dsh-image-bridge

Pure-text LLMs like DeepSeek have a blind spot they cannot see images. Drop a screenshot, architecture diagram, or handwritten formula into the Web GUI and the model simply has nothing to say about the binary blob. dsh-image-bridge plugs that gap: it hooks into the agent/pre-step waterfall event, intercepts image blocks before the model request is assembled, calls a free vision model (GLM-4V-Flash by default) to produce a structured text description, then injects it back into the context so the conversation continues seamlessly.

The core design prioritises resilience. A single model hit with 429 or 503 gets exponential-backoff retries; 404 or exhausted retries trigger a fallback chain (glm-4.6v-flash, glm-4.1v-thinking-flash, glm-4v-flash); if every model fails, the image block is replaced with a failure-notice text and the turn still completes instead of crashing on UNSUPPORTED_CONTENT. Descriptions are cached by attachmentId so the same image is not re-recognised in later steps; images exceeding 10 MB or 40 MP receive an immediate failure notice rather than entering the retry chain; and images nested inside tool-result payloads are transcribed recursively.

Built for users running DeepSeek or other text-only models who still want to understand images in conversation. Zero build step, zero third-party dependencies, Node built-in modules only, and a single API-key environment variable to get started.