dsh-llm-aimux

DeepSeek Harness ships with a built-in llm-pi-ai adapter that covers the major model routes, but once your pipeline touches relay gateways, niche providers, or dozens of simultaneously mounted routes, maintaining separate adapters becomes a drag. dsh-llm-aimux wraps aimux—a Rust-based unified LLM access layer with 325+ built-in providers (including 251 OpenAI-compatible registry entries)—into a single dsh model route, so one adapter spans the vast majority of inference endpoints.

The core experience revolves around hot-swapping and streaming. Add or remove routes via the Web Models page or the llm-aimux section in settings.yaml, and registration.replace takes effect immediately—no process restart needed. Streaming chat supports tool calls (raw JSON arguments) and native reasoning blocks in and out, with token usage broken into cache-read and cache-write slices and attribution headers attached to every outbound request. Routes pointing at self-hosted relays or gateways can override baseUrl and custom headers per route; model metadata is merged from aimux runtime /models discovery plus the community catalogue, feeding Models-page cards and context-window resolution. A streaming chunk-idle watchdog (default 300 s) prevents hung chunks from stalling the call chain.

Built for dsh users who need to mount multiple providers and hot-swap them on the fly; relay and gateway operators who want per-route endpoint overrides; and developers who want to unify 325+ inference endpoints—DeepSeek, Groq, OpenAI-compatible APIs and more—under a single harness.