libinyam/dsh-vision-provider

dsh-vision-provider

English | 简体中文

dsh-vision-provider gives DeepSeek Harness selectable vision choices under one DeepSeek + Vision provider:

DeepSeek + Vision
  GLM-4.6V-Flash
  Qwen VL Max
  GPT-4.1 mini (Vision)

Select only one combination in Harness. The vision model named in that selection is used behind DeepSeek:

Text-only message ───────────────────────────────> DeepSeek V4 Flash

Image message ──> private vision sidecar ──> visual description
                                               │
                                               └──> DeepSeek V4 Flash ──> answer

The vision model does not run as the final answer model. Instead, it appears as part of a selectable DeepSeek combination. DeepSeek still performs reasoning, tool use, and final response generation.

This is a community project. It is not an official DeepSeek or OpenAI package.

Why v0.3.0 exists

Version 0.1.0 added a standalone model named vision-openai. DeepSeek Harness can select only one model for a session, so users had to choose either DeepSeek or the vision model. The two models could not cooperate.

Version 0.2.0 introduced a runtime composite adapter, but the vision model remained hidden in environment configuration and Web UI showed only the vague label DeepSeek V4 Flash + Vision.

Version 0.3.0 brings vision selection into Web UI:

• the plugin reads every model in Settings > Models that advertises image input;
• each vision model becomes a separate selectable DeepSeek combination;
• the combination name shows the vision display name, while its description starts with the exact model ID and provider route;
• text-only requests go directly to deepseek-official/deepseek-v4-flash;
• image-bearing messages are analyzed by the vision model selected in Web UI;
• the visual analysis replaces the raw image before the request reaches DeepSeek;
• DeepSeek remains the model that reasons, uses tools, and writes the final answer;
• repeated tool steps reuse cached image analysis in the current process.

This is a two-model bridge, not native pixel input for DeepSeek. The quality of the final answer depends on both the vision sidecar and DeepSeek.