test_moon_composite.py uses the existing repo sample frames
(21-07-00.jpg east frame with thin cloud cover + full_moon_closeup.jpg
as NASA render stand-in) to produce test_composite_out.jpg without
needing a live NASA API call or moon detection run.
Run: python3 test_moon_composite.py
https://claude.ai/code/session_01HkTxpNSTWtViZzxbrbKytR
Three changes driven by the same goal: make each monthly moon post
look like east captured it at that exact moment.
Tighter illumination windows (±4% of target phase):
Full moon: 95% → 96–100% (was a 5-point band; now hugs exact full)
Quarters: 40–65% → 46–54% (was a 25-point band; now ±4% of 50%)
This ensures the NASA Dial-a-Moon render timestamp is pulled within
4 percentage points of the true phase, making the fetched image
genuinely represent that night's moon. One fresh fetch per event,
~3/month, cached by hour — never reused across months.
Atmospheric cloud veil from east's surrounding sky:
_extract_cloud_veil() samples the annular sky region just outside
east's moon disk (2×–5× radius), scales it to the output frame,
blurs heavily (GaussianBlur r≈output_width/10) so it reads as haze
rather than an upscaled photo, then blends it over the NASA composite
at an opacity proportional to sky brightness:
sky < 5% mean brightness → no veil (clear dark night)
sky ~10% → ~17% veil (thin haze / airglow)
sky ≥ 20% → 40% veil (max, MOON_CLOUD_OVERLAY_MAX_OPACITY)
After the veil pass the NASA moon disk is re-pasted sharply so the
haze sits naturally behind the crisp lunar surface.
This is the honest answer to "east can't capture high-res clouds":
east's real atmospheric fingerprint becomes the veil texture.
New sky-cam.conf keys:
MOON_CLOUD_OVERLAY_ENABLED=true
MOON_CLOUD_OVERLAY_MAX_OPACITY=0.40
#MOON_CLOUD_OVERLAY_BLUR=0 (0 = auto)
https://claude.ai/code/session_01HkTxpNSTWtViZzxbrbKytR
The previous east-flavored composite (tight-crop east region + lunar
texture overlay) was honest but visually limited: 38-px source moon, no
real terminator shadows on quarters, and a faint upscaled-halo
background that was less compelling than just a clean lunar render.
New flow:
- East stays the witness (verifies the moon was visible during the
collection window) and supplies the timestamp.
- moon_dialamoon.py fetches the NASA SVS Dial-a-Moon render for that
exact UTC hour. Free, public-domain, real-physics: correct phase,
libration, and crater shadows for any timestamp.
- moon_composite.render_phase_closeup() places that render at ~92% of
frame height on a 1920x1080 black background -- the long-telephoto
look the user asked for.
- One API call per phase event (~36/year), cached forever in
moon-ref/dialamoon/.
Quarter moons now show real 3D crater shadows along the terminator,
which the prior algorithmic phase-shadow couldn't simulate from a
full-moon reference.
Adds MOON_REQUIRE_EAST_VERIFY=true|false toggle so the user can choose
"only post when east saw it" (default, current behavior) vs "post every
cycle regardless of weather over east."
Drops the install-time lunar reference download (no longer needed) and
the now-unused per-phase phase-shadow path stays in moon_composite.py
for reference / future reuse.
https://claude.ai/code/session_015PBVDESC3KLMbq1LpA6qLn
fade_dur = SUNRISE_TARGET_SECS * 0.085 = 0.850 at the default 10s target.
bc outputs .850 without a leading zero, which ffmpeg's afade filter rejects
with "Unable to parse option value .850 as duration". Pipe bc output through
sed to prepend the zero when the value starts with '.'.
https://claude.ai/code/session_01DANuzLCLhhQ7Li12RTcswY
daily_sunrise_video.sh:
- Insert continuous ambient recording extract as 2nd fallback (after today's
scheduled sunrise audio, before live RTSP). Finds the chunk file in
AUDIO_DIR/<cam>/<date>/ whose start time is the latest one at or before
sunrise, then ffmpeg-copies SUNRISE_TARGET_SECS centred on sunrise_sec.
- Renumber remaining fallbacks: live RTSP → 3rd, yesterday → 4th, library → 5th.
sunrise-audio-capture.sh:
- Fix: retention find pipeline could exit non-zero (directory not yet created
or permission error) and with set -euo pipefail cause the service to report
status=1/FAILURE even after a successful capture. Guard with [ -d ] and
add || true so a retention failure never masks a successful recording.
https://claude.ai/code/session_01DANuzLCLhhQ7Li12RTcswY
- daily_sunrise_video.sh: add live RTSP capture as final fallback in production
audio selection (fires when today, yesterday, and library all miss)
- On audio mix failure, retry once with a fresh live RTSP capture before
giving up — handles corrupt/incompatible pre-recorded files
- When mix still fails after retry, append last 30 lines of
sky-cam-audio-capture.service journal to the ntfy notification body
- Move sunrise-time text overlay from 5% to 15% above the bottom edge
https://claude.ai/code/session_01DANuzLCLhhQ7Li12RTcswY
Two new jobs operate on the SUNRISE_CAM, sharing three Python helpers
(moon_phase, moon_detect, moon_composite) and one cached lunar texture:
- moon-track.sh: nightly batch detects the moon in each east frame,
crops a 480x480 box around it, stitches into an mp4 that holds the
moon roughly centred while clouds and stars drift past.
- moon-phase-monthly.sh: daily check that runs whichever phase composite
is due that day. Handles full moon (posts D+3), first quarter (D+2,
best-effort due to daytime-only geometry from east), and third quarter
(D+2). Picks the frame closest in time to the exact phase moment that
meets quality / altitude / illumination thresholds, then composites
the cached lunar texture into it -- sky/halo/parallactic-angle/timing
real from east, surface detail borrowed from the reference image.
Honest-by-design: a 38-px white blob from a wide-field IP camera cannot
be enhanced into crater detail by software. The composite makes the
borrowing explicit and constrains everything else (when, where, sky,
orientation) to match what east actually saw.
install.sh now downloads the skyfield ephemeris (de421.bsp) and the
default lunar reference (Wikipedia CC BY-SA full-moon photo) on first
run. Both can be overridden via .env.
https://claude.ai/code/session_015PBVDESC3KLMbq1LpA6qLn
Drops clear/cloudy/fog/windy and the wind-speed override — those
conditions sound the same on the recordings, so they don't earn a tag.
Also renames "rainshower" → "rainstorm" and rebalances the rain
mapping so 503/504/522 (extreme + heavy showers) get the storm tag.
Final vocabulary: thunderstorm, rain, rainstorm, snow, snowstorm.
Anything else falls through to the un-tagged filename.
Adds weather-tag.sh, a small failure-safe helper that queries OpenWeather's
free "Current weather" endpoint and prints one tag from a fixed vocabulary
(thunderstorm, rain, rainshower, snow, snowstorm, fog, windy, cloudy, clear)
or nothing on any failure.
ambient-record.sh calls it just before each 30-minute chunk so the filename
reflects conditions at the moment of capture, e.g.
east-08-30-00-thunderstorm.m4a
east-12-00-00-clear.m4a
A foggy dawn no longer mislabels the sunny afternoon files.
Off by default. Enable with WEATHER_ENABLED=true and an
OPENWEATHER_API_KEY entry in .env; missing key / API error / disabled
toggle all fall through to the un-tagged filename, never blocking the
recorder.
The retention block ended with `[ test ] && echo`, which under `set -e`
returns exit 1 when `deleted` is 0 — the normal case on most days. As
the script's last command, that bash exit propagated to systemd, which
marked the unit failed and fired the OnFailure ntfy even though the
audio file was written successfully.
Add `|| true` and an explicit `exit 0` so a no-op retention pass is
treated as success.
Adds a quick-edit cheatsheet covering font, color, size, position, and
shadow/border for the sunrise-time overlay so the layout can be tweaked
without grokking ffmpeg's drawtext syntax.
Replaces the fixed 0.3s fade with SUNRISE_TARGET_SECS * 0.085, so a 10s
clip now fades 0.85s in and out. The fade scales automatically if the
target duration is tuned later.
- Shorten audio fade in/out to 0.3s so they fit comfortably in the 10s clip.
- Lift the sunrise-time overlay by ~5% of frame height to clear the
burnt-in timestamp on the source images.
- When today's sunrise recording is missing, fall back to yesterday's
clip before reaching for the library fallback.
- Print "Validating N frames..." before the loop, including active
filter thresholds when grey-checking is on
- Print "X / N frames checked..." every 500 frames so the user can
see the loop is alive during the long ImageMagick pass
- Print "Validation done: N valid frames" after the loop
- Note "(may take several minutes)" on the Step 1 encode echo
- Show the speed factor and target together: "28.8x → 11.08s target"
https://claude.ai/code/session_01U29A8NpgJ41tboiggZNmk8
When SEASONS_GREY_STDDEV_MIN is enabled, ImageMagick's convert outputs
the stddev/mean values without a trailing newline. read exits 1 on EOF-
without-newline even when it successfully populated the variables. With
set -euo pipefail this kills the script silently on the very first image
that enters the grey-filter code path — producing the observed symptom:
script stops after "Images found:" with no error and no video.
Fix: append || true so read's EOF exit code is absorbed. The variables
are still set correctly; only the exit code changes.
https://claude.ai/code/session_01U29A8NpgJ41tboiggZNmk8
When no corrupt or grey frames exist (bad=0, grey=0), the two
conditional-echo lines
[ "$bad" -gt 0 ] && echo "WARNING: ..."
[ "$grey" -gt 0 ] && echo "INFO: ..."
exit with code 1 because [ 0 -gt 0 ] is false and && short-circuits.
With set -euo pipefail the script treats that 1 as an error and exits
immediately with no output — producing exactly the "ran without making
a video, no error, no notification" symptom on any normal day where all
frames are valid.
Fix: append || true to each line so the expression always exits 0.
https://claude.ai/code/session_01U29A8NpgJ41tboiggZNmk8
Without explicit duration in the concat file, ffmpeg assigns timestamps
from EXIF metadata or defaults — producing a variable frame-rate raw
video with uneven frame spacing. Applying setpts without fps= then
preserved that unevenness into the final clip, causing the observed
pause-then-jump-minutes/hours behaviour.
Fix 1: write `duration 0.04` (1/25 s) for every JPEG in the concat
list, making the raw video a true 25 fps CFR stream. The duplicate
final-entry is appended as required by the ffmpeg concat demuxer so the
last content frame keeps its full display duration.
Fix 2: add `fps=25` to the setpts filter in step 2, resampling to
standard 25 fps CFR regardless of the speed factor, which guarantees
smooth, uniform playback in the final clip.
https://claude.ai/code/session_01U29A8NpgJ41tboiggZNmk8
Conditional expressions like `[ "$count" -gt 0 ] && echo ...` return 1
when count is 0 (nothing to delete). Under set -euo pipefail with no
explicit exit 0, that false becomes the script's exit code. systemd sees
failure, OnSuccess= for sunrise2mm.py never fires, and the Mattermost
upload is silently skipped every day.
Fix: add || true to all end-of-script conditional-echo statements and to
the cam_audio cleanup line, then add an explicit exit 0 so the success
path always exits cleanly.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
File size can't reliably reject bad frames (200KB can be good or bad).
Remove SEASONS_MIN_FRAME_BYTES. Instead, the grey/green uniform-colour
filter (SEASONS_GREY_STDDEV_MIN) now skips the expensive ImageMagick
check for files above SEASONS_LARGE_FRAME_BYTES (default 1 MB) since
those are always good. Only sub-1MB files get the content check, making
the filter practical on days with 40k+ images.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
43k images × ffprobe process launch = 30-90 min before encoding starts.
Every corrupt file seen in practice has been 0-byte (caught instantly by
[ ! -s ]). Non-zero corrupt files are rare; if ffmpeg hits one the
existing drift warning flags the short output. Removing ffprobe restores
the fast path: one stat() call per image.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
Covers: installing git, first clone, pulling updates, stash vs reset
--hard, checking diffs, switching branches, and a note that .env is
never touched by git pulls.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
Allows processing any past date instead of always yesterday:
./4-seasons.sh east 2026-04-19
Useful for backfilling days with existing images.
Systemd (no date arg) continues to default to yesterday.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
duration=INTERVAL in the concat file made a 0.1fps raw video, which
x264 couldn't speed up 4000x cleanly and produced 20s of the last
frame instead. The original script used plain 'file path' entries
with no duration, letting ffmpeg default to 25fps internally. The
speed_factor calculation against raw_duration compensates correctly
either way, but only the standard-fps raw video survives the extreme
setpts transform. Also removed the coverage_pct check which was
based on images*INTERVAL and became meaningless without duration.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
Accessing ${images[-1]} on an empty array under set -u causes a bash
error. Guard with an explicit check so a fully-corrupt image directory
exits with a clear message instead of a cryptic crash.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
fps=25 caused ffmpeg to discard 15 of every 16 source frames, making
the final clip jump by ~2.5 minutes per frame. The original script
had no fps filter, so all source frames were preserved. Without it,
the setpts speed-adjust keeps every frame; the effective framerate in
the output rises proportionally (e.g. ~400fps for a full day at 10s
intervals into an 11s clip), which is fine for H.264 and all
downstream montage steps.
https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ
Claude