@dsh-external/dsh-pseudo-vision
为仅支持文本的提供商添加可选择启用的图像兼容同级路由，并在委托前将图像块转换为受预算限制的本地增强 OCR、颜色统计、像素扫描和元数据证据。
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:DDDFXYqiming/dsh-pseudo-vision#484202492027a89f2bd64fec80dd52b46d3c7fc2Pseudo Vision 以 @dsh-external/dsh-pseudo-vision 发布，当前版本为 0.7.0。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 →简体中文 | English
Gives text-only providers in DeepSeek Harness a layer of "tool vision". Image attachments are broken down into OCR text + color statistics + pixel scan + metadata on the LLM dispatch path, so any text-only model can read an image through those words. Everything runs locally, with only local OCR and pixel statistics.
deepseek-official route, which already handles images natively. Other text-only providers get sibling routes named dsh-pseudo-vision/<provider> through the bridgeProviders whitelist (shown as · Pseudo Vision in the model selector), or all at once via bridgeOtherProviders.inputModalities: ["text", "image"], which is what lets the request pass the host image-admission gate.<pseudo-vision-context>, then delegates to the original provider.fullEvidenceTurns (default 2) user turns get the full pipeline; older-turn images degrade automatically to compact evidence (metadata + colour + scan, no OCR) with a vision_ocr(file_path=…) re-fetch pointer, so the model can pull the text back on demand. Re-attaching an old image in a new message restores its full tier.maxImages (default 8) bounds full-tier conversions, maxTotalEvidenceChars (default 96 000 characters, ≈24K tokens) bounds the combined evidence text. Over the limit the plugin degrades instead: unconverted images keep an explicit [图片 N 未转换…] placeholder and a [⚠️ 图片处理摘要] summary line tells the model exactly which images did not take effect.| Tool | Purpose | Implementation |
|---|---|---|
vision_ocr | Extract every text line (with normalized coordinates) | tesseract.js (chi_sim + eng), pipeline below |
vision_color_stats | 9-bucket pixel share + average luminance | sharp + histogram |
vision_pixel_scan | Row + column multi-bucket scan emitting focusY/focusX | sharp raw pixel access |
vision_meta | Dimensions, format, color space, 4-corner + center samples | sharp metadata |
通 知 → 通知) and strips leading icon symbols.[数字复核 N 处] audit block records each fix.The key point in the OCR pipeline is that the tesseract.js PSM argument must be a Number; the string
"3"breaks full-page detection.
The GitHub route is the recommended one, assuming network access. When schannel or pnpm blockers hit on Windows, fall back to a local path.
# GitHub (recommended)
dsh plugin --profile web add github:DDDFXYqiming/dsh-pseudo-vision
# Local path when schannel / pnpm blockers hit on Windows
git clone https://github.com/DDDFXYqiming/dsh-pseudo-vision.git
cd dsh-pseudo-vision && pnpm install && pnpm build
dsh plugin --profile web add <local absolute path>
The GitHub route triggers the prepare script, which builds lib/ from source. On the first add, pnpm >= 10 refuses to run build scripts of git dependencies: copy the exact package key pnpm prints into the profile's pnpm-workspace.yaml, then re-run add:
allowBuilds:
'@dsh-external/dsh-pseudo-vision': true
Treat this approval as "let this package run code on your machine at install time". Pin a commit (github:DDDFXYqiming/dsh-pseudo-vision#<sha>) if you want later pushes to stop changing what gets built.
It works out of the box with the default configuration. The deepseek-official route keeps handling images natively. Other providers get no sibling route by default, so opt in explicitly.
- id: dsh-pseudo-vision
config:
bridgeProviders: ["kimi-for-coding"] # only this provider gets a sibling route
ocrBudget: auto # also small | normal | large | mega
ocrNoResize: false # true: skip budget resize/upscale
evidenceMaxChars: 32000 # per-image character cap on model-visible evidence text
# tessdataDir: "D:/tessdata" # offline traineddata dir; wins over the PV_TESSDATA env fallback
maxImages: 8 # full-tier image count cap per request (1-32)
maxTotalEvidenceChars: 96000 # combined evidence character cap (16000-320000)
fullEvidenceTurns: 2 # recent user turns kept full; older degrade (1-8)
You can also set bridgeOtherProviders to bridge every provider except the excluded list in one go. The tradeoff is one extra entry per model in the selector, so think before turning it on.
bridgeOtherProviders: true
Once configured, sibling routes show up as dsh-pseudo-vision/<provider> (· Pseudo Vision). Text-only models go through the local pseudo-vision conversion, while native vision models stay untouched.