dsh-vision-guard

Workhorse models like deepseek-v4-pro on DeepSeek Harness are text-only. Paste a screenshot and the model simply cannot see it; worse, once an image block enters the session log every subsequent turn replays that image to an upstream gateway that only accepts text, producing a 400 unknown-variant error that bricks the conversation permanently. dsh-vision-guard is a transparent guardrail plugin designed to eliminate both failure modes at the root.

Its core design is a two-stage pre-log rewrite: image content is converted into model-readable text at the agent/pre-step stage, so it never touches the session log and the deadlock path is severed at the source. Sessions already deadlocked before installation recover on the next turn because replayed history images are rewritten at request time. The bundled vision_analyze tool lets the model pick an engine per task - local character OCR or a configured vision model - covering image reading, PDF text layers with embedded images, docx/pptx parsing, and video frame OCR. Anti-deadlock is a hard invariant: even if the entire vision pipeline is unavailable, output degrades to placeholder text and never falls back into the 400 loop.

It is built for users whose primary model is a text-only workhorse who still need to drop screenshots, PDFs, and office documents into the workflow daily. The plugin carries zero external dependencies, runs on pure Node builtins, and holds no API keys of its own - it reuses the vision-model route and credentials already configured in your dsh profile. All you need is one image-capable model on your subscription to act as the eyes while your reasoner keeps doing the thinking.