DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Wraps Qwen's dynamic-resolution vision pipeline into a DSH plugin, enabling text-only models to perform image analysis, OCR, object grounding, and long-screenshot recognition with progressive tool exposure and multi-provider failover.
Integrates Ollama's native /api/chat protocol into dsh for local or remote inference, with one-click model discovery, streaming chat, tool calls, multimodal, reasoning traces, and direct parameter passthrough.
A community-maintained DeepSeek Harness bundle serving conversation models through Databricks AI Gateway with streaming and tool-use support, installed out-of-tree with zero patches to core.
A product-design copilot for DeepSeek Harness that turns a one-line request into a reviewable, executable, and verifiable delivery plan via S/M/L routing and a hard quality gate.
Folds thinking and tool-call blocks so answers render directly, and adds an expandable session-activity bar above the composer showing live Think / Edit / Bash / Read counts.
Zero-dependency geodesic toolbox for dsh: great-circle distance, bearings, destination points, rhumb lines, path intersections, and spherical polygon areas with exact spherical trigonometry.
A DeepSeek Harness plugin that connects xAI Grok models with auto-discovery, mixed-protocol routing (Chat Completions & Responses), and OAuth token refresh.
A Bayesian stock prediction plugin that fuses multi-indicator signals (MACD, RSI, volume, etc.) to produce posterior win-probabilities, 0-100 score ratings, and multi-horizon trend analysis.
Switch DSH session models with a single natural-language sentence across Web, TUI, and IM interfaces, preserving full conversation context for seamless task continuation.
X-Ray observability for DeepSeek Harness — attributes per-request context cost to each plugin and inspects prompt sections, tool schemas, and skill pricing.
End-to-end toolchain for DSH to run, validate, and draft reports on the ml-quant-trading protocol v1 CPU benchmark.
Local-first, evidence-preserving tool-output compression for DeepSeek Harness that condenses oversized logs, JSON, HTML, and text into compact model-ready context without any LLM call.