YOGEMOW/DeepSeek_Prism
30Last commit Aug 15, 2026
DeepSeek_Prism DSH plugin
DeepSeek_Prism is a DSH plugin that enables text-only DeepSeek models (e.g., deepseek-v4-flash) to understand images. It calls an external vision API to extract visible facts, compresses them into a compact VEP/1 evidence packet, and feeds it back to the main model for reasoning. The plugin supports automatic image scaling, multi-provider failover, and local caching.
How to install the DeepSeek_Prism DSH plugin
dsh plugin --profile <name> add @yogemow/deepseek-prism-dshThis source command needs manual review. Copying does not run it.
DeepSeek_Prism DSH plugin data source
DeepSeek_Prism DSH plugin snapshot date: Aug 16, 2026
discovered
What the DeepSeek_Prism DSH plugin can do
- Trigger forced recognition when the model cannot read an image (Unsupported format / binary).
- Produce VEP/1 compact evidence output (≤520 chars, ~50–150 tokens) by default.
- Support five detail modes for structured reports: page restore, issue location, error logs, text tables, chart data.
- Auto-scale images over 2048px using built-in sharp (libvips) before upload, supporting AVIF/TIFF/SVG conversion to PNG.
- Provide multi-provider presets (SiliconFlow, Zhipu, ModelScope, Alibaba, OpenRouter, Groq) with automatic fallback.
- Cache SHA-256 results with 24-hour TTL and 1000-entry limit, with `--no-cache` bypass option.
Where the DeepSeek_Prism DSH plugin fits
- Enable a text-only model to read screenshots and extract UI elements or error messages.
- Analyze charts and graphs from images, returning structured data for further processing.
- Process scanned documents or tables where the model cannot directly parse the image format.
- Automatically convert uploaded images in a conversation into textual evidence for the model.
- Debug vision API calls and test different providers without modifying the main model.
Who the DeepSeek_Prism DSH plugin is for
- Developers using DeepSeek Harness with text-only models who need occasional image understanding.
- Users who want to avoid switching to a multimodal model but still benefit from image analysis in their workflow.
DeepSeek_Prism DSH plugin limitations
- Requires an external vision API key (e.g., SiliconFlow API key) – not a local vision model.
- Only supports Node.js >= 18; no other runtime dependencies but sharp may be missing in pure CLI environments.
- VEP output is a compressed summary, not a full image description; detail may be lost for complex scenes.
- The plugin is still in early development (v0.7.1) and may have edge cases with large or animated images.
DeepSeek_Prism DSH plugin: from the repository README
Quoted from the YOGEMOW/DeepSeek_Prism README, the upstream source of the DeepSeek_Prism DSH plugin. Copyright remains with the original authors.
为纯文本 DeepSeek 模型（如 `deepseek-v4-flash`）提供按需识图能力的 Codex / DeepSeek Harness（DSH）双平台 Skill：图片由外部视觉 API 提取可见事实，压缩为低 Token 的 VEP/1 视觉证据包回传主模型，由主模型继续完成推理、规划与决策。 ## 版本选择 | 版本 | 定位 | 说明 | | --- | --- | --- | | **[v0.7.1](https://github.com/YOGEMOW/DeepSeek_Prism/releases/tag/v0.7.1)** | **零补丁版** | DSH 当前主线：自包含 Cordis 组合包，`prism_see` 工具 + 纯文本模型图片准入 VEP 降级 + 技能运行时注册 + 设置卡片；harness 本体零改动、上游更新零冲突、卸载零残留 | | **[v0.6.2](https://github.com/YOGEMOW/DeepSeek_Prism/releases/tag/v0.6.2)** | **实用版** | DSH 完整 UI 体验：原图保留展示 + VEP 折叠链接/识别进度卡片 + Web 设置卡片可编辑；需应用 `harness-patch/dsh-prism-harness.patch`（可选增强） | | **[v0.2.0](https://github.com/YOGEMOW/DeepSeek_Prism/releases/tag/v0.2.0)** | **Codex 推荐 skills 版** | Codex 平台的技能本体（SKILL.md + `scripts/vision.mjs` + references），安装到 `~/.codex/skills/deepseek-prism` | ## 功能 - 强制触发协议：主模型无法直接读图（Unsupported format / 无法读取 / binary）时，立即调用 `scripts/vision.mjs` 识图。 - VEP/1 紧凑证据：默认输出 ≤520 字符（约 50–150 tokens），字段按优先级裁剪。 - `--detail` 五模式分节报告：页面还原（A1–A7）/ 问题定位 / 报错日志 / 文本表格 / 图表数据。 - 自动分级：小图/简单任务默认 VEP/1；长内容（代码截图、长日志、文档、宽/高比大的图）自动走 `--detail` 完整通道；超长内容自动续写并合并。 - 程序化输出：`--raw` 输出清洗后的原文；`--full` 输出 `{raw, parsed}` JSON 信封。 - 输入预处理：解析图片宽高（含 AVIF/TIFF/SVG 的 sharp metadata 回退），超过
Read the full READMERepository license: MIT
DeepSeek_Prism DSH plugin questions
How do I install DeepSeek_Prism for DSH?
Run the command `dsh plugin --profile <name> add @yogemow/deepseek-prism-dsh` in your terminal. Replace `<name>` with your DSH profile name (e.g., `web`). The plugin will install the latest version (v0.7.1) automatically. After installation, restart your DSH web service for the changes to take effect.
Do I need an API key to use this plugin?
Yes, you need an API key from a supported vision provider. The plugin defaults to SiliconFlow, so you'll need to set `SILICONFLOW_API_KEY` either as an environment variable or in a `.env` file. You can also configure other providers like Zhipu, ModelScope, or Groq using the `VISION_PROVIDER` and `VISION_API_KEY` environment variables.
What image formats are supported?
The plugin supports common image formats like PNG, JPEG, GIF, WebP, and also handles AVIF, TIFF, and SVG by converting them to PNG using the sharp library. For animated GIFs, it preserves all frames during scaling. The image must be accessible via a local file path or a URL.
What is VEP and why does it matter?
VEP stands for Visual Evidence Packet. It's a compact format that compresses visible facts from an image into a low-token representation (typically 50–150 tokens) so the text-only model can use the information without exceeding context limits. The default VEP/1 output is a structured summary with fields like source, model, action, target, summary, and details.
Can I use this plugin without the DSH web interface?
Yes, the plugin works in a CLI environment as well, but you need to ensure the `prism_see` tool can be invoked by the model. If you don't have a web interface, you can still install the plugin and use it via the DSH CLI. However, the automatic image upload conversion in the chat requires the web frontend; for CLI, you'll need to manually call `vision.mjs see` with the image path.