dsh-read-image-jpeg-fallback

In the DSH + LM Studio openai-completions path, the read_image tool call can succeed, yet the very next provider request build chokes because LM Studio rejects the PNG or WebP image attachment. dsh-read-image-jpeg-fallback is a zero-configuration compatibility shim that fixes exactly that transport mismatch. It does not replace read_image, does not add an OCR step, and does not route the image through a separate vision model. Instead, it transparently re-encodes PNG/WebP blocks to opaque sRGB JPEG (quality 90) via sharp just before the outgoing request is assembled, so the main multimodal model still receives the image itself.

The plugin hooks into DSH's tools/post-execute event. For accepted read_image results containing PNG or WebP image blocks it reads the stored attachment, converts it locally, writes the new JPEG through DSH's attachment store, and swaps only the model-visible image block reference. The canonical tool result and the built-in read_image implementation remain untouched. The conversion is fail-open: if reading, converting, or saving goes wrong, the original result is preserved, a warning lands in the DSH log, and a successful tool call is never turned into a plugin error.

This is the right fit when your main model already natively accepts image input, you talk to LM Studio through its OpenAI-compatible route, and the only failure point is the media format at request-build time. It is not the right fit if your provider already digest the formats DSH emits, or if your main model is text-only and what you actually need is an OCR or vision-fallback pipeline.