dsh-mimo-vision-hint

When an agent meets an image mid-workflow, it usually falls back to read_image and parses the pixels inside its own context window — burning main-session tokens and leaving a dedicated vision model on the bench. dsh-mimo-vision-hint fixes this with the lightest possible approach: it registers no tools, exposes no services, and simply injects a short instruction into the system prompt that tells the agent to dispatch image-recognition tasks to an opencode-go mimo-v2.5 subagent via the workflow tool. If the subagent call fails, the agent gracefully falls back to read_image, so the pipeline never stalls.

Because the plugin is prompt-only, it needs no isolate realm, no runtime dependencies, and no extra configuration. Once mounted into an agent preset, the instruction takes effect for every new session immediately. The prerequisites are minimal — a working opencode-go / mimo-v2.5 provider-model pair and a workflow tool in the active preset — and if either is missing the injected text simply steers the model straight to read_image with no side effects.

It is designed for DeepSeek Harness deployments that already serve mimo-v2.5 and want to offload vision tasks from the main conversation without adding services, schemas, or extra moving parts. If you only glance at the occasional screenshot, you can skip it; if you want every image to default to a dedicated vision subagent, this one-line prompt plugin does the job with zero overhead.