dsh-vision-helper
Deployment-level vision plugin for DeepSeek Harness: registers the vision_analyze tool backed by a configurable multimodal model, with a settings-page UI. Zero dependencies.
AI 분석
核心用途是为纯文本主模型补充视觉分析能力。适合需要处理图像分析任务的用户。支持在设置页面配置服务商、模型、温度及最大图片尺寸，零外部依赖，开箱即用。
설치
검증된 bundle이 없거나 호환성 검사에 실패했습니다. 먼저 저장소 설명을 읽어 주세요. 전체 README 읽기 ↗
Configuration
Via the settings page (设置 → 视觉助手): provider, model, temperature, max output tokens, max image edge. Equivalent config document at $DSH_HOME/dsh-vision-helper.json (beside settings.yaml):
{
"provider": "opencode-go",
"model": "mimo-v2.5",
"temperature": 0.2,
"maxTokens": 1024,
"maxEdge": 4096,
"mode": "auto"
}
provider/model empty = auto-select; maxEdge limits the longest image edge in px. mode controls the automatic-invocation guidance injected into the agent's system prompt:
auto(default) — smart guidance: if the main model is multimodal (image delivered in conversation, or readable directly viaread_image), analyze it natively with no extra cost; callvision_analyzeonly when the main model cannot view images (text-only model); ask for the path when missing.force— mandatory: image tasks MUST use the tool, no direct answering or refusal. Use this only when you explicitly want the dedicated vision model (e.g. a text-only main model, or a specialized vision model) — with a multimodal main model it incurs extra API calls.off— no guidance injected; the tool stays registered but the agent decides on its own.
Changes are read per use — no restart needed (the guidance section re-registers on save).