DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A portable ModelProvider enabling DeepSeek V4-Pro to handle image attachments by captioning via Flash Vision and streaming text-only responses, compatible with dsh-std v0.15.
Repairs streaming tool-call responses during model inference in DeepSeek Harness to ensure complete and usable outputs.
Auto-detect and configure reasoning levels and vision input modalities for self-hosted llm-pi-ai gateways.
Offline, deterministic verification that a hash-linked change ledger settles within one declared UTC maintenance window, bound to a single receipt, artifact, environment, policy, and plan.
An on-demand advisor tool that delegates planning, correction, or stop-signal review to a separately configured reviewer model per agent scope.
ADHD-friendly output shaping for DSH: action-first replies, numbered steps, zero fluff, with live on-off toggles and restart persistence.
Pushes system notifications to macOS, Windows, and mobile devices when a DSH task completes, errors, or blocks, with plain-text layout and built-in SVG/PNG icons.
A global model command palette for DeepSeek Harness Web with fuzzy search, provider filtering, favorites, reasoning effort control, and OpenRouter media tools.
Local soundscape plugin for DSH that plays thinking, tool, error, and emotion-aware audio cues with ducking-based mixing and per-tool sound mapping.
Adaptively routes each user message to the optimal reasoning effort (low/high/max) via MiniMax scoring, with tiered token ledgering and per-reply metadata badges, safe on GLM-5.3 forced-thinking models.
A zero-config DSH shim that transparently re-encodes PNG/WebP image blocks to JPEG for LM Studio openai-completions routes, preserving the native multimodal flow.
Shows DeepSeek peak/off-peak status, countdown, and session cost in the composer dock; hover for full rate breakdown, token buckets, and a cache-hit sparkline.