good-boy4069/dsh-vision-guard ↗★ 0
dsh-vision-guard
Transparent image guard + vision analysis for DeepSeek Harness: text-only models read pasted images without the 400 session deadlock.
AI 분석
核心用途是为不支持多模态的纯文本模型补充图片读取能力。适合经常需要向 Agent 粘贴截图，且需要限制每日视觉调用额度以防超支的用户。
설치
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:good-boy4069/dsh-vision-guardConfiguration
All fields optional (defaults shown). The vision route must point to a model that accepts image input:
| Field | Default | Meaning |
|---|---|---|
visionProvider / visionModel | opencode-go / minimax-m3 | The vision route. Point it at an image-capable model on your subscription |
ocrTimeoutMs | 45000 | Per-image OCR timeout |
budgetPerDay | 200 | Daily OCR cap (runaway-cost guard), state stored under $DSH_HOME |
cacheMaxEntries | 500 | OCR result cache size cap (LRU eviction) |
maxOcrTokens | 2048 | Vision call output cap |
stateFile | ~/vision-guard-state.json | Budget state file (~ = dsh home) |
ocrPrompt | verbatim transcription | Custom instruction |
passthrough | [] | Raw-image whitelist: [{provider, model}] — only add routes you have tested to accept images |
vision_analyze side: the OCR engine is a required per-call argument (engine), chosen by the model per task — local = local tesseract (free, characters only), vision = the configured vision model. There is no localOcr config key.