The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When configuring the pi-ai route, developers often need to manually adjust the context window, maximum output tokens, image input support, and reasoning parameters. These parameters are usually scattered across configuration files or API calls, making manual maintenance costly and error-prone. The dsh-model-capabilities plugin adds a “Model Capabilities” page to DSH Web Host, providing a visual configuration interface to centralize management and correction of these parameters.

Core Features

The plugin mainly includes the following features:

  • Parameter editing: Provides a visual editing interface for the llm-pi-ai route, supporting configuration of the context window, maximum output tokens, image input, and the actual reasoning value.
  • Automatic loading: Automatically loads the model catalog from OpenRouter as a reference source for configuration data.
  • Smart matching: Uses a smart matching algorithm (based on token boundaries and rare-word weighting) to generate default drafts for new models or missing fields.
  • Custom mapping: Supports custom reasoning value mapping, allowing UI options (such as off/low/high) to be mapped to the strings the model actually accepts (such as none/low/extra).
  • Safe correction: Strictly follows the principle of “not overriding existing corrections,” fills only blank fields, and preserves manually edited drafts.
  • Browser memory: Provides browser memory by caching processed models for 6 hours to avoid repeated actions.

Installation and Enablement

To install the plugin, run the following command:

dsh plugin --profile web add github:Kalospacer/dsh-model-capabilities#main

After installation, restart DSH Web Host and then find the “Model Capabilities” page in the Settings menu.

Typical Usage

On the “Model Capabilities” page, the workflow is as follows:

  1. Select llm-pi-ai in the route filter bar.
  2. Browse the model list and click a specific model to enter the editor.
  3. Check the values (context window, maximum output tokens, etc.) and use the “Fill blanks” button to populate blank fields.
  4. Click “Save this model” or “Save all” to write the configuration to DSH settings.

The browser remembers processed model configurations, so you can use them directly after refreshing the page without re-entering them.

Use Cases and Notes

This plugin only takes effect for the llm-pi-ai configuration route. The configuration is written to DSH’s own settings. Saved configurations are retained after the plugin is uninstalled and will not be lost.

When using this plugin, note the following:

  • Data source differences: Before saving, verify the values actually accepted by the route, because OpenRouter is an aggregated catalog and does not necessarily guarantee that the native API values of all connected routes are identical.
  • Reasoning mapping: The reasoning value mapping must be explicitly specified (for example, off: none). Any option other than off must have an actual value to send.
  • Field inheritance: Leaving a numeric field blank means it inherits the default value. Saving after clearing a saved override removes the override for that field.
  • Draft mechanism: Unsaved changes are retained only on the current page and will be lost if you refresh or leave the page.

This plugin simplifies the configuration workflow for model parameters in DSH using visual tools, making it suitable for developers who need fine-grained control over pi-ai route parameters. For more details and source code, please visit GitHub or the community catalog.