WJZ-P/dsh-model-capabilities

This is a plugin built specifically for DeepSeek Harness (DSH), designed to extend the default model configuration page with an additional input type selector. Before this plugin, DSH could only guess if a custom OpenAI-compatible model supported multimodal image input based on default provider settings, which often led to mismatched capability declarations and incorrect behavior when handling image attachments. The plugin integrates directly into the existing model creation and editing workflow, letting users explicitly set what types of input the model can accept, aligning the configuration with the model’s actual capabilities.

To use the plugin, you first open DSH’s model settings page, then either create a new custom model or edit an existing one, and expand the advanced configuration section for the target model entry. You will see the new input type selector added by this plugin, and can choose from four options: inherit default, text only, image only, or both text and image. Once you save the configuration, DSH will use this new declaration to check if the model supports image input when you attach files to prompts. This plugin is most useful for users who add custom multimodal models to their DSH instance, as it prevents incorrect capability guessing.

You can install the plugin via npm by running the DSH CLI command dsh plugin --profile web add dsh-model-capability directly into your DSH web profile. The plugin is distributed under the open source MIT license, and development requires Node.js version ^22.19.0 or >=24.0.0 and the pnpm package manager. It currently targets DSH 0.1.0-rc.5’s web client extension interface, and follows DSH’s plugin conventions with proper package manifests and a pre-built browser bundle. The plugin only modifies the input field of the target model via DSH’s public extension slot, so it is low-impact and does not take over core DSH functionality.