dsh-vision-router Visual Memory: Eyes for Text-Only DeepSeek Harness Agents
dsh-vision-router is a DeepSeek Harness (DSH) plugin that gives text-only agents eyes by routing image turns to a separate vision model while DeepSeek stays the reasoning brain. It installs with one command, needs no Python and no API key, and ships 14 deep vision tools plus cached visual memory so a text-only agent genuinely remembers earlier images without re-spending vision calls. What is dsh-vision-router and why text-only DeepSeek agents need eyes DeepSeek Harness is a local agent framework built around DeepSeek models. Many of its most popular models, including deepseek-v4-flash, are text-only: they declare inputModalities: ["text"] and simply cannot accept image attachments. That is a real limitation for agent work, because a growing share of tasks — reading a screenshot, checking a rendered UI, verifying a chart, OCRing a receipt — are fundamentally visual. ...