DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Enables text-only models to see images by running a vision-side analysis before session entry, keeping the conversation model and selector completely unchanged throughout.
Fixes OpenAI-compatible gateways (e.g. LM Studio) that reject non-PNG/JPEG/GIF image MIME types by transcoding to PNG before the model request.
Map any key combination (e.g. remote F13-F24) to instantly switch models and reasoning intensity in DSH, with zero mouse interaction and no pop-ups.
Adds a one-click prompt optimization button to the Web chat input, silently rewriting your draft in place with undo support and a dedicated model selector.
Dynamically routes each step's reasoning effort for the DeepSeek flash model based on task complexity, running silently in the background without changing the model itself.
Automatically selects reasoning level and tool access per request using local rules—no classifier model, GPU warm-up, or per-prompt switches required.
Zero-config streaming prompt optimizer with three refinement levels, a JSON-based vocabulary, and one-click skill export.
A Windows computer-use plugin for DSH that drives mouse and keyboard with native vision screenshots, UI element reference clicks, window focus control, and DPI-correct coordinates.
A generic model orchestrator that discovers live models, profiles their capabilities, and routes each unit of work to the best match automatically.
Turns character, weapon, and scene descriptions into structured AI-video fight prompts with action causality and spatial movement first, targeting Seedance 2.0/2.5 and MiniMax H3.
Injects a polish-prompt template into the DSH composer with one click; the main agent rewrites it using full context, with zero dependencies and no auto-send.
A local reverse proxy that caps images per request to fix single-GPU vLLM vision backends while preserving SSE streaming.