Flow

Run flow and deliver the requested media. Do not stop after writing a prompt unless the user explicitly asks for prompt text only. In a video project, video-production decides who runs what: the agent runs stills and 360p draft batches through make.py and reviews them; the user runs only the 1080p finals, from the command the agent gives.

When a job is stuck or flow is broken

There is no cancel command. Don't kill and relaunch in a loop.
NOT_FOUND in the final summary: flow has already spent its one regeneration on another account, so rerunning unchanged won't help. Put a new "seed" on that job and rerun. Clips Flow's content filter fails after accepting them are no longer reported as lost: they come back as UNSAFE_GENERATION with Google's reason (see below).
BATCH_WORKER_FAILED: rerun the same command.
No progress for about 10 minutes, or the same failure twice after following the hint: stop. Tell the user at once, in a few lines they can hand to whoever fixes flow: the exact command, the last lines of output, the error code, and what looks wrong (a stuck profile lock, an expired login, a Google error). Don't keep waiting, work around it or offer a menu of options.

Waiting on a run

Run the flow (or make.py) command itself as the background task; the harness tells you when it exits. That is the only wait you need.
Never poll with pgrep -f/pkill -f on a pattern: the waiting shell's own command line contains the pattern, so the loop never ends (video-3: two loops ran 2.5 hours, looking like a flow hang) and pkill -f kills the shell that started it. To wait on or stop a process, use its PID (tail --pid=<PID> -f /dev/null, kill <PID>).
Before reporting done, check your background tasks and stop any of yours still running.
For a quick look at what flow is doing (batch saved/total, jobs in flight, last error, free and premium credits as last seen), run flow summary. It reads local files only and never spends credits.
Write each scene's stills or clips into a folder named for it (-o scene-3/scene-3-v2.jpg). The user's status bar groups jobs by folder, showing what is done and what is generating, so a folder per scene keeps that readable.
flow stop stops every running flow command (not only yours) and closes Labflow's browser. Use it only when the user asks to stop everything; to stop your own run, kill <PID>. Accepted jobs resume on rerun.