dsh-vision-helper

DeepSeek Harness agents are driven by plain text at their core, yet real workloads are full of screenshots, UI wireframes, and data charts that a text-only model simply cannot see. Users are forced to describe images by hand while the agent guesses from words alone — dsh-vision-helper exists to close that gap.

The plugin registers a single vision_analyze tool: it accepts a local file path or a data URI, forwards the image to a user-configured multimodal model (any OpenAI-compatible vision endpoint), and returns structured analysis text. The entire package is zero-dependency — no build step, no symlinks, no npm install — and once installed it becomes a persistent, session-wide tool across all profiles. A smart system-prompt injection handles the rest: in auto mode the agent checks whether the main model can already see images and skips the extra call when possible, only invoking vision_analyze when the main model is text-only.

If your workflow regularly requires the agent to genuinely read a screenshot or chart, or you have already wired up a multimodal route and want a dedicated vision model to handle recognition instead of letting the main model guess, this plugin is built for that use case. Every setting takes effect immediately, survives restarts, and demands no ongoing maintenance — ideal for DSH users who prefer a lean, drop-in utility.