kajisho5/ffmpeg-skill v3.8.8

on GitHub

Automated release: version and notes generated from pull requests merged since 3.8.7.

Fixed

• Cache keys see every file an output depends on (#458 A80, A81, A100, A277). batch.py's
  and render.py --cache's file key read the modification time in whole seconds, so a same-size
  rewrite of a file's middle within one second was served the old output; it is nanoseconds now. A
  rewrite whose mtime is set back to the old value, or a filesystem that keeps whole seconds, still
  keeps the key (no full hash: --force re-runs). A project recipe's key now covers every file the
  project names (cues, music, brand and its logo and font, overlays, chapters, the other clips), and
  is computed on every --watch pass instead of once. render.py --cache serves the chapters
  stage of an inline chapters list when the delivery under it came from the cache; a delivery copied
  without an export stage is a new file and still misses. Every cache key changes once with this
  release, so the first run after it runs everything again.
• render.py's work directory belongs to one run (A77, A98, A99). The default directory is new
  for every run (<output>_work_<pid>_<random>), so a later run with the same PID no longer deletes
  an earlier run's kept preview. A re-run with the same --work replaces the stage files an earlier
  render wrote there and nobody changed since (recorded in .ffskill-render-stages.json) instead of
  refusing them; any other existing file there is still refused, with --cache too (a cache hit
  no longer replaces it: the stage runs and refuses it, cache.miss_reasons says
  destination_not_owned). A --dry-run writes nothing into
  --work and removes a --work directory it had to create.
• An output is named for what is written into it (A83, A84, A108). batch.py names a .webm
  source's output .mp4 (unless the recipe sets ext), where every WebM item was refused
  output_container_codecs. render.py with no export stage refuses an output named for another
  container than the file it would copy (.webm, .mkv for an MP4 intermediate), kind: input,
  before the first stage when it can tell, at the copy otherwise; it used to complete and verify an
  MP4 named .webm. .mov, .m4v, .m4a and .3gp are accepted for an MP4. sequence.py --dir .
  without -o writes <folder>_sequence.mp4 beside the folder instead of a traceback.
• Caller text no longer adds lines to a file render.py writes (A260, A261, A263, A264). An
  inline chapter title holding a line feed, carriage return or NUL is refused ("Intro\n0:01 Evil"
  was two chapters, verified), and metadata.py refuses a NUL in a --chapters file; other control
  characters are kept. An EDL replaces line breaks and control characters in every name with a
  space, and FCPXML the characters XML cannot hold. The social pack's table escapes | in a file
  name and report.py --pack reads it back (a | shifted every column and marked passing
  destinations failed). A JSON file with a lone surrogate escape ("\ud83d") is refused as invalid
  in a project, recipe, brand, plan, --words or transcript file, instead of a UnicodeEncodeError traceback with nothing on stdout;
  surrogates in command-line arguments are not covered.
• An EDL refuses what CMX 3600 cannot number (A265, A266): more than 999 events, or a record
  past 24:00:00:00, is refused naming .fcpxml/.otio; a source timecode wraps at midnight
  (23:59:59:00 plus 2 s is 00:00:01:00, not 24:00:01:00).
• render.py and batch.py do what their flags say (A82, A87, A90, A96, A102, A267).
  --stop-after export stops after the delivery (the chapters and check stages ran, and a non-object
  "check" was a traceback). An executed plan reports its output as an absolute path, not relative
  to the plan's own directory. batch.py's glob matches names without regard to case on every
  platform (*.mp4 found nothing of GOPR0001.MP4 on Linux), letters inside a [...] class
  excepted, and a glob that matches nothing is a note. --export-timeline probes each file once.
  --list-templates and --init print a JSON document under --json. The audiogram template takes
  --brand and draws --title as a lower third. references/scripts.md now says that a pack is
  written into the directory of -o, lists the ten presets that turn normalize on, and no longer
  names a --codec render.py does not have; examples/brand.json lost a loudness block no tool
  reads. Measured on FFmpeg 6.1.1 only.
• fix: the remaining findings of the render.py, batch.py and sequence.py audit (#458, 25 items) (#508)

What's Changed

• fix: the remaining findings of the render.py, batch.py and sequence.py audit (#458, 25 items) by @kajisho5 in #508

Full Changelog: v3.8.7...v3.8.8