@raullenchai/dsh-provider
A native Rapid-MLX model provider for DeepSeek Harness that reads a model's context length, quantization and reasoning support directly from a running Rapid-MLX server instead of a hand-maintained settings file.
This adapter connects DeepSeek Harness (dsh) to a local Rapid-MLX inference server through Rapid-MLX's own /v1/models endpoint, replacing the generic openai-completions provider's hand-written configuration. It reads the server's reported context window, whether a model actually supports a reasoning parameter, and the server's memory-fitted token ceiling, so switching models on the Rapid-MLX server does not require updating dsh's settings by hand.
Because it reads the server's real capacity rather than an advertised or copied number, context-compaction timing follows what the machine can actually hold. The project describes itself as tracking a moving target: it was verified end to end against dsh 0.1.0-rc.7 on an M3 Ultra, and is re-verified at the protocol and unit-test level against rc.8, whose adapter contract is additive only.
It requires Node.js 22.15 or newer and is published to npm.