dsh-llm-kimi

DeepSeek Harness exposes a unified LlmAdapter seam for every upstream provider, yet teams wanting to call Moonshots Kimi K3 family had to hand-roll the integration. dsh-llm-kimi ships a ready-made direct-fetch adapter that streams Kimi Code chat completions in their OpenAI-compatible dialect and maps every token into standard Harness StreamChunks, so orchestration layers stay provider-agnostic.

Under the hood the adapter resolves base URL, model catalog, and credential on every request, meaning a rotated key or a new endpoint takes effect on the next call with zero restarts. Image input is first-class: the adapter inlines base64 data URLs through the Harness attachment service and enforces pixel and byte budgets before any network I/O. Reasoning effort is exposed as low / high / max for the K3 family and locked on for the coding models. During multi-turn tool sessions the adapter replays the Harness reasoning block into the reasoning_content field Kimi expects, keeping conversations valid. Production safeguards include a five-attempt retry policy, a 300-second stream idle timeout, and a 131K default output cap that yields to per-model limits.

It is aimed at engineering teams whose pipeline already plugs into the Harness llm seam and who need Kimi K3 or Kimi for Coding with multimodal image support, multi-step tool orchestration, and hot-reloadable configuration. Adding a provider entry to cordis.yml is the only integration step required.