DSH Plugins
返回列表
🤖

dsh-image-router

模型推理 更新于 2026.09.13

在终端中运行以下命令:

dsh plugin install zhiwuli0228/dsh-image-router

将以下提示词粘贴到 DeepSeek Harness 对话框中:

在 DeepSeek Harness 终端中执行 dsh plugin install zhiwuli0228/dsh-image-router 即可安装,源码仓库为 https://github.com/zhiwuli0228/dsh-image-router

插件介绍

Text-only models in DSH hit a dead end the moment an image enters the conversation. The usual workaround is to swap to a multimodal model mid-session, which means a jumpy model selector, a [model changed] banner, and a session history that looks like two different models are taking turns talking. dsh-image-router sidesteps all of that: before you hit send, a vision model you designate quietly turns the image into a paragraph of text and substitutes it into the prompt. Your conversation model never changes, the selector stays put, and the log records a text analysis instead of an image block. From the model's point of view it simply received a longer piece of text.

Feeding the vision model is flexible. If you already have a multimodal model configured under Settings and Models, pick it from a dropdown that only lists image-capable entries. If you have a separate vision API, fill in base URL, model name, and API Key, and the plugin translates that into a single upstream route without touching any of your existing provider definitions. Beyond inline images, any model can call the describe_image tool to analyze an image file on disk on demand, because the image never enters the session and thus never hits the route-must-declare-image gate. API Keys are written to a credential store once and stripped from every parsed result, config section, and card draft; a failed vision call simply lets the original request proceed so the host surfaces the real error rather than masking it.

Who is it for? DSH users whose primary model is text-only but who regularly deal with screenshots, annotated diagrams, or document images. Developers who want a single, consistent model context across a session and refuse to be interrupted by mid-conversation model swaps. And anyone with a self-hosted or third-party vision endpoint who wants to plug it into their DSH workflow without reconfiguring existing providers. Everything is set through a GUI card, takes effect on the very next prompt, and requires neither hand-written YAML nor a restart.

使用场景

  • 纯文本模型需要处理截图或图表时,自动旁路调用视觉模型生成文字描述替代图片块
  • 已有独立视觉 API 端点,想无缝接入 DSH 会话而不改动现有 provider 配置
  • 模型在会话中按需读取磁盘图片文件,不受路由必须声明图片模态的限制

适合人员

  • 主力模型为纯文本、但日常涉及截图或文档图片的 DSH 用户
  • 要求会话模型上下文始终一致、拒绝中途切换的开发者
  • 持有自托管或第三方视觉端点、希望低成本接入 DSH 工作流的用户