Image observations

MODEL "Image observations"
DESCRIPTION "Analyze a supplied image observation with an installed image-capable model package. Requires the referenced model release and an input observation."
VERSION "1.0.0"
TAGS "ai", "images", "observations"

# Supply a grid-ai-input/1 object through the existing input API.
# The producer chooses snapshot, historical, or realtime timing and retains
# source assets when later replay is required. No capture starts on load; supply the input before resolving the output.
# The same input accepts selected camera/window frames and video clip offsets.
# mediaTimeMs locates a frame in a recording; it is not capturedAtMs event time.
# Chat can gather a bounded time range from a local video. Those selected frames
# use the same observation contract without inventing a recording date.
# Select AiObservation in Data, then Add source → Observe selected input to
# follow an already-running image producer with fresh, distinct captures.
# This action starts a read-only Agent session; it does not start the producer.
input AiObservation = BLANK

# Install this exact image-capable release after it is published.
# The same signed AI capability serves Agent and workbook requests.
output ImageAnalysis = PREDICT("model:grid/qwen3-vl-4b-instruct@1.0.0", AiObservation)

END MODEL