dsh-read-image-jpeg-fallback
DSH 配置插件：将 read_image 的 PNG/WebP 附件转换为不透明的 sRGB JPEG，以便 LM Studio 的 openai-completions 端点接受它们
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:wuyiwuaa1-ship-it/dsh-read-image-jpeg-fallback#864761d82d8cfdbc63843faf8507f418dbe8c317Read Image Jpeg Fallback 以 dsh-read-image-jpeg-fallback 发布，当前版本为 0.1.1。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。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →A zero-configuration DSH image compatibility shim for LM Studio.
It keeps DSH's native read_image + multimodal-model flow intact, but transparently converts PNG/WebP tool attachments to opaque sRGB JPEG before the next provider request is built.
Use it when your main model already supports images, but LM Studio's openai-completions route rejects a DSH image attachment with an error such as:
400 'url' field must be a base64 encoded image
This is a media-format compatibility plugin, not a vision model, OCR plugin, or image-to-text fallback.
In some DSH + LM Studio paths, read_image can produce a PNG/WebP image block that later reaches the provider request in a format LM Studio does not accept.
This plugin fixes only that transport mismatch:
read_image
↓
DSH image attachment
↓
PNG / WebP ?
├─ no → pass through unchanged
└─ yes → opaque sRGB JPEG (quality 90)
↓
same multimodal model
The model still receives the image itself. There is no OCR step, no auxiliary VLM call, and no conversion to a text description.
| Approach | What happens to the image | Main model path |
|---|---|---|
| This plugin | PNG/WebP is re-encoded to JPEG | Preserved; the same multimodal model receives the image |
| Vision fallback plugins | Image is sent to another VLM and usually converted to text | Main model receives a description/result |
| OCR / image-reader plugins | Image is analyzed by dedicated tools | Main model consumes tool output |
| Separate LM Studio vision tools | Agent calls a dedicated vision tool/model | Bypasses the native read_image → current-model path |
This plugin is intentionally narrow: it is useful when the native DSH vision path is already the behavior you want and only the image encoding is incompatible.
The plugin hooks into DSH after a successful read_image tool call and converts model-visible image blocks as follows:
image/png → opaque sRGB JPEGimage/webp → opaque sRGB JPEGIt does not replace the built-in read_image tool or patch DSH core source.
The conversion itself is local. The resulting JPEG then follows your existing DSH/provider request path.
This plugin is a good fit when all of the following are true:
openai-completions route;read_image works until the following model request is built, then fails on the image attachment;You probably do not need it if your provider already accepts the image formats DSH sends, or if your main model is text-only and you actually need a vision fallback/OCR solution.
Requires:
0.1.1-rc.2>= 22pnpm available on PATHdsh plugin --profile web add dsh-read-image-jpeg-fallback
dsh plugin --profile web add github:wuyiwuaa1-ship-it/dsh-read-image-jpeg-fallback
To pin the current main branch explicitly:
dsh plugin --profile web add github:wuyiwuaa1-ship-it/dsh-read-image-jpeg-fallback#main
dsh plugin --profile web add .\dsh-read-image-jpeg-fallback-0.1.1.tgz
Restart DSH Web after installation.
No configuration is required.
Use read_image normally. When its successful result contains a PNG or WebP image block, the plugin creates a JPEG replacement for the model-visible attachment before the next provider request is built.
There is no new model-facing tool to learn and no change to the normal agent workflow.
The plugin is designed to fail open.
If reading, converting, or saving the replacement image fails:
read_image result is preserved;Verified with:
0.1.1-rc.2openai-completionssharp 0.35.3Compatibility with other DSH versions, operating systems, or providers is not guaranteed unless explicitly tested.
On some Windows + pnpm setups, removing the plugin may hang after pnpm has already completed its filesystem changes, particularly when sharp is being pruned.
If this occurs, pinning sharp directly in the DSH profile avoids the issue:
cd <DSH_HOME>\profiles\web
pnpm add sharp@0.35.3
dsh plugin --profile web remove dsh-read-image-jpeg-fallback
Restart DSH Web afterwards.
The plugin uses DSH's tools/post-execute hook.
For accepted read_image results containing PNG or WebP image blocks, it:
sharp to opaque sRGB JPEG at quality 90;The canonical tool result and built-in read_image implementation remain untouched.
The goal is deliberately small: provide a transparent image-format compatibility layer until the provider/host path accepts the original media format directly.
If upstream DSH or LM Studio gains equivalent handling, this plugin may become unnecessary for that configuration.
MIT