Starter time-lapse video
This skill turns the starter photos since a feed into one looping video on the Pixel, playing in VLC without the screen shifting between passes.
Read the zephyr skill first for the key scheme, the bake log filenames, and the event_time tags.
Read the starter-peak-from-photos skill for where the photos live and how they are named, because this skill reuses its list_frames.py.
Inputs
zephyr is required and has no default.
It accepts only the Zephyr key form, such as 1 or 3-2.
hours defaults to the whole time since the feed.
When given, such as "the last 3 hours", keep only the photos within that many hours of the newest photo.
direction defaults to forward.
reverse plays newest first, as in "play it backwards".
probe defaults to 2.
The feed time is never typed by default.
It is the event_time of the feed starter entry in the bake log, resolved through the zephyr skill.
If no feed entry has a time, stop and ask for it.
Before starting
Docker must be running, because ffmpeg runs in a container and is never installed on the host.
docker info --format {{.ServerVersion}}
If that fails, run open -a Docker, then check again.
Use the session scratchpad for every temporary file, or mktemp -d when there is none.
SCRATCH below means that directory.
Collect the photos and temperatures
Always fetch fresh, because new photos and probe readings keep arriving while the starter is still going.
FEED is the feed time as YYYY-MM-DDTHH:MM:SS.
uv run --no-active --script /Users/mtm/pdev/taylormonacelli/volcanicviper/skills/starter-peak-from-photos/list_frames.py --start FEED > SCRATCH/frames-list.txt
Write the probe readings into the scratch directory, never over chargingcheetah's own merged_data.parquet.
uv run --no-active --directory /Users/mtm/pdev/taylormonacelli/chargingcheetah main.py parquet --start FEED --probe 2 --output SCRATCH/temps.parquet
The probes are the diminutivedragon hardware, and the parquet holds one reading about every 30 seconds in UTC.
Annotate the frames
Clear the frame directory first with find SCRATCH/frames -mindepth 1 -delete when it already exists.
uv run --no-active --script /Users/mtm/pdev/taylormonacelli/volcanicviper/skills/starter-timelapse-video/annotate_frames.py SCRATCH/frames-list.txt --feed FEED --temps SCRATCH/temps.parquet --out-dir SCRATCH/frames
Add --hours N when hours was given.
The script converts the readings to America/Los_Angeles and writes frame_00000.jpg onwards, oldest first.
Every frame carries these properties, which the user settled on and should not drift:
- A black banner across the top.
- A large line such as
9h 34m since feed, or26m since feedunder an hour, in the colour of the current hour. - A smaller line such as
hour 10 now 4:21 AM fed 6:47 PM. - A third line such as
avg 69°F since feed, the mean of every probe reading from the feed up to that photo's time, rounded to the nearest whole degree with no decimals. - A 16px border around the whole frame, not around the jar, whose colour steps through a 16-colour palette once per elapsed hour since the feed, so it changes at the feed minute past each hour rather than on the clock hour.
- The camera's own date stamp stays in the bottom right corner.
Encode one pass
The pass is 15 frames per second, so each photo, taken about 2 minutes apart, lasts one frame.
docker run --pull never --rm --entrypoint /usr/local/bin/ffmpeg --volume SCRATCH:/work lscr.io/linuxserver/ffmpeg:latest -loglevel error -y -framerate 15 -i /work/frames/frame_%05d.jpg -c:v libx264 -preset fast -crf 28 -pix_fmt yuv420p /work/pass.mp4
For reverse, add -vf reverse before -c:v.
Do not add a held last frame, because the user found the pause confusing.
Do not add chapters unless asked.
Repeat the pass inside one file
VLC's own repeat reloads the file at every end, and each reload drops out of full screen so the status and navigation bars flash and the picture shifts.
Repeating the pass inside the file makes each restart just the next frame, so VLC only reloads once per file.
Choose the repeat count so the file runs about 10 minutes: count = ⌈600 / pass seconds⌉, where pass seconds is the frame count divided by 15.
Read the frame count from the pass rather than counting files by hand.
docker run --pull never --rm --entrypoint /usr/local/bin/ffprobe --volume SCRATCH:/work lscr.io/linuxserver/ffmpeg:latest -loglevel error -count_frames -show_entries stream=nb_read_frames -of csv=p=0 /work/pass.mp4
For example, 484 frames is a 32s pass and ⌈600 / 32.27⌉ = 19 repeats, and 91 frames is a 6s pass and ⌈600 / 6.07⌉ = 99 repeats.
-stream_loop takes the number of extra plays, which is the count minus one.
docker run --pull never --rm --entrypoint /usr/local/bin/ffmpeg --volume SCRATCH:/work lscr.io/linuxserver/ffmpeg:latest -loglevel error -y -stream_loop LOOPS -i /work/pass.mp4 -c copy /work/loop.mp4
The copy is not re-encoded, so the file grows in step with the count, at roughly 2MB per pass for a full night and 0.6MB per pass for 3 hours.
Check before sending
Cut the first frame, the last frame of the first pass, and the first frame of the second pass, cropped to the banner, and read the image.
docker run --pull never --rm --entrypoint /usr/local/bin/ffmpeg --volume SCRATCH:/work lscr.io/linuxserver/ffmpeg:latest -loglevel error -y -i /work/loop.mp4 -vf "select='eq(n,0)+eq(n,LAST)+eq(n,FIRST2)',crop=720:210:0:0,tile=1x3" -frames:v 1 /work/check.png
LAST is the frame count minus one and FIRST2 is the frame count.
The banners should run from the first window time to the newest photo time and straight back, with the average line present.
Send it to the Pixel
The file name is zephyr<KEY>-starter-loop.mp4, with -<N>h before -loop when hours was given and -reversed before -loop for reverse, such as zephyr1-starter-3h-loop.mp4.
Copy it into the Syncthing drop folder.
cp SCRATCH/loop.mp4 /Users/mtm/pixel-sync/NAME
Read its size with ls -l, then wait in the background until the phone's copy has the same size.
until /Users/mtm/Library/Android/sdk/platform-tools/adb -s 192.168.1.223:45147 shell ls -l /storage/emulated/0/Sync/NAME 2>/dev/null | grep -q SIZE; do sleep 5; done
Get the adb serial from adb devices first, because the IP and port change, and see the pixel9a-wireless-adb memory when the phone is offline.
VLC has permission to read videos only, not all files, so it cannot open the Sync folder.
Copy the file into Movies and register it with the media library.
/Users/mtm/Library/Android/sdk/platform-tools/adb -s SERIAL shell cp /storage/emulated/0/Sync/NAME /storage/emulated/0/Movies/NAME
/Users/mtm/Library/Android/sdk/platform-tools/adb -s SERIAL shell content call --uri content://media/external/file --method scan_file --arg /storage/emulated/0/Movies/NAME
Play it in VLC
Start VLC fresh at the beginning, because VLC otherwise resumes where the last video stopped or hands the file to whatever VLC screen is open.
/Users/mtm/Library/Android/sdk/platform-tools/adb -s SERIAL shell "am start -S -a android.intent.action.VIEW -d file:///storage/emulated/0/Movies/NAME -t video/mp4 -p org.videolan.vlc --el position 0 >/dev/null; sleep 3; dumpsys window | grep mCurrentFocus; screencap -p /sdcard/Download/claude-check.png"
Pull the screenshot, read it, and confirm the video fills the screen in portrait with the banner readable, then delete the screenshot from the phone.
Never launch apps with monkey, because it turns the phone's auto-rotate back on.
VLC settings this relies on
Google Photos and Files by Google both loop videos with no way to stop it, which is why VLC is the player.
VLC's "Video screen orientation" must be Portrait, under More, Settings, Video, because the default "Automatic (sensor)" ignores the phone's rotation lock.
The player's repeat mode should be repeat one, set in the player's "…" menu under "Repeat mode" until its icon shows a 1, so the file also restarts after its own 10 minutes.
VLC needs only android.permission.READ_MEDIA_VIDEO, granted with pm grant org.videolan.vlc android.permission.READ_MEDIA_VIDEO, and never all-files access.
If the picture is squashed or stretched and choosing "Best fit" by long-pressing the aspect ratio button does not fix it, pm clear org.videolan.vlc resets VLC, after which the video permission and Portrait orientation must be set again.
Check a setting only when something looks wrong, not on every run.
Driving the phone
Check the focused window inside the same adb shell command as each tap, and stop when it is not VLC, because the user often switches back to the Claude app and a blind tap lands there.
Never run a tap in parallel with the focus check that is meant to guard it.
uiautomator dump fails on VLC's main screen, which never goes idle while it looks for network shares, so read that screen from a screenshot instead.
VLC does not redraw while paused, so read a paused position from dumpsys media_session rather than from the picture.
When done, post the Ping notification described in the return-to-claude-app memory.
Output
Report in a few lines what was made and verified: the window as clock times, the photo count, the pass length, the repeat count and total length, the file size, the file name, and that it is playing in VLC.
Say plainly when a step was not verified, such as a screenshot that was not taken.
Errors
If Docker will not start, the photo folder has no frames since the feed, the parquet is empty, or the phone is not reachable over adb, stop and say which, with the command and its output.
Do not install ffmpeg, adb, or anything else on the host to work around a failure.