DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A micro-tier compaction observer that clears whitelisted re-runnable tool results to placeholders below the summarize threshold, reducing expensive model calls without mutating requests.
Launch concurrent dsh sessions and stream-monitor chain-of-thought to classify gray, normal, or anomalous models in real time, reporting the gray-model ratio.
Auto-switches to a vision model when images are present, then falls back to your original model when the turn ends.
A draggable dial for DSH Web GUI that shows real-time model output speed (tok/s) and expands to reveal precise speed, TTFT, and cumulative token details.
Adds a star button to the DSH composer that diagnoses your draft with a rule engine and rewrites it into a structured, high-quality prompt using the session model, no extra API key required.
A DeepSeek Harness plugin for managing connections to local and OpenAI-compatible model servers, featuring llama-swap and Ollama support, image input, reasoning effort levels, and built-in model presets.
Adds image recognition to DSH sessions by calling any OpenAI-compatible vision model API.
Add a searchable model picker, recent-model quick switch, image-capability badges, and inline image upload to your chat input for faster model switching and multimodal interaction.
Fixes the missing finish_reason / [DONE] stream-termination defect for models such as muse-spark-1.2 and gpt-5.6-luna in DSH, eliminating spurious TRANSPORT errors and wasted retries.
Gives text-only Agents an external visual eye via a local MCP server, delegating image understanding, generation, and editing to Agnes AI.
Gives text-only agents vision capabilities by calling DeepSeek VL, GPT-4V, or any OpenAI-compatible endpoint to describe, locate, detect, and crop image elements.
Runs configured type-checkers and linters after successful file edits, injecting diagnostics into the next model request for immediate error visibility.