dsh-volcengine

Most LLM plugins in the DeepSeek Harness ecosystem target a single provider, yet many teams routinely reach beyond it to model families hosted on Volcano Engine Ark, such as DeepSeek, GLM, Doubao, Kimi, and MiniMax. dsh-volcengine bridges that gap in one install: Ark models appear in the Harness model picker immediately, and you can point the agent at any of them mid-session without spinning up a proxy or writing a custom adapter.

The plugin works on two levels. First, it configures two provider routes via cordis.patch.yml: a pay-per-use channel (/api/v3) covering DeepSeek-V4, GLM-5, and the full Doubao-Seed lineup, and a Coding Plan channel (/api/coding/v3) for Kimi, MiniMax, and GLM models subscribed through that plan. Context windows reach up to 1M tokens, and several models accept image input alongside text. Second, it ships two lightweight tool plugins that talk to Ark's OpenAI-compatible endpoints with plain fetch and a Bearer key. volcengine_chat handles one-shot completions including reasoning and multimodal vision (pass an images array to send pictures to Doubao-Seed or DeepSeek-V4.1-Flash visual models). volcengine_image_generate calls the Seedream / Jimpeng text-to-image endpoint and returns the generated image directly.

It is built for developers already running a DeepSeek Harness workflow who want Ark models in the pool without adding SDK dependencies, signing logic, or extra infrastructure. If you also need TTS, ASR, video generation, Embeddings, or Rerank from Ark, the plugin's roadmap spells out the next extensions so the same project can grow with your stack.