kajisho5/ffmpeg-skill v2.17.4

on GitHub

Automated release: version and notes generated from pull requests merged since 2.17.3.

Fixed

• A font file in a folder with a non-ASCII name draws on Windows. drawtext hands fontfile= to FreeType, which opens it with the ANSI fopen on Windows. A --font-file (or a brand's font_file) under a folder such as フォント or compat dir ünïcode failed to load, and overlay.py, graphics.py and grid.py died with "could not parse fontconfig pattern". Measured on the gyan.dev 9.0.2 runner: the same font under an ASCII folder loaded, spaces included. Such a path is now given in its 8.3 short form, or copied into the run's private drawtext folder under an ASCII name when the drive keeps no short names (font_path_copied). Found when the Windows tests first ran in parallel.
• verify.py --timeout 0 no longer times out every step at once. 0 means no limit, as it does on every other tool; verify.py passed it to Python as a 0-second limit. Its own per-step default (600 s) is now stated in --help.
• The contract says render accepts an audio-only input. It always did (--template podcast and --template audiogram take a WAV/M4A), but contract --json listed it as video_required: true, audio_only: false.

Documentation

• A read-through for statements that contradicted the code or each other:
  • README: join.py always re-encodes, so it is not a stream-copy tool. The no-Node install copies the same files npx ffmpeg-skill does (templates/ and package.json were missing, so no --template rendered). doctor's ok is false when anything any tool requires is missing, with the full required list. The CI matrix now names Ubuntu 26.04's FFmpeg 8.0.1. The contract's deprecated list is described as it is in 2.x. The demo count is gone. The release bumps five files. All nine test groups are listed. The Quick start installs FFmpeg first and says how --transcribe gets a speech engine.
  • references/scripts.md: join is not one of the tools that measure under --dry-run.
  • references/gotchas.md: intermediates are superfast since 2.16.
  • docs/design-decisions.md: probe is none in the contract, not analysis_only.
  • docs/roadmap.md: 2.0.0 is shipped. The 1.19 to 1.21 themes are marked superseded.
  • SKILL.md: references/scripts.md holds most flags, and --help lists all. The description now states the FFmpeg floor next to Python's (FFmpeg 6.1+ and Python 3.10+, with 5.x and 3.9 deprecated).
  • SKILL.md: three rules stated twice now appear once: --overwrite consent lives in step 7 only, "never report success without the output probe" in step 6 only, and a stray blank line is gone. 157 bytes and one line of headroom under the 30,000-byte and 220-line budgets.

Tests

• npm test and CI run the end-to-end modules in parallel (tests/run_parallel.py): the same tests as tests/test_all.py, with each module, and two shards of each of the four largest, in its own process and OUT directory. On a 4-core Linux machine all 936 tests took 21 minutes instead of 35. FFmpeg already uses several threads per encode, so more processes stop helping beyond that. tests/test_all.py still runs them one after another.
• Fix statements that contradicted the code or each other; run the e2e tests in parallel; non-ASCII font paths on Windows (#365)

What's Changed

• Fix statements that contradicted the code or each other; run the e2e tests in parallel by @kajisho5 in #365

Full Changelog: v2.17.3...v2.17.4