Introduction

The LLM Seam plugin ecosystem in DeepSeek Harness (DSH) allows developers to extend model support. dsh-llm-mimo is an adapter for Xiaomi MiMo v2.5. It resolves the HTTP 400 INVALID_REQUEST errors present in the original fork and supports the newer dsh-llm dependency.

Core Features

  • Thinking chain support: Correctly recognizes MiMo’s reasoning_content field and converts it to the harness’s ReasoningBlock format.
  • Multi-turn conversations: Preserves the assistant message’s reasoning_content during tool-calling turns, as required by MiMo.
  • Tool calling: Serializes tool results into OpenAI-compatible role: 'tool' messages, sending "" when the text is empty.
  • OpenAI-compatible API: Uses MiMo’s OpenAI-compatible chat completion endpoint.
  • Full SSE streaming support: Supports full Server-Sent Events streaming and chunk assembly.

Installation and Activation

The plugin declares a dsh.bundle manifest, and one-step installation via the command line is recommended.

dsh plugin --profile web add github:dfhxxc666/dsh-llm-mimo --config.minimum-release-age=0

After installation, select Xiaomi MiMo → MiMo-V2.5-Pro in the GUI model selector to use it.

Configuration

  • API Key: MIMO_API_KEY must be configured in ~/.dsh/.credentials.yaml.
  • Models: Supported models are mimo-v2.5-pro or mimo-v2.5.
  • Dependencies: The dependency @deepseek-ai/dsh-llm must be version ^0.1.1-rc.2 or higher.

Notes

  • Non-bundle type: The plugin is not a bundle type. If installed manually, an llm-mimo block must be inserted into cordis.patch.yml.
  • Manual restart: After installation, the DSH process must be manually restarted via restart-dsh.ps1 for the changes to take effect.
  • Version dependency: Ensure that @deepseek-ai/dsh-llm is version ^0.1.1-rc.2; otherwise, errors may occur.