dsh-media-skills
适用于 DeepSeek Harness v0.1.6 的免费图像读取和生成技能：视觉问答、插图生成、提供商回退，且不附带密钥。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:MJorgin/dsh-media-skills#f71f9517390b35ddfadbeae8b038ff0fd9d8f59c
Media Skills 以 dsh-media-skills 发布，当前版本为 0.4.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A lightweight DSH plugin that contributes two bring-your-own-key media skills:
vision-review — describe images, run OCR, inspect screenshots, detect UI issues such as overlapping or overflowing text, and output optional structured evidence.media-tools — generate illustrations, avatars, backgrounds and banners through SenseNova U1 Fast or SiliconFlow Kolors.The plugin uses the DSH v0.1.6 skill-provider lifecycle and does not patch DSH core, modify model settings, register providers, or write hidden configuration.
Why · Install · Configure keys · Usage · Manual directory install · Verification · FAQ
English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Deutsch · Português · Русский
DeepSeek Harness v0.1.6 already supports modern image attachments and file workflows for models that accept image input. This bundle focuses on two complementary jobs that remain useful after native attachment support:
| Need | Skill | How it helps |
|---|---|---|
| Explicit screenshot QA | vision-review | Checks rendering completeness, overlap, overflow, misalignment, watermarks and visual consistency. |
| OCR and image-to-text | vision-review | Turns screenshots, photos and scanned content into text, with an optional structured JSON contract. |
| Provider failover | vision-review | Uses configured engines in a predictable chain and reports each failed attempt. |
| Image asset production | media-tools | Generates usable image files through a configured SenseNova or SiliconFlow key. |
The plugin leaves model routing to DSH and its Models UI. That makes it compatible with v0.1.6 runtime enabling, disabling, uninstalling and restarting without leaving global state behind.
ctx.skills.registerProvider(...).llm-pi-ai settings mutation and all model-route seeding.SKILL.md, preventing copy drift.<= v0.1.1-rc.2; they are not required for v0.1.6.Open DSH's Plugin Manager and add:
github:MJorgin/dsh-media-skills
Then restart the profile.
For the common web profile:
dsh plugin --profile web add github:MJorgin/dsh-media-skills
Replace web with the DSH profile you actually use. Restart that profile after installation so the new bundle is mounted.
No build step is required: the package ships ready-to-run ESM and Python scripts and has no dependency installation or prepare script.
Keys are never stored in this repository. Skill scripts read environment variables first, then:
~/.dsh/secrets/media-tools.env
~/.codex/secrets/media-tools.env # legacy fallback
vision-review can also read compatible keys from:
~/.dsh/.credentials.yaml
media-tools reads environment variables and the two media-tools.env files; configure its keys explicitly in one of those locations.
| Key | Used by | Notes |
|---|---|---|
GLM_API_KEY | vision-review | Primary Zhipu glm-4v-flash engine; verify current Zhipu pricing/free-tier terms. |
DEEPSEEK_API_KEY | vision-review | Optional paid DeepSeek vision model; also read from DSH credential storage. |
SILICONFLOW_API_KEY | Both skills | Qwen3-VL for review and Kolors for generation. |
SENSENOVA_API_KEY | Both skills | SenseNova vision model for review and U1 Fast for generation. |
GEMINI_API_KEY | vision-review | Optional Gemini fallback; may require GEMINI_PROXY on some networks. |
Example secrets file:
# ~/.dsh/secrets/media-tools.env, chmod 600
GLM_API_KEY=...
SILICONFLOW_API_KEY=...
SENSENOVA_API_KEY=...
GEMINI_API_KEY=...
This plugin does not add a model to the DSH model picker. To have a normal DSH conversation natively accept images, configure a multimodal model/provider in DSH's Models settings. Then use DSH's native attachment flow.
Use vision-review when you want a dedicated scripted review/OCR workflow, a provider failover chain, or structured evidence instead of only a multimodal chat response.
Ask DSH to use vision-review, or run its script from the skill directory:
python3 scripts/vision.py screenshot.png
python3 scripts/vision.py a.png b.png --structured
python3 scripts/vision.py screenshot.png --provider=siliconflow-qwen
python3 scripts/vision.py --doctor
The default prompt checks rendering completeness, overlapping/misaligned/overflowing text, color hierarchy, watermarks and obvious visual bugs. For a specific task, pass a focused prompt:
python3 scripts/vision.py page.png --prompt="检查按钮、标题和图表是否有重叠，并指出具体位置"
The failover chain joins engines only when their keys are available. Optional --structured output includes summary, OCR, reading-order layout, semantics, visual notes and uncertainty.
python3 skills/media-tools/scripts/generate.py "a cinematic Chinese palace in the clouds, realistic, grand scale" palace.jpg 16:9
SenseNova is used when SENSENOVA_API_KEY is available. Otherwise the script uses SiliconFlow Kolors when SILICONFLOW_API_KEY is available. SenseNova sizes can be supplied as exact dimensions or common ratios; the script maps them to the nearest supported size.
Plugin installation is recommended because this repository contains multiple skills. The v0.1.6 filesystem provider scans only one directory level under a skill root, so cloning the repository directly into ~/.dsh/skills/ will not discover nested skills/*/SKILL.md files.
To install manually, link each skill separately:
git clone https://github.com/MJorgin/dsh-media-skills.git ~/.dsh/bundles/dsh-media-skills
mkdir -p ~/.dsh/skills
ln -s ~/.dsh/bundles/dsh-media-skills/skills/vision-review ~/.dsh/skills/vision-review
ln -s ~/.dsh/bundles/dsh-media-skills/skills/media-tools ~/.dsh/skills/media-tools
Restart DSH after creating the links.
Run the complete local checks:
npm test
This runs DSH bundle manifest validation, runtime provider registration/loading through a fake DSH-like context, JavaScript syntax validation, and Python compilation for both skill scripts.
The old core patches remain available for users maintaining legacy DSH builds:
They apply to historical builds through v0.1.1-rc.2. New v0.1.6 users should not apply them.
dsh-media-skills/
├── package.json # DSH bundle manifest and test commands
├── cordis.patch.yml # Cordis plugin insertion
├── index.js # Registers the bundled skill provider
├── skills/
│ ├── vision-review/ # Image analysis and screenshot QA
│ └── media-tools/ # Image generation
├── scripts/ # Bundle validation helpers
├── examples/ # Example images and test card
└── docs/ # Setup guides, translations and historical notes
Do I need a core patch on DSH v0.1.6? No. Configure a multimodal model in DSH for native image conversations, or use the skill scripts for dedicated review and generation workflows.
Does the plugin add a model to the model selector automatically? No. DSH v0.1.6 provides model and plugin management; the plugin only registers skills and never mutates model settings.
Are all providers free? Provider pricing and free-tier policies can change. GLM-4V-Flash and Kolors have been free-tier friendly, while DeepSeek usage is paid. Check the provider's current terms before relying on a workflow.
Are API keys included? No. Keys stay in your environment, DSH credential storage, or local secrets files.
Where should I send sensitive internal screenshots? Only to providers approved by your organization. Avoid Gemini or other external providers for internal documents unless company policy permits them.
More details: examples/README.md.