Kimi connector for DeepSeek Harness¶

@phillarmonic/dsh-llm-kimi is a Kimi K3 (Moonshot AI) connector plugin for the DeepSeek Harness llm capability seam.

It registers a direct-fetch LlmAdapter for the kimi-code provider route and streams Kimi Code chat completions, an OpenAI-compatible dialect, as harness StreamChunks. Connection facts resolve per request, so a changed base URL, model catalog, or API key reaches the next request without a restart.

Why this plugin¶

• Native Kimi Code endpoint. Talks to https://api.kimi.com/coding/v1/chat/completions with bearer auth and genuine attribution headers.
• Streaming with reasoning. Streams text and reasoning deltas, replays reasoning_content on tool-call turns so multi-turn tool sessions stay valid.
• Image input. All four Kimi Code models accept images; the plugin reads them through the harness attachment service and inlines them as base64 image_url data URLs.
• Safe defaults. Fixed sampling by default (no temperature, top_p, or n), sensible token and context defaults, and an idle-timeout watchdog on every stream.
• Configured from cordis.yml. Every deployment-varying choice is a validated config field that reloads live.

Quick start¶

Install it as a bundle into a DeepSeek Harness profile:

dsh plugin --profile <name> add @phillarmonic/dsh-llm-kimi

Export your Kimi Code key and restart the profile:

export KIMI_CODE_API_KEY=sk-...
dsh --profile <name>

Then select provider kimi-code and a model such as k3 when you run a task.

Continue with Installation, Configuration, and Image input.

At a glance¶