視覚ツール · DeepSeek Harness プラグイン — 48 ページ

Zero-token offline math input for DeepSeek Harness: handwriting recognition, screenshot OCR, LaTeX editor, and inline LaTeX rendering.

Convert Office documents (.doc/.docx/.xls/.xlsx/.ppt/.pptx) and PDFs (incl. scanned, via local OCR) to structurally-formatted Markdown, powered by MarkItDown. CLI + dsh agent tool.

DeepSeek Harness Web GUI 插件:会话输入框 📎 按钮——图片直传官方附件管线;文本类原样暂存;文档(docx/pptx/xlsx/pdf/rtf/odt/epub/csv…)经本地 anydoc 转为 Markdown 后暂存,消息只插图短回执。Composer picker that converts documents to Markdown.

DSH Markdown 阅读插件：点击会话中的 MD/图片文件链接时，在浏览器右侧「详情栏」内以真正的三栏布局（侧边栏 | 会话区 | MD 文档）直接渲染阅读，而不是交给系统默认程序。目录 TOC、复制原文、字号缩放、自动跟随磁盘变更、图片内联预览。双半身插件：宿主半身提供只读 HTTP 路由，浏览器半身拦截点击并渲染面板。

Media Forge: agent + skills workflow for media-generation APIs, with a sidebar panel for loading saved generation skills into a conversation

Generate images and videos in DSH chat via OpenAI-compatible providers configured in Model settings; dedicated Settings menu and workspace media_gen output.

Plan and execute Atlas Cloud image, video, audio, and 3D workflows in DeepSeek Harness.

Deterministic aggregate media budgets and safe request projections for DeepSeek Harness (DSH)

All-in-one inline media for DeepSeek Harness chat: ```mip-img / ```mip-video / ```mip-audio fences render images (cards + lightbox + grids, TGA→PNG), native <video> and <audio> players right inside the message (house pattern of img-preview/dsh-mermaid/pbr-render). Image route whole-reads local workspace images; media route streams with HTTP Range seeking and an ffmpeg transcode fallback (cached) for containers browsers cannot play (mkv/avi/flv/wma...). Trust-fenced (loopback + webRuntime.trustedHosts) so LAN/mobile access keeps working. Supersedes the separate img-preview + video/audio fence plugins.

Medical imaging specialist tools for deepseek-harness: MAIRA-2 and MedGemma report generation for chest X-ray/CT/MRI/retinal (fundus) images, TotalSegmentator organ segmentation, BiomedCLIP zero-shot ultrasound classification, and BiomedParse text-prompted segmentation across X-ray, CT, MRI, ultrasound, and retinal (fundus) images.

DSH plugin: meme/reaction picker for DeepSeek Harness agents. No memes, no soul. Picks the right reaction image from the kagura-agent/memes library at runtime — semantic tag matching over 26 categories.

DeepSeek Harness 思维导图插件：AI 回答自动整理为可编辑 Markdown，支持导出文件和生成图片；核心逻辑已内置，单包发布。

DeepSeek Harness 插件：以 MindSee 为后端，为 DeepSeek 提供图片相关能力

明眸 VisionBridge - 自研视觉桥：瞎子模型收图时自动调用视觉模型识别，把识别文字喂回主模型，无感、可配置、升级不丢

MiniMax 文生图插件：把一句话画面描述生成图片并保存到工作区，作为 `image-gen` 工具提供给模型调用。

多模态路由 for DeepSeek Harness：纯文本模型的每一步请求里的每张图片（用户上传 / 工具返回 / MCP 渲染）都先经多模态理解模型转述为文字再作答，图片类报错自动转述重试，两类模型全程交叉协作。Multimodal router for DSH: every image in every step of a text-only model's stream is transcribed to text by a multimodal understander first; image-related failures auto-recover with a transcription retry.

MiniMax CLI bridge for DeepSeek Harness (dsh): mmx-backed web search fallback and transparent image understanding, with first-run onboarding and a settings card.

MiniMax multimodal capability hub for DeepSeek Harness (DSH). One mmx_multimodal tool covers describe/image/video/speech/music/cover/search/quota; optional web_search takeover; built-in client enhancement renders inline players in the Web GUI.

Automatic per-task model hot-switching for DeepSeek Harness (dsh): image-aware tasks route to the vision model automatically, every other task keeps your default model. Zero extra tokens, no context disturbance.

dsh web 插件：自动配置模型思考档位与能力，并按 provider+model 记住最后选择的思考档位

DSH Settings section to edit per-model input modalities (text/image) AND reasoningEfforts (incl. custom wire values), for all configured providers, models and modelOverrides at once — a companion to the lighter input-only WJZ-P plugin. Works on DSH Web and DSH Desktop.

Explicit model input-modality selection for the DeepSeek Harness Web UI

Model tags, reasoning capabilities, image routing, and vision proxy support for DeepSeek Harness

Capability metadata for hand-declared gateway models, resolved at runtime by bare model name (context window, output cap, reasoning levels, vision) — matched against the official DeepSeek route's catalog, the installed pi-ai catalog, opencode and models.dev, without writing to your settings.