kajisho5/ffmpeg-skill v3.8.14

on GitHub

Automated release: version and notes generated from pull requests merged since 3.8.13.

Fixed

• verify.py never writes over another of its inputs (review of #521). The names it
  generates (<stem>_cut.mp4, _acc, _fit, _cap with its .srt/.ass, _cues.txt, ...)
  were checked by each step against its own input only, so verify.py FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 in it replaced clip_cut.mp4 with the cut of
  clip.mp4 and then verified the replacement. Every name is checked against every input, a link
  or another case of the name included, before anything is written: kind: input,
  error.reason: output_is_input, naming both paths. batch.py already refused this (#428).
• --json-brief alone prints the JSON document on every tool (review of #521). It implies
  --json, but the tools that branch on --json themselves did not see it: probe.py --field
  printed a bare string or an empty line, check.py its text table, report.py the document and
  then a bare path, verify.py its Markdown report and no document, and cropdetect.py and
  silence.py --list a bare document with no status.
• redact.py --blur-strength 1 blurs colour too (review of #521). The chroma radius was half
  the luma radius rounded down, 0 at strength 1, so detail carried by colour alone stayed
  readable; it is at least 1 wherever the chroma plane can hold it (a region 6 px or more on its
  short side).
• The agent eval notices a second edit and refuses a run folder it would see (review of
  #521). evals/run_agents_25.sh compared git status alone, so a run that edited a tracked file
  already modified, or an untracked file already there, went unreported; the checkout's state now
  adds the tracked files' diff against HEAD and the untracked files' hashes. A RUN_DIR inside
  the checkout that git does not ignore is refused (exit 2) before anything is staged: the run's
  own transcript read as a change to the checkout.
  The state also carries HEAD and the index, so a run that commits or only stages its edit is
  seen; an untracked link is its target's name (a dangling one no longer fails the snapshot); and
  a snapshot that cannot be taken stops the script (exit 3) instead of comparing equal to another
  failed one.
• A BT.2020 source with no transfer tag is no longer written as HLG (#463 A153). probe reads
  BT.2020 primaries with no transfer tag as "BT.2020 SDR", and every re-encode on the 10-bit line
  (x265, SVT-AV1, libaom, ProRes, VideoToolbox) tagged the output arib-std-b67. It is tagged
  bt2020-10, the SDR curve BT.2020 defines, and reads back as BT.2020 SDR. A Dolby Vision
  source with no transfer tag keeps the HLG tag it had.
• color.py --to-sdr --force on a file that is not HDR (#463 A161). The note said "tone-mapped
  untagged input as PQ" for every such file, a BT.709-tagged one included, which the chain reads
  as BT.709; it now names the transfer and primaries the chain read the file as. On a file with no
  colour tags at all the run failed as kind: ffmpeg (zscale on FFmpeg 6.1: "no path between
  colorspaces"); a frame missing a tag is labelled with what the chain assumes first.
• fit.py --method trim from the start keeps subtitle and data tracks (#463 A158): it added
  -ss 0.000, which the tool treats as moving the timeline. A trim with --from-center still
  seeks and drops them, as decided.
• An -o without an image extension is a stem in every look.py mode (#463 A159): --compare
  and the contact sheet used it verbatim and failed as kind: ffmpeg; they now write
  <stem>_vs_<other>_<i>_<t>s.png and <stem>_sheet.png, as --at already did.
• A media output with no file extension is refused before ffmpeg runs (the sweep of A159).
  Every tool that writes media handed it to ffmpeg, which failed as kind: ffmpeg ("Unable to
  choose an output format"), and the dry run said completed. It is kind: input,
  error.reason: output_extension_missing, dry runs included, unless the command names its format
  with -f. Side files (a report, an SRT) are not checked; look.py takes such an -o as a stem.
• proxy.py never writes a proxy wider than its source (#463 A162): the default --width 640
  upscaled a smaller source, and a --scale that rounded to 0 became scale=0, which FFmpeg reads
  as the source's width. The width is capped at the source's (even) width and floored at 2, with a
  notes line when either applies.
• sphere.py's default view has square pixels (#463 A163): --v-fov defaulted to 60 whatever
  --h-fov and the frame were, and 90 x 60 at 16:9 squashed the view 2.6 % vertically. Without
  --v-fov the vertical field of view now follows from --h-fov, --width and --height
  (58.7 at the defaults); a pair given that is not square-pixel is drawn as given, with a notes
  line.
• stabilize.py --zoom is the only zoom (#463 A156): vidstab's own "optimal zoom" was always
  on, so --zoom 0 did not keep the original framing and --crop black rarely had an edge to
  fill. The default output now keeps the framing and shows the edges stabilizing reveals.
• The youtube presets stay within YouTube's frame-rate limit (#464 A115): they keep the
  source's rate, so a 120 fps source was exported at 120 fps and check.py --platform youtube
  failed it. A source above the platform's fps_max is exported at its rate over the smallest
  whole divisor within the limit (120 to 60, 119.88 to 59.94, 100 to 50), with a notes line.
• export.py --max-mb with --normalize stays within the size it was encoded to (#464 A112):
  the loudness pass re-encoded the sound at 192k (raised to 256k or 320k on a true-peak overshoot)
  against a budget built on the preset's 128k. Under --max-mb it runs at the preset's bitrate.
• probe.py --field prints JSON, not Python (#464 A123): a boolean, a missing field or an
  object came out as True, None or {'codec': ...}. A string is printed as itself and any
  other value present as JSON; a missing field is still an empty line, as it always was. With
  --json or --json-brief every line is JSON (a string is quoted, a missing field is null).
• verify.py keeps its outputs apart and where it says (#464 A121, A119, A120). Two clip.mp4
  and a clip_1.mp4 were given the output stem clip_1 twice; an indexed stem now skips every
  stem already taken, compared case-folded. --overwrite and --fast reach every step, so a
  rerun into the same --out replaces the earlier outputs instead of failing every step.
  --keep without --out kept nothing (the temporary folder was deleted at exit) and named no
  folder; it keeps the outputs in a new folder, named in the new out_dir (also set for --out,
  null when the outputs were deleted) and on stderr.
• One tool's contract without the other 41 (#467: A220). SKILL.md and references/flags.md
  sent an agent to contract --json, about 950 KB, for one tool's dry_run. contract --json --tool NAME keeps only the named tools under tools (repeatable; cut or cut.py), in the
  order given, and every other key as it is: about 55 to 85 KB for one tool. An unknown name is
  refused (exit 2) with the list of tools; the two pages now name the flag.
• verify.py walks every media extension the skill reads (#467: A218). A folder argument
  skipped .ts, .3gp, .wmv, .gif, .ogg, .opus, .aif/.aiff, .caf and .wma: verify
  kept a list of its own. It uses _common's list less the stills, as batch.py does.
• The npm package no longer ships Python bytecode (#467: A222, A27, A36). With npm 10,
  package.json's files wins over .npmignore, so a checkout that had run the tests packed
  59 scripts/__pycache__/*.pyc files carrying the machine's paths. files ends with negated
  entries for them, and tests/release_check.sh fails a tarball that has one.
• Dead code and unused imports in the runtime (#467: A224): 39 unused imports, the unused
  loudness.py loudnorm string and graphics.py list, and two functions nothing called
  (_platforms.has_frame, _contract._ff_list) are gone, and a test fails on an import a shipped
  module never uses.
• The gallery prints the command its picture was made with (#467: A243, A244).
  cropdetect_crop read the rectangle from the wrong level of cropdetect's document and cropped a
  hard-coded one labelled "detected"; scenes_highlights printed a cut list (0.00-1.50,4.00-5.50)
  the build never ran. Both run the measured value, docs/demos.md prints the one the build
  measures (0.00-2.00,2.00-4.00), and a build that measures anything else stops.
• The agent eval runs the shipped skill and notices a changed checkout (#467: A271, A274).
  evals/run_agents_25.sh linked the whole repository in as the skill, the eval's answer keys
  included, and with every permission check off a run could edit the tracked tree unnoticed. It
  installs the skill with bin/install.js --dir, and exits 1 naming the runs during which the
  checkout's git status changed.
• fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521)

What's Changed

• fix: audit #463, #464 and #467 open items across media tools, delivery and packaging by @kajisho5 in #521

Full Changelog: v3.8.13...v3.8.14