What it can see
Um Bot receives a compact text summary of the open project and sequence: project and sequence names, clip names and timings, a timestamped transcript when one has already been produced, and findings from the latest Um Out pass. It answers about that context, so suggestions can point at timecodes instead of generic editing principles.
The model does not receive media files or local media paths. If the answer is not in the context it can see, it should say so rather than invent a clip, line, or timecode.