Vision tools · DeepSeek Harness plugins — Page 15

Agent Skills and Rules Workbench for DeepSeek Harness better-sidebar

DeepSeek Harness plugin + browser extension bridge: lets DSH agents read and control the user's browser, like Kimi WebBridge or Claude's browser extension. Node side: browser_* tools over a localhost WebSocket; browser side: a Manifest V3 extension.

DSH 对话体验增强套件 + 四工作台（零 DSH 源码改动，纯插件注入）。回合呈现：思考 / 工具调用聚合 chip（实时走秒 + 分步跟随滚动）+ 步骤卡 / 总结卡（回合收口才出现）+ 共享活动抽屉（思考语义分组 + 工具调用树）。正文增强：proto-tabs 可交互卡片 + diagram 流程图围栏（JSON 转 SVG）+ 生图画廊 + 重试行影子。界面与工具：会话头部视图标签上移右上角 + 桌面壳窗口控制留位与主题同步 + 对话截图（常驻无头浏览器渲染 markdown/shiki/mermaid，可内嵌本地 HTML）+ download 下载工具（wire 工具 + 实时进度条/速度/ETA）+ KR 对话双栏大盘（右栏保留任务 / 思考 / 工具调用 / 记忆四张卡，顶栏由 KR_PANEL_HEADER_VISIBLE 门控）。融合的原 dsh-triad 四工作台（座位与路由前缀原样保留，dsh-triad 自此退役）：自动沉淀的长期记忆、定时自动化、用量与账号趋势、技能与 MCP Server 管理。

Give DeepSeek Harness agents live control of your ComfyUI canvas (local or cloud): embed it as a split-screen tab, then read/edit/run workflows, debug errors, fetch output images, and batch-sweep parameters straight from the chat. Ships the ComfyUI-side bridge node.

Vision-only desktop automation plugin for DeepSeek Harness (fusion refactor)

Windows Computer Use for DeepSeek Harness: window-bound screenshots, robust OCR, verified clicks, pure-OCR mode, pluggable vision models.

DeepSeek Harness 功能坞插件（dsh-dock）：模块化功能中枢——每个功能一个 features/ 模块（可单独提取成包发布，又可经 dockBridge 装回面板）。含用量记录（LLM 调用记账/费用统计）、模型设置、图片理解代理、模型余额、任务动画（8 种氛围动效 + 桌面伙伴 3D 场景、完成通知与提示音、钉钉/飞书推送，开关独立、配置持久化）；侧栏入口 + 可拖拽弹窗（最大化/最小化/缩放）。

Focus Mode for the dsh web GUI: a full-screen reading overlay that hides chrome and folds the AI tool-call flow into summaries, reusing the official Markdown / image primitives.

DeepSeek Harness 娱乐工作台插件：小说 / 漫画 / 图片 / 音乐 / 小游戏 / RSS，多列卡片首页 + 全局音乐播放器，本地优先存储。

Application-level Vision-Language-Model (VLM) analyzer for DeepSeek Harness: analyze_image tool with primary/backup OpenAI-compatible endpoints, automatic failover, and an auto-saving web settings page.

Image viewer: a real image tab type for the vncode right bar (dsh-rightbar). PNG, JPEG, GIF, WebP, AVIF, BMP, ICO, SVG and TIFF open in a viewer that FITS the pane, zooms on a ladder (5%-800%) with fit-to-pane and 100%-actual-pixels, wheel-zooms at the pointer with Ctrl/Cmd, and is dragged to pan with a real grab cursor - the image is laid out at its true size in a scrollable box, so a zoom moves the layout and never a CSS transform, and a zoomed picture stays scrollable to its edge. A checkerboard shows PNG transparency, pixelated rendering past 300% makes pixel inspection honest, and the status line names the file's dimensions, size, format and the pixel under the pointer with its colour. Bytes are read through the harness's own workspaceFiles remote, so no new host route, no new surface and nothing shipped is disabled: the type registers in the extension band and REPLACES the shipped preview's bare intrinsic-size image only for image addresses. Alpha.

Image-generation tool plugin for DeepSeek Harness: qwen-image-3.0 (Aliyun MaaS) and doubao-seedream (Volcano Ark) providers

GPT Image 2.5 generation through Codex subscriptions or API keys, with progressive previews for DeepSeek Harness

让纯文本主模型（DeepSeek V4 等）也能接收图片附件：图片投影为显式占位文本，由主模型委托视觉子代理经 read_attachment/read_image 读取。零核心补丁，纯插件实现。

Kimi WebBridge for DeepSeek Harness: drive the user's real browser (navigate, click, fill, snapshot, screenshot, evaluate, network, upload, PDF) through the local Kimi WebBridge daemon

DeepSeek Harness plugin that turns a local llama.cpp (llama-server) into a first-class model provider: process & router management, model catalog sync with mmproj vision pairing, auto-start on demand, stop-then-load model switching, and a sidebar terminal-output monitor panel.

DeepSeek vision bridge for the dsh web GUI: route image attachments to a vision model (pi-ai / llama.cpp Qwen3-VL) and continue the conversation with the text description on a text-only LLM (DeepSeek)

DeepSeek 小鲸鱼姬桌宠与摸鱼伴侣插件 (Chibi Whale Maid Desktop Companion & Relaxation Hub for DeepSeek Harness)

Media the agent can actually work with: media_probe (ffprobe's reading of any image, video or audio file, translated into a report and a browser-playability verdict with the exact ffmpeg command that would fix it), media_run (a real ffmpeg or ffprobe command as an ARGV array - no shell, a deadline, an output cap, and no overwrite unless asked), and media_frames (frames out of a video at named timestamps, sampled evenly, or tiled into one contact sheet). Ships the pin for a static ffmpeg/ffprobe copy per platform (SHA-256 verified, installed into $DSH_HOME/dsh-media/bin on first need, used only when the machine has none of its own), two bundled skills (ffmpeg-cli, ffprobe-cli), and the host routes dsh-video's tab reads: a Range-capable file stream, the same probe summary as JSON, and a background remux/transcode that turns a container a browser cannot open into a cached MP4.

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

DSH 插件：给非官方（自定义）提供方的模型自动填充推理级别、最大上下文、输出上限与图片模态等参数（数据来自 models.dev），并提供兼容性提升、记住推理级别等用户体验优化。

模型信息补全：按模型名自动补全上下文、输出上限、思考档位和图片能力

Multimodal eyes and hands for DeepSeek Harness: vision transcription, OCR, and text-to-image via OpenAI-compatible backends, with an in-conversation generated-image card.

Local OCR for DeepSeek Harness: paste/attach an image, get its text via PP-OCRv5 + ONNX Runtime, fully offline. TUI (cc-tui) and Web. / DeepSeek Harness 本地 OCR 插件：图片转文字，PP-OCRv5 + ONNX Runtime，完全离线，支持 TUI 与 Web。