rapid-mlx-dsh-provider: A DSH plugin that lets DeepSeek Harness auto-sync local Rapid-MLX model facts

rapid-mlx-dsh-provider is a native Rapid-MLX plugin (DSH plugin) for DeepSeek Harness, maintained by raullenchai under the Apache-2.0 license, last updated in 2026-08. It solves the problem that when DSH connects to a local Rapid-MLX server via the generic openai-completions route, model facts like context window and reasoning capability must be hand-maintained in settings.yaml, leading to stale numbers and wrong compaction timing. This plugin reads model facts directly from Rapid-MLX's /v1/models endpoint, so dsh always follows the server's real state without manual edits or re-running setup.

What core capabilities does the rapid-mlx-dsh-provider DSH plugin offer?

The plugin automatically syncs model metadata, so switching models requires no config changes; it only shows off/low/medium/high reasoning options for models that actually have a reasoning parser; and it times compaction based on the model's real context window (default thresholdRatio 0.8), avoiding premature or late compaction from stale hand-written values. It also reads fields like recommended_sampling, tool_call_parser, and is_hybrid, but not all are acted on yet—the current version is a working skeleton, unpublished to npm, verified end-to-end against dsh 0.1.0-rc.7.

How do I install and configure rapid-mlx-dsh-provider?

You need Node ≥ 22.15 (dsh relies on Node's Zstd stream API) and a running Rapid-MLX server. Install with:

dsh plugin --profile web add github:raullenchai/rapid-mlx-dsh-provider
export RAPID_MLX_BASE_URL=http://localhost:8000/v1     # optional; this is the default
dsh web

Then point the agent at the route in settings.yaml:

agent-default-model:
  provider: rapid-mlx
  model: qwen3.6-35b-8bit

What is the verification status and known limitations of rapid-mlx-dsh-provider?

Verified against dsh 0.1.0-rc.7 on an M3 Ultra: installs and activates as a profile layer, registers the rapid-mlx route, and handles plain chat, a single tool call, and a multi-step bug-fix task (completed in 36 s on qwen3.6-35b-8bit). Not yet implemented: auto-applying recommended_sampling, failing fast on models that cannot emit tool_calls, and memory-aware capacity calculation.

Which DeepSeek Harness versions and platforms does rapid-mlx-dsh-provider support?

The plugin targets Apple Silicon (MLX ecosystem), depends on dsh 0.1.0-rc.7 and Node ≥ 22.15, and is a working skeleton without a cross-version compatibility promise.

See more DSH plugins at dsh-plugin.org