dsh-llm-sampling

In multi-model inference pipelines built on the DeepSeek Harness, callers must manually supply temperature, topP, topK, and other sampling fields on every request, which is tedious and prone to omissions that cause output quality to drift. dsh-llm-sampling elevates sampling policy to a deployment-level concern: declare a complete parameter set for a named provider/model route in settings.yaml, and the plugin overrides every upstream agent/request proposal in the waterfall, guaranteeing a consistent generation distribution on that route.

The plugin touches no prompt text or tool schema and adds zero tokens; it only reshapes the sampling distribution itself. When reasoningEffort is explicitly set to off, the off profile overlays the default; unconfigured routes pass through untouched. This lets you pin distinct topK and minP combinations per model and get deterministic parameter fallback on reasoning-mode switches, without scattering conditional logic across business code.

Ideal for teams that need uniform generation behavior across multiple providers and models; platform engineers who want to lift sampling strategy out of call-site code and into the infrastructure layer; and organizations that require an auditable, per-route parameter record for A/B experiments or compliance.