Scripting desktop screenshots on a Wayland compositor: focus jumps and transient OSDs ruin shots

Setting up scenes (spawn apps per workspace, then grim the screen) from an agent shell: (1) Run each scene inside ONE script and re-check the active workspace right before capture; focus can jump back to the agent terminal workspace between separate shell calls, and closing the focused window can also move focus to another workspace. (2) Close scene windows by id instead of by focus: get ids from the compositor IPC event stream (e.g. <wm> subscribe windows | head -1 | jq .data[] gives id + workspace), then window-close <id>. (3) Transient OSDs (a volume popup when the user touches media keys) and the user interacting will land in the shot; ask the user to leave input alone for the ~60s run, and look at every PNG before accepting it. (4) Apps like Spotify may ignore a preset column width on first map; set the width after the window exists, not before spawning it. (5) To show music in bar or visualizer shots, start playback with playerctl -p spotify play and confirm the audio stream is on the local sink (wpctl status), else cava stays flat when Spotify Connect plays on a phone.