dsh-image-pathify

A DeepSeek Harness plugin that lets models which cannot view images (for example deepseek-v4-flash) handle images pasted into chat: before a message is sent to such a model, the image is replaced with a local file path, and the built-in analyze_image tool reads that file and returns a text description through a configured vision API. Requires DeepSeek Harness >= 0.1.1-rc.1 (readme states it was verified on 0.1.5-rc.1); the readme says earlier hosts should keep using dsh-image-pathify@0.1.8. Lice

• Before sending to a model that cannot view images, images are replaced with a local file path line; chat records and UI thumbnails stay unchanged.
• Adds a built-in analyze_image tool: the model reads the saved file and gets a text description from the configured vision API.
• Models that can already view images are unaffected: images are sent as-is and analyze_image is not in their tool list or system prompt.
• read_image is rejected on models that cannot view images, with a hint to use analyze_image instead.
• Multiple images can be passed in one analyze_image call via images, so all are included in a single vision request.
• Settings page (Settings → Plugins → Image Recognition) covers API key, vision model, vision base URL, disableThinking, maxTokens, models and relaxAdmission; sav
• API key is written to $DSH_HOME/.credentials.yaml; other image-pathify fields go to $DSH_HOME/settings.yaml under the image-pathify section.
• Works with any OpenAI-compatible vision endpoint; when the installed version lags npm, the card header shows the new version and copies an upgrade command.