dsh-local-llm-controller

Using a local large model inside DSH (DeepSeek Harness) has long meant juggling terminal commands to launch llama-server, wiring up ports and auth headers, then refreshing provider settings by hand. dsh-local-llm-controller folds that entire workflow into a single card on the DSH settings page: point at your llama.cpp directory and model folder, hit Start, and the local model appears in the session picker. Stop releases the port; status, errors, and recent logs are all visible right on the card.

Version 2.0 introduces dual model slots A and B, each backed by its own folder of GGUF files. Every slot exposes eight editable launch-parameter rows covering text vs. vision mode and short vs. long context presets, so you can freely tune -ngl, -t, -c, sampling options, and more. Model names and provider keys are derived automatically from the selected GGUF filename and can be renamed later in the Model settings page. In vision mode, an mmproj file is detected by name and mounted automatically—no manual CLI assembly needed.

This plugin is built for users who already run llama.cpp locally and want to drop Qwen, Llama, Mistral, or other GGUF-quantized models straight into their DSH chat sessions, especially when managing multiple models, switching context lengths, or working with vision multimodality. The plugin ships without a llama-server binary and does not download model weights; it focuses solely on bridging DSH to a local inference server.