DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Enables official Chinese document mode for DSH: AI responses are written and rendered in formal govt-doc layout (title, salutation, signature) with a draw-to-feedback canvas; toggling off restores plain output.
Route specific provider/model pairs through HTTP or SOCKS5 proxies at the transport layer without rewriting baseURL, with a Settings UI.
A sidebar plugin that generates style-tagged reply candidates from your own system prompt on the latest AI message, with one-click fill into the composer and no auto-retry.
Generates diverse candidates via multi-temperature sampling, then uses a judge call to select and return the best answer with reasoning.
Adds per-model reasoning-effort settings and an in-picker slider to the DSH Web UI, supporting seven levels from off to max without manual config editing.
A read-only plugin for inspecting the effective prompt of a live session, displaying prompt sections, runtime contexts, variables, tools, and injected messages.
An installable DSH plugin that classifies requests and routes them to configured provider/model pairs with ordered technical-failure fallback.
Signs in to GitHub Copilot via device-code flow and registers a github-copilot LLM route for DeepSeek Harness using official dsh seams.
Extends DSH Desktop to accept arbitrary file attachments via drag-and-drop, with the agent reading files by path through a core patch.
An adaptive model routing plugin for DeepSeek Harness that dynamically selects the optimal inference model based on task requirements.
Manage local GGUF models and serve them via llama-server for inference without Ollama, with Hugging Face downloads, CUDA/CPU build selection, and a multilingual UI.
Auto-routes each user turn to the optimal DeepSeek model and reasoning level via keyword regex or semantic classification, supporting mixed, regex-only, and semantic-only modes.