image_inspect supplies visual understanding to a main model that is not
vision-compatible. The tool accepts an original-image reference and one
question, then asks the session-selected visual_image model in an isolated
completion. It never accepts a provider/model override and never takes over the
main task. The result states the original image dimensions and a no-crop basis;
the vision model is asked for normalized coordinates relative to that original
so they map back regardless of provider-side resizing. Compaction summaries
carry image_ref tokens forward, so the reference remains usable after
compaction.
The host hides the tool when the main request is vision-compatible or no vision
target is selected.