Introduction

In the DeepSeek Harness (DSH) ecosystem, connecting to custom gateways (such as vLLM, LM Studio, or self-hosted OpenAI proxies) typically requires a dedicated adapter. dsh-llm-openai-completions is an OpenAI Completions-compatible adapter for custom gateways. It fills the gap beyond llm-deepseek and llm-pi-ai, specifically handling protocol differences in non-standard gateways, especially compatibility issues related to role definitions and thinking modes.

Core Features

The plugin mainly addresses the following issues:

  1. Role definition compatibility: Forces the use of role: "system", avoiding errors from custom gateways when they receive the non-standard developer role (Unexpected message role 400).
  2. Thinking-mode driven behavior: Thinking behavior is driven by compat.thinkingFormat in the model configuration, supporting multiple formats such as Qwen3.6 (enable_thinking), Qwen Chat Template, and Qwen3.8 (reasoning_effort).
  3. Response-side parsing: On the receiving side, separates Qwen3-style thinking content (vLLM renders it inside content) into independent reasoning blocks, preventing thinking text from being mixed into the main response body.
  4. Vision input support: Supports receiving user-uploaded images (single or multiple) and serializing them into an OpenAI-compatible multi-part content array (image_url).

Installation and Enablement

Installation

Install the plugin using the DSH command line (npm registry recommended):

dsh plugin --profile web add dsh-llm-openai-completions -w

After installation, restart dsh web to activate the plugin:

dsh web

Configuration and Enablement

Enable the plugin in ~/.dsh/profiles/web/cordis.patch.yml and specify the list of providers for which the plugin should take over proxy handling:

llm-openai-completions:
  enabled: true
  providers:
    - local-35b

Typical Usage

In the llm-pi-ai configuration, specify api: openai-completions and configure the baseURL and model parameters. The plugin takes over the streaming behavior for that provider based on the configuration.

llm-pi-ai:
  providers:
    local-35b:
      api: openai-completions
      baseURL: http://192.168.100.242:8200/v1
      models:
        - id: Qwen3.6-35B-A3B
          reasoningEfforts: { off: null, high: 'high' }
          compat:
            thinkingFormat: qwen        # 使用 enable_thinking，不使用 reasoning_effort

Notes

  1. Project maintenance has ended: According to an announcement dated 2026-10-01, this project has been discontinued and removed. The reason is that dsh-thinking-levels no longer requires this plugin.
  2. Version status: The version on npm remains 0.1.0. Although package.json shows 0.2.0, the 0.2.0 branch with vision/image input support was never published and will not be published.
  3. Dependencies: Vision input support depends on ctx.attachments (attachment storage) provided by the host to read image bytes and perform Base64 encoding. Text-only models are not affected.