为 DeepSeek Harness 提供图片识别、视觉检查与图片生成能力，并自动注册可配置的视觉模型。
- 语言
- Python
- License
- MIT
- 分支
- main
安装
$ dsh plugin --profile web add github:akqwpeter-prog/dsh-media-skills在终端中运行以上命令，通过 dsh CLI 安装此插件。可在右上角切换 Profile。 第一次用 dsh？看这篇新手教程
对话式安装
帮我安装 DeepSeek Harness 插件 akqwpeter-prog/dsh-media-skills：先查看仓库 https://github.com/MJorgin/dsh-media-skills 确认安全性，然后执行安装命令并验证插件加载成功。
把这段指令粘贴给 DSH Web GUI 里的助手，由它代你完成安装与验证。
一句话定位
这个插件把读图、视觉检查和图片生成接入 DeepSeek Harness，可在模型选择器使用视觉模型，也可在已适配的宿主中把图片转成文字后交给纯文本模型。API Key 需要由用户自行准备。
核心能力
- 读取本地图片和截图，概括内容、提取文字，并整理版面与对象关系。
- 检查界面截图中的文字重叠、内容溢出、元素错位、配色问题和水印等视觉异常。
- 输出摘要、完整 OCR、阅读顺序、对象关系和不确定性组成的结构化结果，方便后续程序读取。
- 自动缩放并分批处理多张图片，单次读图请求最多发送 5 张，避免超过服务限制。
- 调用 SenseNova 或 SiliconFlow 生成插画、头像、背景和横幅，并保存到指定文件。
- 注册智谱与商汤的视觉模型路线，并在已配置密钥时按智谱、SiliconFlow、SenseNova、Gemini 及自定义服务的顺序故障转移。
技术实现
- 语言: JavaScript ESM + Python 3
- 关键依赖: Python 3 标准库、Pillow 图片处理库、Node.js 标准库、OpenAI 兼容服务接口
- 架构模式: Cordis 插件入口注册技能提供方并补写缺失的模型路线；技能加载本地说明文件，再调用 Python 脚本处理图片或生成图片
- 入口文件:
index.js
适用场景
适合需要查看截图、核对页面排版、提取图片文字，或在聊天中生成素材的用户。日常单图问答可以直接使用视觉模型；批量检查、脚本化处理，以及让纯文本模型理解图片，则更适合走读图技能或已打补丁的贴图链路。
前置依赖与兼容性
| 依赖 | 最低版本 | 说明 |
|---|---|---|
| DeepSeek Harness | 0.1.0-rc.7～0.1.0-rc.8（完整贴图） | 技能和模型路线未声明最低 DSH 版本；纯文本会话自动转述需在对应版本源码应用宿主与客户端补丁 |
| Node.js | 未声明 | 入口使用 ESM，并读取 Node.js 内置文件系统与 URL 模块；仓库没有 npm 运行时依赖 |
| Python 3 | 未声明 | 两个技能脚本由 python3 启动；README 徽章标注 3.9+，但依赖系统未强制校验该版本 |
| Pillow | 未声明 | 读图脚本处理 JPEG 压缩时动态导入 Pillow；缺少时运行 vision-review 的诊断命令会提示手动安装 |
| 服务商 API 与网络 | 不适用 | 至少需要可访问的读图服务与对应 API Key；生成图片还需图片生成服务的 Key |
| 平台 | 跨平台 | 未声明操作系统或 CPU 限制，也不依赖原生模块；实际运行仍取决于 DSH 宿主、Python 和网络环境 |
安装方式
dsh plugin --profile web add github:akqwpeter-prog/dsh-media-skills
配置项
| 配置 | 类型 | 说明 | 默认值 |
|---|---|---|---|
GLM_API_KEY | 密钥（字符串） | 智谱读图服务的访问凭证；未配置时不能使用该读图路线 | 未配置 |
SENSENOVA_API_KEY | 密钥（字符串） | 优先用于图片生成，也可在读图主路线失败时使用 | 未配置 |
SILICONFLOW_API_KEY | 密钥（字符串） | 用于图片生成，并在读图主路线失败时参与故障转移 | 未配置 |
GEMINI_API_KEY | 密钥（字符串） | 智谱或其他备用服务失败后，尝试使用 Gemini 读图 | 未配置 |
SENSENOVA_VISION_MODEL | 模型名（字符串） | 覆盖商汤读图时使用的模型 | sensenova-6.8-flash-lite |
SILICONFLOW_VISION_MODEL | 模型名（字符串） | 覆盖 SiliconFlow 读图时使用的模型 | Qwen/Qwen3-VL-8B-Instruct |
GEMINI_MODEL | 模型名（字符串） | 覆盖 Gemini 读图时使用的模型 | gemini-3.6-flash |
GEMINI_PROXY | 代理地址（字符串） | 仅供 Gemini 联网使用；未设置时会沿用系统 HTTPS_PROXY | 未显式设置 |
VISION_FALLBACKS | JSON 数组 | 追加自定义读图服务；每项提供服务地址和模型，可选填写密钥变量名、输出上限及结构化输出开关 | [] |
脚本优先从环境变量读取密钥，其次读取 ~/.dsh/secrets/media-tools.env，最后兼容读取 ~/.codex/secrets/media-tools.env。插件不会把密钥写入仓库；视觉模型路线使用 DSH 凭据存储中的同名变量。
常见问题
Q: 安装后需要额外配置吗？
A: 需要。读图至少要配置智谱 API Key，图片生成还要配置 SenseNova 或 SiliconFlow API Key；其他服务可按需用于故障转移。密钥应放在环境变量或本机凭据文件中，不要写入仓库。
Q: 纯文本模型可以直接贴图吗？
A: 可以在已适配的 DSH 宿主中实现。仓库只提供模型路线、读图与生图技能；直接贴图还需要在 DSH 0.1.0-rc.7 或 rc.8 对应源码应用匹配的宿主和客户端补丁，否则纯文本会话可能继续拒绝图片。
Q: 图片会保存在哪里？
A: 读图脚本不会把输入图片保存到仓库或指定输出目录。直接贴图时，宿主会把原图留在会话附件库；生成图片只写入调用者给定的目标路径。
Q: 图片会发送给谁？
A: 图片会发送到当前读图服务，生成任务会发送到 SenseNova 或 SiliconFlow。Google 免费服务的请求数据可能被用于改进其产品，处理证件、内部资料或客户截图前应先确认服务商条款。
Q: 是否支持 macOS、Windows 和 Linux？
A: 插件没有操作系统或 CPU 限制，也没有原生模块。系统需要能够运行 DSH、Python 3 和 Pillow，并能访问所选服务；其余限制来自 DSH 宿主机。
Q: 报 1210 或输入输出总量超限怎么办？
A: 智谱读图服务要求单次输出不超过 1024，输入与输出合计不超过 16384。插件已按该限制配置；长会话可新建后重试，并确认自动注册的模型配置没有被改成更高的输出上限。
Q: 读图服务不可用时如何排查？
A: 可运行 vision-review 脚本的 --doctor 选项，检查 Python、Pillow、密钥和各个服务的连通性。也可配置第二个视觉服务或通过 VISION_FALLBACKS 追加自定义路线。
Q: 如何卸载？
A: 可运行 dsh plugin --profile web remove dsh-media-skills。插件不会主动删除已写入 DSH 设置的模型条目；不再使用时可手动删除对应模型配置并重启 DSH。
上手难度
进阶 — 技能调用本身简单，但完整贴图能力需要准备 API Key、理解 DSH 模型设置，并在对应版本源码中手动应用宿主与客户端补丁。
已知问题与限制
- 直接贴图自动转述不在插件代码内，仍依赖 DSH 本体的视觉转述补丁；仓库提供的补丁只验证了 rc.7 与 rc.8，不能混用不同版本补丁。
- 智谱单次读图请求最多处理 5 张图片；脚本会把更多图片拆成多批，但智谱输出上限为 1024，输入与输出合计上限为 16384。
- 仓库没有声明 Python 依赖，也不会自动安装 Pillow；缺少 Pillow 时读图功能无法完成图片压缩。
- 自动注册模型路线时，只会在对应路线不存在时补写智谱和 SenseNova 配置，不会写入 API Key，也不会替用户配置 SiliconFlow 或 Gemini 路线。
- 随附的 DSH 补丁文档指出，图片投影逻辑仍缺少单元测试；现有证据主要来自补丁回环检查和指定包的类型检查。
🎨 dsh-media-skills
Give DeepSeek Harness eyes — and a brush. Read images in any chat, generate new ones, all with free models.
DeepSeek Harness is brilliant at reasoning — but a text-only model can't see the image you just dragged into the chat. This bundle fixes that with two free skills, a free vision model route, and a vision engine failover chain:
- 📎 Paste to read — paste, drag, or pick an image in any session; the free vision model turns it into text your current model understands. (Powered by the DeepSeek Harness core auto-description path — see docs/HARNESS_PATCH_EN.md; this bundle contributes the vision model route and the skill it relies on.)
- 👁️
vision-review— analyze images and screenshots, catch UI visual bugs, detect watermarks, turn images into text. - 🎨
media-tools— generate illustrations, avatars, backgrounds and banners with a free, watermark-free model. - 🔀 Engine failover — GLM-4V-Flash → SiliconFlow Qwen3-VL → Google Gemini (AI Studio) → any OpenAI-compatible endpoint, with ModLens-style structured evidence output.
No hardcoded keys, no paid API, no file saving, no session switching.
Why · Quick start · See it in action · Usage · Keys & privacy · FAQ · Examples
English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Deutsch · Português · Русский
🤔 Why
Most DSH vision plugins only read images — and many push you through a shared third-party endpoint. dsh-media-skills takes a different stance:
| This bundle | Typical vision-only plugin | |
|---|---|---|
| Read images for free | ✅ Zhipu GLM-4V-Flash | ✅ |
| Generate images for free | ✅ SiliconFlow Kolors | ❌ usually absent |
| Auto model route in the picker | ✅ installed automatically | sometimes |
| Keys committed to the repo | ❌ never — keys stay local | ⚠️ often required |
| Docs in multiple languages | ✅ 9 languages | ❌ usually English only |
| Privacy | ✅ you choose the provider; images only go to your provider | shared free endpoints can see your images |
Why bring your own free key instead of a built-in anonymous endpoint? Privacy and reliability. Your images go only to the provider you choose, under your account and your rate limits — no shared third-party service in the middle.
✨ What you get
| Capability | What it does | Model | Cost |
|---|---|---|---|
| 🖼️ Paste-image reading | In a text-only session, paste, drag, or pick (add-image button, restored by the client-ux patches) an image into the composer; it is described by the vision model (GLM-4V-Flash with SiliconFlow Qwen3-VL failover, 15s per route) and handed to the current model as text beside a live thumbnail. (Harness-core feature on rc.7/rc.8: requires the api-proxy admission patch + the rc.8 client-ux patch — see docs/HARNESS_PATCH.md / HARNESS_PATCH_EN.md, patch files included for rc.7 and rc.8; this bundle supplies the vision route + skill it depends on) | GLM-4V-Flash + Qwen3-VL | Free |
| 🧠 Vision model route | 「智谱 GLM-4V-Flash（视觉）」 appears in the model selector automatically — pick it for a new conversation and talk about images directly | Zhipu GLM-4V-Flash | Free |
👁️ vision-review | Analyze / recognize / describe images & screenshots; catch UI visual bugs (overlap, overflow, misalignment); detect watermarks/logos; turn images into text. Optional --structured mode returns ModLens-style evidence JSON (summary, full OCR, reading-order layout, entities/relations, uncertainty). Engine failover chain: GLM-4V-Flash → SiliconFlow Qwen3-VL / Google Gemini (auto-join with free keys) → any OpenAI-compatible endpoint | GLM-4V-Flash + Qwen3-VL + Gemini | Free |
🎨 media-tools | Generate images, illustrations, avatars, backgrounds, banners | SiliconFlow Kolors | Free, no watermark |
⚡ Quick start
dsh plugin --profile <name> add github:MJorgin/dsh-media-skills
-
Get two free keys (~2 minutes, no payment):
- Zhipu — open.bigmodel.cn → API Keys (
glm-4v-flashis free) - SiliconFlow — siliconflow.cn → API Keys (Kolors is free)
- (optional third) Google Gemini — aistudio.google.com → Get API key; joins the vision failover chain automatically
- Zhipu — open.bigmodel.cn → API Keys (
-
Add them in the Web GUI (Settings → Models → the zhipu-vision provider's API Key field), or use the credentials file:
# ~/.dsh/.credentials.yaml (chmod 600) GLM_API_KEY: <your key> -
Restart
dsh web, then hard-refresh (Cmd+Shift+R).
Verify: the model selector shows 智谱 GLM-4V-Flash（视觉）. If your Harness build supports paste-image reading, the input bar also has a 📎 Add image button — paste an image in any session and it arrives as a text description.
Full walkthrough and troubleshooting: docs/SETUP_VISION_EN.md.
📸 See it in action
Paste an image in a text-only session → the free vision model describes it → your model answers. The same bundle also generates new images on demand.
How it works in one picture:
🚀 Usage
Three ways to read images:
| Way | How | When |
|---|---|---|
| A. Paste directly (recommended) | In any session, click the 📎 button / drag / paste an image and send | Everyday image questions — no file saving, no model switching |
| B. Vision model session | New conversation, pick 智谱 GLM-4V-Flash（视觉）, paste images and chat | Multi-turn image conversations, native read_image |
| C. Files + skill | Put the image in the workspace and say “read this image with vision-review” | Batch review, scripted workflows |
Descriptions follow your message language (Chinese message → Chinese description; English message → English description; no text → Chinese).
Also just say:
- “Look at this image / check this screenshot for visual bugs” →
vision-review - “Generate an image of …” →
media-tools
🔑 Keys & privacy
Keys are never stored in this repo. Skill scripts read, in order: environment variables → ~/.dsh/secrets/media-tools.env → ~/.codex/secrets/media-tools.env (legacy fallback). The vision model route reads GLM_API_KEY from DSH's credential store.
Where to get the keys (all free): Zhipu — open.bigmodel.cn → API Keys (glm-4v-flash). SiliconFlow — siliconflow.cn → API Keys (Kolors). Google (optional, joins the vision failover chain automatically) — aistudio.google.com → Get API key.
# ~/.dsh/secrets/media-tools.env (chmod 600, one KEY=value per line)
GLM_API_KEY=...
SILICONFLOW_API_KEY=...
GEMINI_API_KEY=... # optional
Your images are sent only to the provider you configure — never to this repo, never to a shared anonymous endpoint.
Privacy note on Gemini: Google's free-tier key comes with data-use terms — requests may be used to improve Google products. For sensitive images (IDs, internal docs, customer data), prefer the direct domestic engines (Zhipu / SiliconFlow).
❓ FAQ
Does paste-image reading require a DeepSeek Harness core patch?
The auto-describe pipeline lives in the Harness core (api-proxy image-admission logic; see docs/HARNESS_PATCH_EN.md). This bundle ships the model route + skills: the vision model works on any DSH build, but paste-image reading requires a Harness build with that core support — see FAQ Q1 in docs/SETUP_VISION_EN.md.
Why not just use a built-in free endpoint with no key at all? We prefer to let you own the route: your images go to the provider you pick, under your rate limits, with no shared middleman. The keys are free and take about two minutes to create.
Is media-tools really free?
Yes — SiliconFlow Kolors is free and watermark-free. If a model is temporarily disabled, the skill lists available models and you can switch.
🎁 Examples
Sample material to try instantly — 6 AI-generated images with their prompts, plus a purpose-built vision test card (title, buttons, bar-chart values) for checking reading accuracy:
🗺️ Layout
dsh-media-skills/
├── package.json # dsh.bundle manifest
├── cordis.patch.yml # plugin layer
├── index.js # registers skills + seeds the zhipu-vision model route
├── skills/
│ ├── vision-review/ # image reading
│ └── media-tools/ # image generation
├── examples/ # sample images + vision test card
├── docs/
│ ├── screenshots/ # demo mockup & how-it-works diagram
│ ├── SETUP_VISION_EN.md # detailed setup guide (English)
│ ├── SETUP_VISION.md # 详细配置指南（中文）
│ ├── HARNESS_PATCH_EN.md# core patch notes (English)
│ ├── HARNESS_PATCH.md # 本体补丁说明（中文）
│ ├── COMPARE_MODLENS.md # 与 ModLens 的对比/共存（中文）
│ └── lang/ # READMEs in 9 languages
├── scripts/make-banner.py # regenerates docs/social-preview.png
└── docs/social-preview.png
🧩 Using ModLens alongside?
Both this bundle and ModLens give text-only models vision. Installed together they do not conflict: ModLens intercepts pastes first (path → modlens_read_image tool), and this bundle's api-proxy fallback handles anything it doesn't take over. See docs/COMPARE_MODLENS.md (中文) for the full comparison, the paste routing order, and how to point ModLens at the same free Zhipu endpoint.
🤝 Join the DSH plugin ecosystem
DeepSeek Harness developer preview is still in its testing phase for Harness developers; core plugins and base APIs will keep iterating. We look forward to exploring the upper limits of intelligence together with developers worldwide, on top of open-source, open, reusable, and composable infrastructure.
- dsh-plugin topic
- Quickstart
- DeepSeek Harness repo
- dsh-agent-conductor — 同作者的指挥家：在 DSH 里派活给 11 种外部 agent CLI（Codex / Claude Code / TraeCode…）
This repo is tagged
dsh-pluginand listed in the awesome-dsh-plugin curated list. PRs, issues and translations are welcome.
📄 License
查看使用指南 →
该插件的安装步骤、关键要点、FAQ 与兼容性说明（基于已收录字段派生）。
收录徽章
[](https://deepseek-plugin.org/plugins/akqwpeter-prog/dsh-media-skills)把这段 markdown 粘贴到你的 GitHub README，链接回本插件详情页。徽章只声明已被本站收录，不代表安全认证。