DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A standalone DSH plugin that gives each session an independent reviewer model to observe the primary transcript and inject severity-ranked advisory feedback without interfering with the primary loop.
Injects a mandatory Simplified Chinese reasoning directive so the model always thinks in Chinese while replying in the user's language and keeping code unchanged.
Add a configurable vision model to text-only main models, converting chat images to text for replies and supporting OCR, charts, and screenshots.
Offline local document intelligence for PDFs, Office files, and scanned images, converting them to Markdown, plain text, or JSON without Docker or API keys.
oh-my-dsh is a curated DeepSeek Harness distribution that provides an opinionated coding environment with LSP navigation, lazy MCP activation, and approval-gated skills.
A DeepSeek Harness bundle plugin that upgrades built-in search and fetch with fused multi-engine search, X search, focused fetching, deep research, and parallel research.
Adds configurable reasoning effort levels to hand-declared llm-pi-ai models and sets a default reasoning effort for subagents.
Automatically converts dragged or pasted images and files into workspace paths for text-only models, bypassing native attachment limits while preserving the native UI experience.
An instant, near-lossless context compaction drop-in replacement for DeepSeek Harness that uses deterministic text processing to achieve zero-cost, millisecond compression without LLM calls.
Unified agent preset for DeepSeek V4 Pro/Flash, anchoring RL-aligned states with full Standard tools, dual-channel support, and long-session stability.
Intelligently shrinks agent tool outputs based on semantic shape and meaning, preserving critical errors and summaries to prevent silent data loss from naive truncation.
Provides structural code analysis tools for DeepSeek Harness, enabling the agent to accurately query declarations, call chains, and impact scopes via a symbol graph instead of inefficient text search or LSP.