash |
off |
Typical Usage¶
Based on the task type, the plugin automatically routes to the corresponding model:
* Complex structure scenarios (such as refactoring or design): use deepseek-v4-pro(max).
* Small scripts or small methods: use deepseek-v4-flash(max) or high.
Use Cases and Notes¶
- This plugin runs with DSH process permissions. Please review the source code and license before installation.
- The plugin version is
0.1.0-rc.1. It may have discrepancies with the Harness version requirements; please confirm compatibility. - When keyword matching does not hit, the Agent’s original model is preserved.
Summary¶
After installing and enabling dsh-model-controller, the plugin automatically selects a model in each user turn based on the configured rules. This significantly reduces the maintenance cost of model selection and improves the adaptation of task processing.