| Files | Last commit | Last update |
|---|---|---|
fix: align HDR RGB conversion with its matrix tag (#560) Use the tagged BT.2020 matrix on the 10-bit HDR RGB path. The original implementation and regression fixtures were prepared in Claude Code. Codex verified the full CI and the combined tree on Windows with FFmpeg 9.0.2. Closes #544. Refs #559. これはCodexが処理しました | 8 hours ago | |
fix: captions and text drawing, the open items of audit #457, and #505 (#511) * fix: captions and text drawing, the thirteen open items of audit #457, and #505 - A65: Gurmukhi, Gujarati, Telugu, Kannada, Malayalam, Sinhala, Khmer and Myanmar are detected (they were read as Latin, drawn in the Latin font and reported verified); caption, graphics and grid pick a covering font, overlay --text refuses them as it refuses Devanagari. - A54: phrase wrap keeps the fullwidth ,;:. (and closing brackets) off a line's start in every pass, including cues of three lines or more and the pieces of a split cue. - A62: an emoji PNG in a left/right caption is placed from the Style's side margins. - A59, #505: karaoke waits for the first word, and --offset moves the word timings. - A56: --animate slide ends on the line's own anchor for every position and margin. - A57: --karaoke-style word takes one --animate envelope across its word events. - A58: the size fit measures against the --karaoke-style word budget. - A55: --mode mux neither wraps nor splits the soft track; --max-lines/--wrap there are refused. - A69: the emoji placeholder restores only what it changed, not the whole Style. - A64: doctor's emoji mode is mono when libass draws a glyph and no colour family is listed. - A70: font_used.file under --bold names the face fontconfig matches at bold weight. - A71: in a burn --language is a hint only; in a mux its refusal names --language. The wrap reads ja_JP as ja. - A68: an SRT burn whose cue holds {text} goes through the generated ASS, which draws it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: README says which emoji.mode none carries a warning An explicit --emoji none is a choice and gets an info line; emoji_not_drawn is the automatic fallback when nothing here draws the glyph. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * review: narrow four doc claims, skip the bold-face test where fc-match is missing Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 5 days ago | |
fix(fit): report structured retime details (#597) Expose the requested speed ratio and whether fit actually retimes the clip, including in saved plans. Refs #582 and #581. これはCodexが処理しました | 8 hours ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: the open items of audit #462 (cut, smart cut, silence) and #459 (audio, sync, multicam, scenes, waveform) (#516) * fix: the audit #462 items still real in cut.py, the smart cut and silence.py silence.py - A40: a silence at the start or the end of the sound keeps no margin on that side (FFmpeg 6.1 closes a trailing silence at EOF, so `inf` never caught it) - A41: --edl is written once the cut exists, like --timemap; the same for scenes.py --highlights --edl (after --sheet) and multicam.py --edl / --write-project (after the render) - A42: cut_points are measured on the file's clock (PCM padded from 0) - A43: a frame-snapped bound never lies before 0 (late picture) - A48: --list / --filler-list take the picture from the -o given - A49: --max-cuts counts removal ranges, does not refuse a list, and gives a reason that still holds cut.py / _smartcut.py - A44: --smart copies the last GOP of a cut that runs to the end of the video - A45: a .mkv smart cut is checked against Matroska's 1/1000 time base - A46: a re-encoded single-range audio extraction pads a late sound from 0 - A51: --remove leaves out a leftover with no frame of the picture, with a note Docs: A47 (the time map does not drift), design-decisions section "After 3.8.5 — audit #462", CHANGELOG. One regression test per class. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: audio, sync, multicam, scenes and waveform findings of the 2026-11 audit (#459, 14 items) audio.py: --duck-amount is the depth in dB (a sidechain gate keyed by the speech, A134); --limit and --voice strong's limiter run on the mix when a bed is added (A129); --music-fade-out ends at a short bed's end (A142); --bitrate (and loudness.py --audio-bitrate) reaches an audio output's codec, refused for .wav/.flac (A143); the dynamics help gives ffmpeg's own defaults in dB and a knee's full 0..18.06 dB range (A135). sync.py: the confidence's runner-up is the best other peak (A126); --help states the correction the code applies (A133). multicam.py: overlapping --switch ranges are refused (A132); --fix-drift retimes the picture with the sound (A130). scenes.py: a --highlights range lies inside its scene (A139); the beat grid keeps its refit on a --beat-range bound (A146). waveform.py: an audio input's default output is <name>_waveform.mp4 and an audio -o is refused (A137). Shared: a filter graph over 100,000 characters (8,000 on Windows) reaches ffmpeg in a file, so scenes.py --sheet of thousands of scenes runs (A138); decode_pcm_mono() returns an array('f'), 4 bytes a sample (A140). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #462 and #459 sit under Unreleased Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #462 and #459 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the cut/silence/audio audit items Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam's project states a cut between its clips; a sheet with no scene plans one tile Review of #516: --write-project named no transition, so render.py crossfaded 0.5 s at every switch and the project ran short of the direct render; and scenes.py --sheet --dry-run on a pending input recursed on an empty term list. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: temporary diagnostic for the empty mix on FFmpeg 9 (removed before merge) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam --switch refuses a range shorter than a millisecond The sibling of scenes.py --highlights' empty range: the cut list is written to the millisecond, so such a range was a 0.000-0.000 line cut.py --segments refuses. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam folds a filled range under a millisecond into its neighbour Review of #516: the gap fill made a camera-0 range before --switch "0.0004-3:1" that the cut list wrote as 0.000-0.000. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the FFmpeg 9 diagnostic probes the written file (temporary, removed before merge) The graph writes 576078 bytes on 9.0.2 in every variant; it is ffprobe that cannot read the file. This round records the header, ffprobe's own error and whether the file decodes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a written WAV is probed as a WAV (FFmpeg 9.0.2 read some as MPEG-TS) FFmpeg 9.0.2's MPEG-TS probe outscores the WAV one when the PCM holds 0x47 at a transport packet's spacing, so ffprobe failed the output check of a file the run had written whole ("End of file"). probe() passes -f wav for a file that opens with a RIFF/WAVE header. The two changes made on the earlier reading of this failure (beds mixed in float, no -shortest on audio outputs) are reverted, and the temporary diagnostic test is removed. Inputs: #518. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a bed's fade-out ends where the bed is cut Review of #516: with --length-samples past the main track the fade sat after the bed's trim, and a bed whose sound starts late in its file was faded on its length instead of its timestamps. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the WAV header is read from a regular file only riff_wave_args() opened whatever path probe() was given, and a named pipe blocks on open: the contract test that passes a FIFO (refused before anything opens it) hung every POSIX CI job. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a refused stream costs only itself, and an output is what its name says (#554) * fix: a refused stream costs only itself, and an output is what its name says (#539) The leftovers of #515, #517 and #520, each reproduced on FFmpeg 6.1.1 with lavfi fixtures and swept before the fix (tests/sweep.py and by hand). - run_keeping_subtitles() retried without every subtitle and data stream when one of them failed, so an MP4 with a mov_text track and a timecode track lost the subtitles (into .mkv, and on 6.1.1 into .mp4). It now has export.py's rungs: the data stream is retried away alone first. The fallback records `dropped`, and dropped_non_av_streams is read back. - metadata.py copies every stream, so a stream the output container does not take as a copy failed in the muxer after a dry run that completed. It is refused before ffmpeg runs (kind input). An MP4 with a timecode track, which 6.1.1 would not copy into .mp4, is written without the copied track; the muxer writes it again. - A command that encodes a picture into an audio-only extension is refused in run() (reason output_extension_audio_only), dry runs included; the tools completed with H.264 in .m4a and failed in the muxer for .wav. - audio.py and loudness.py write a late sound's lead as silence in a sound-only output, which played that much early before. - graphics.py's 0.3 s fades and the lower third's slide are cut to a window shorter than they are, on drawtext and on the ASS route. Not changed, recorded in docs/design-decisions.md: caption.py's burn and waveform.py dropping the subtitle track, cut.py -o x.mka, and color.py --retag's re-encode fallback (an existing test pins its field). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings and FFmpeg 9 CI failures of the #539 change (#554 review) Each finding was measured on FFmpeg 6.1.1 before it was changed. - waveform.py's refusal of an audio extension carries the reason output_extension_audio_only, as README says of the picture tools. - The contract's dropped_non_av_streams text states the case it holds for (a source with a subtitle and a data stream). - audio.py reads the lead of the track --audio-stream selects, not the first one's: a second track 0.476 s late lost its lead. - metadata.py judges .m4a as the MPEG-4 family (SubRip into .m4a failed in ffmpeg after a dry run that completed), and its timecode retry into .m4v passes -write_tmcd 1: the ipod muxer wrote no timecode track. - graphics.py: the lower third's slides meet on a centisecond, the instant the ASS route can write (the text led the bar by 11 px at 60 fps); a window the slides fill writes no rest line (its emoji showed for one frame); the sticker's pop is cut with its fade on drawtext too. From the CI log on FFmpeg 9.0.2, not run locally: - The padded lead is cut into 4096-sample frames (asetnsamples, p=0): the FLAC encoder refused the one large frame the pad produced. - The fade's alpha is held in 0..1. This removes a negative alpha on a frame just before --start; the bright first frame seen on 9.0.2 has an alpha of exactly 0, so this may not clear it (see design-decisions). - The new tests no longer assert how many data tracks a muxer writes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the short-window fade test says what its odd frame holds when it fails (#554 CI) The chapter chip's first frame came out far brighter than the rest on FFmpeg 9.0.2 only, which cannot be run locally. On a failure the test now reports each frame's brightest and darkest pixel, the frames before the window and the filter graph, so the next CI run shows what was drawn. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the fade test counts decoded pixels, and three more review findings (#554 review) - tests: the short-window fade test took each frame's brightness from signalstats' YAVG. On FFmpeg 9.0.2 (CI) that named three frames 6.94 above the rest while the same frames' YMAX was 1 above black, which no picture can be: the sum is the filter's, not the frame's. The level is now the mean of the decoded gray pixels; a frame that really flashes still fails, and a failure prints the raw signalstats lines and frame types around --start. - graphics.py: the sticker's pop factor is held in 0..1, as its fade is. - waveform.py: an audio-only output extension is refused before the level measurement, which runs ffmpeg over the whole input (a silent input with --on-silent fail stopped there first). - metadata.py: a planned sound input that is not written yet (a dry run of a later stage) has a placeholder picture, not a measured one, and is no longer refused for it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: require execution evidence for result claims (#588) Derive verification from measured artifacts and lossless from successful stream-copy writes. Keep pending dry-run measurements out of confirmed results. これはCodexが処理しました | 17 hours ago | |
fix: leftovers of the audit, the open items of #510 (#532) * fix: leftovers of the audit, the open items of #510 - broll.py --insert with a still and a --from past 0 shows it for the whole cutaway (looped and bounded by -t); grid.py takes a still cell for the grid's length and refuses a grid of stills only; overlay.py with a still main picture under a looped --image ends with it instead of running to its timeout. - A font file that is not there or not a font is refused (kind: input) from --font-file (overlay --text, graphics) and a brand font_file (those two, and caption.py when it burns). - A brand file key no tool reads is refused with the nearest known key, the key quoted as JSON; "loudness" gets a note. The check is in load_brand(), so waveform.py and a render project's stages refuse it too. - report.py --pack takes only the table render.py --template all writes. - A test pins a silent or too-short output's not_measured loudness step (behaviour since 3.6.7, no source change). Left for a product decision (recorded in design-decisions): render.py's up-front filter check, look.py's HDR path without zscale, cut.py --dry-run's copy plan, and the four #429 batch/render items. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the still loop bound, the still length guard, the font check and the pack table rows Review of #532, each reproduced on 05e40c7 first: - broll.py: a still B is bounded past the end of the trim (two loop frames after B's first frame at or after --from plus the window), not at --from + --duration, which cut off the one frame of a short window (--from 0.041 --duration 0.033 needs the frame at 0.08). - broll.py: the check of B's length past --from skips a still; image2 gives a TGA a one-frame duration and it was refused as too short. - grid.py: a grid whose first cell is a still is named .mp4 with no -o; white_grid.png failed in the image muxer. - graphics.py --template progress draws no text and no longer reads the font file; caption.py with no input video writes the SRT only and no longer reads a brand font_file. - report.py --pack: the line under the header must be the five-column delimiter row, and every later line is a row (`| | |` was skipped as a delimiter instead of refused as a short row). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
feat(mcp): advertise a named selection of tools (#604) Allow a smaller advertised catalogue without removing tools from direct calls. Document measured payload sizes. Closes #573. これはCodexが処理しました | 8 hours ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a late-starting sound in a copied extraction is reported, subtitle tracks kept by crop and stabilize, an overlay's fade-out cut to its window (#542) * fix: a late-starting sound in a copied extraction is reported, subtitle tracks kept by crop and stabilize, an overlay's fade-out cut to its window (#515, #517, #520) #515: an audio-only stream copy (cut.py -o x.m4a without --accurate) of a source whose sound starts after --start begins at the sound's first packet, so the silence before it was left out and the result only said the file was short. The copy is kept (a start off by less than --tolerance is kept, as a keyframe snap is; a lead of --tolerance or more already re-encodes, and the re-encode pads it, A46) and an audio_lead_dropped warning {seconds, segments} and a note say how much went, in a dry run too. Whether the copy should re-encode instead is left open. #520: crop.py and stabilize.py left subtitles to FFmpeg's default stream selection, which on 6.1 drops an MP4's mov_text track; they now map the probed picture and sound and run through run_keeping_subtitles(), as fit.py and color.py do, and report dropped_non_av_streams. The helper converts a text track to the output container's own format (subtitle_codec_for(), the rule export.py and a crop of several tracks already used) instead of copying it and dropping it on a container change. test_delivery's source_carry test used crop.py as the tool with no field for the loss; it uses reverse.py now. #517 item 2: overlay.py --fade N with --end E <= N left the image layer's fade-out out; the fade-out is now min(N, E - start) for image and text layers. #517 item 1 (aif/aiff/caf/wma outputs) needs a product decision (support or refuse) and is not changed. Swept: tests/sweep.py with an MP4 holding a mov_text track and an MKV holding SubRip against a plain clip, before and after; a late-sound clip against a plain one (no row differs: the defect is in the written sound), so every tool that writes an audio output was run by hand to .m4a and .wav and its leading silence measured; the aif/aiff/caf/wma outputs through cut, audio, silence, join and loudness, dry run and real run, from a video and an audio source. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a dry run reports a dropped lead whatever --tolerance is, and the subtitle claims name the retry that drops them (#542 review) A run judges the written length, which is short by up to a packet less than the lead, so a --tolerance between the two kept the copy and warned while the dry run left the warning out for any lead of --tolerance or more. The dry run reports it and its note says a run re-encodes and pads instead when the copy comes out more than --tolerance short. CHANGELOG, design-decisions and references say that a data stream the container refuses beside a subtitle track makes the retry drop both. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: a copy that comes out --tolerance or more short is re-encoded, not only more than it (#542 review) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix(cut): verify the total duration of copied segments (#589) Closes #557. これはCodexが処理しました | 17 hours ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: align HDR RGB conversion with its matrix tag (#560) Use the tagged BT.2020 matrix on the 10-bit HDR RGB path. The original implementation and regression fixtures were prepared in Claude Code. Codex verified the full CI and the combined tree on Windows with FFmpeg 9.0.2. Closes #544. Refs #559. これはCodexが処理しました | 8 hours ago | |
fix: align HDR RGB conversion with its matrix tag (#560) Use the tagged BT.2020 matrix on the 10-bit HDR RGB path. The original implementation and regression fixtures were prepared in Claude Code. Codex verified the full CI and the combined tree on Windows with FFmpeg 9.0.2. Closes #544. Refs #559. これはCodexが処理しました | 8 hours ago | |
fix: a variable-frame-rate cut lays its grid from its start; a hold or pad at the start keeps its last frame (#549) * fix: a variable-frame-rate cut lays its grid from its start; a hold or pad at the start keeps its last frame (#538) Reproduced on FFmpeg 6.1.1 with make_vfr_flash_beep() (27 fps average) against make_flash_beep() (30 fps): - cut.py --start 0.45: flash minus beep 24 ms (plain and --accurate); --start 0.2: 33 ms. fit.py --method trim --from-center (same class): 24 ms at --duration 3.1, 37 ms at 3. Cause: the fps conform laid its grid from the first decoded frame (time base 1/27), then the output-side -ss dropping the seek margin landed between slots and -fps_mode cfr -r rounded again. Fix: cfr_filter(meta, seek=m) shifts the picture so the seek is 0, conforms with start_time=0, moves it back on AVTB. After: -13, -4 and -13 ms (each slot shows the frame nearest start + k/27); frame counts unchanged (96, 103, 84). - freeze.py --at 0 --hold 1: 134 frames of 135. Same in pad.py --start 1 (sibling, 134/135). Cause: tpad's start pad passes its input's end on without the pad, and the fps conform after it ended at the last frame's start (tpad,fps=30 on a 30 fps clip: 149 of 150). Fix: conform first, hold/pad round(T*R) frames of the delivered rate, sound delayed by n/R. After: 135, 122 (0.5 s), 109 (--hold 0.05); flash minus beep 18-19 ms (the grid's own). Sweep: tests/sweep.py VFR --baseline CFR printed no differing row before and after. Every output-side -ss and every tpad start pad in scripts/ read; measured pad --end, freeze at the end, fit trim from the start (already right). Left: fit.py --method trim --from-center --fps R on VFR (24 ms at --fps 27; no -r on its encoder line, so a trailing setpts is not safe on FFmpeg 7+). Tests: test_a_variable_rate_cut_lays_its_grid_from_its_start, test_a_variable_rate_hold_or_pad_at_the_start_keeps_every_frame (9 subtests fail before, pass after); pad rows of test_a_variable_rate_conform_keeps_each_picture_within_half_a_frame updated. Run on 6.1.1: test_editing 286 OK (1 skipped), test_cut_copy 176 OK, test_contract 420 OK (3 skipped). FFmpeg 7.1/8/9 not run here (CI). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the conform rounds each frame's time onto the grid; a slot in a gap keeps the frame before it (#549 review) Measured with numbered frames on FFmpeg 6.1.1: cut at 2.2, inside the VFR fixture's second gap, slots 0-2 show the frame of 2.067 and slot 3 the frame of 2.3, though 2.3 is nearer slot 2. The CHANGELOG, design decision and cfr_filter()'s docstring said each slot shows the nearest source frame; the test's model of the shown frame is the filter's rule now (same expectations on this fixture: the gap's neighbours are dark). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: a refused stream costs only itself, and an output is what its name says (#554) * fix: a refused stream costs only itself, and an output is what its name says (#539) The leftovers of #515, #517 and #520, each reproduced on FFmpeg 6.1.1 with lavfi fixtures and swept before the fix (tests/sweep.py and by hand). - run_keeping_subtitles() retried without every subtitle and data stream when one of them failed, so an MP4 with a mov_text track and a timecode track lost the subtitles (into .mkv, and on 6.1.1 into .mp4). It now has export.py's rungs: the data stream is retried away alone first. The fallback records `dropped`, and dropped_non_av_streams is read back. - metadata.py copies every stream, so a stream the output container does not take as a copy failed in the muxer after a dry run that completed. It is refused before ffmpeg runs (kind input). An MP4 with a timecode track, which 6.1.1 would not copy into .mp4, is written without the copied track; the muxer writes it again. - A command that encodes a picture into an audio-only extension is refused in run() (reason output_extension_audio_only), dry runs included; the tools completed with H.264 in .m4a and failed in the muxer for .wav. - audio.py and loudness.py write a late sound's lead as silence in a sound-only output, which played that much early before. - graphics.py's 0.3 s fades and the lower third's slide are cut to a window shorter than they are, on drawtext and on the ASS route. Not changed, recorded in docs/design-decisions.md: caption.py's burn and waveform.py dropping the subtitle track, cut.py -o x.mka, and color.py --retag's re-encode fallback (an existing test pins its field). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings and FFmpeg 9 CI failures of the #539 change (#554 review) Each finding was measured on FFmpeg 6.1.1 before it was changed. - waveform.py's refusal of an audio extension carries the reason output_extension_audio_only, as README says of the picture tools. - The contract's dropped_non_av_streams text states the case it holds for (a source with a subtitle and a data stream). - audio.py reads the lead of the track --audio-stream selects, not the first one's: a second track 0.476 s late lost its lead. - metadata.py judges .m4a as the MPEG-4 family (SubRip into .m4a failed in ffmpeg after a dry run that completed), and its timecode retry into .m4v passes -write_tmcd 1: the ipod muxer wrote no timecode track. - graphics.py: the lower third's slides meet on a centisecond, the instant the ASS route can write (the text led the bar by 11 px at 60 fps); a window the slides fill writes no rest line (its emoji showed for one frame); the sticker's pop is cut with its fade on drawtext too. From the CI log on FFmpeg 9.0.2, not run locally: - The padded lead is cut into 4096-sample frames (asetnsamples, p=0): the FLAC encoder refused the one large frame the pad produced. - The fade's alpha is held in 0..1. This removes a negative alpha on a frame just before --start; the bright first frame seen on 9.0.2 has an alpha of exactly 0, so this may not clear it (see design-decisions). - The new tests no longer assert how many data tracks a muxer writes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the short-window fade test says what its odd frame holds when it fails (#554 CI) The chapter chip's first frame came out far brighter than the rest on FFmpeg 9.0.2 only, which cannot be run locally. On a failure the test now reports each frame's brightest and darkest pixel, the frames before the window and the filter graph, so the next CI run shows what was drawn. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the fade test counts decoded pixels, and three more review findings (#554 review) - tests: the short-window fade test took each frame's brightness from signalstats' YAVG. On FFmpeg 9.0.2 (CI) that named three frames 6.94 above the rest while the same frames' YMAX was 1 above black, which no picture can be: the sum is the filter's, not the frame's. The level is now the mean of the decoded gray pixels; a frame that really flashes still fails, and a failure prints the raw signalstats lines and frame types around --start. - graphics.py: the sticker's pop factor is held in 0..1, as its fade is. - waveform.py: an audio-only output extension is refused before the level measurement, which runs ffmpeg over the whole input (a silent input with --on-silent fail stopped there first). - metadata.py: a planned sound input that is not written yet (a dry run of a later stage) has a placeholder picture, not a measured one, and is no longer refused for it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: leftovers of the audit, the open items of #510 (#532) * fix: leftovers of the audit, the open items of #510 - broll.py --insert with a still and a --from past 0 shows it for the whole cutaway (looped and bounded by -t); grid.py takes a still cell for the grid's length and refuses a grid of stills only; overlay.py with a still main picture under a looped --image ends with it instead of running to its timeout. - A font file that is not there or not a font is refused (kind: input) from --font-file (overlay --text, graphics) and a brand font_file (those two, and caption.py when it burns). - A brand file key no tool reads is refused with the nearest known key, the key quoted as JSON; "loudness" gets a note. The check is in load_brand(), so waveform.py and a render project's stages refuse it too. - report.py --pack takes only the table render.py --template all writes. - A test pins a silent or too-short output's not_measured loudness step (behaviour since 3.6.7, no source change). Left for a product decision (recorded in design-decisions): render.py's up-front filter check, look.py's HDR path without zscale, cut.py --dry-run's copy plan, and the four #429 batch/render items. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the still loop bound, the still length guard, the font check and the pack table rows Review of #532, each reproduced on 05e40c7 first: - broll.py: a still B is bounded past the end of the trim (two loop frames after B's first frame at or after --from plus the window), not at --from + --duration, which cut off the one frame of a short window (--from 0.041 --duration 0.033 needs the frame at 0.08). - broll.py: the check of B's length past --from skips a still; image2 gives a TGA a one-frame duration and it was refused as too short. - grid.py: a grid whose first cell is a still is named .mp4 with no -o; white_grid.png failed in the image muxer. - graphics.py --template progress draws no text and no longer reads the font file; caption.py with no input video writes the SRT only and no longer reads a brand font_file. - report.py --pack: the line under the header must be the five-column delimiter row, and every later line is a row (`| | |` was skipped as a delimiter instead of refused as a short row). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix(join): accept tab-separated ffconcat directives (#598) Support tab-separated ffconcat directives and preserve actual tabbed filenames in plain lists. Follow-up to shiroha777's PR #596. これはCodexが処理しました | 8 hours ago | |
feat: fit reports where the picture sits; a still where a clip is expected is refused, stabilize takes RGB input (#543) * fix: a still where a clip is expected is refused as an input, stabilize takes RGB input, and fit reports where the picture sits (#531, #519) #531: - A still picture where a length is needed is refused, kind input, naming the still and insert.py (_common.emit.refuse_still): cut.py (was kind output, or 0.04 s verified), fit.py --duration (0.00x; --method trim wrote one frame verified), freeze.py --at past 0, speedramp.py, loop.py (a length of 0.000) and metadata.py --chapters (kind ffmpeg, found by the sweep). freeze --at 0 and look --at 0 still show the one frame. - look.py --at past a still's frame, or past either clip under --compare, is refused as a clip past its end is (was kind output). A still is read without -ss, so --at 0 on a TGA writes its frame. - stabilize.py stabilizes a YUV 4:4:4 copy of an RGB source (vid.stab aborts on RGB); the output's pixel format is the encoder line's, with a note. - caption.py --fonts-dir reads the folder for the asked face; without one the font is reported as fontconfig resolves it, a substitution included. - Not changed, recorded as a product decision: export.py STILL and join.py --transition none with a still. #519: fit.py's result has `picture` after a reframe: the rectangle measured by ffmpeg running the frame's own filters on a white and a black frame, checked against the written file in the test. Swept: tests/sweep.py still.png/still.jpg against a clip, rgb24 qtrle and gbrp H.264 .mov against YUV .mov, plus join, render, verify, sync, multicam and the --compare and prores paths by hand. Siblings left are listed in docs/design-decisions.md "After 3.8.23". Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir reports the face libass matches, fit's cropped rows at the vertical scale Review of #543, seven findings: 1. font_used wording (docs/contract.md, _contract.py, CHANGELOG, references/scripts.md, design-decisions): a folder without the face is reported as a run without --fonts-dir; substituted only when fontconfig gives another family (an installed family: false; Windows: by_name, false). Test: an installed family missing from the folder reports resolved/false (added to the existing fonts-dir test). 2. fonts_dir_face matched any of family/full/PostScript names from fc-scan. Checked by drawing on libass 0.17.1 with fontconfig given no fonts: family is name ID 1 only (not ID 16: "Inter" drew nothing from Inter-Black.otf), full name for TrueType and PostScript name for CFF/Type 1, not the other way. Names are now read from the file's name table (_sfnt_names) and the outline type from fc-scan %{fontformat}. 3. Of a family's faces the one nearest the asked weight (400/700, upright) is reported, for both values of --bold (DejaVuSans + DejaVuSans-Bold: Bold was reported, Regular drawn). 4. fc-scan is given the folder's own files (no subfolders, no dot-files, chunks of 200, no shell); a face in DIR/nested matched before and libass never loads it. Test for 2-4: test_a_fonts_dir_face_is_matched_as_libass_matches_it (fails 6 subtests on the old code). 5. fit.py --fit crop converted top/bottom with the width's scale: 360x640 to 360x202 now reports 219 rows a side, not 220, checked against the written file (a 219-row black band is cut whole, a 220-row one is not). Case added to test_fit_reports_the_picture_rectangle_it_wrote. 6. Not changed: an RGB HDR/BT.2020 source gets BT.601 YUV under BT.2020 tags on every re-encoding line (stabilize, fit, denoise, crop measured alike), and a gbr-tagged one fails every re-encoding tool (sweep); the shared cause is the HDR line, recorded in design-decisions "After 3.8.23" with the claims of this PR narrowed to it. 7. The RGB stabilize test asserts three channels before comparing them. (a) design-decisions #519 entry: only fit.py reports picture; pad.py pads in time, crop.py's rectangle is the caller's, render's frame stage and export's reframing presets are not measured yet. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir names a file only where libass's pick is certain, font_used names one face Review round 2 of #543: A. font_used names one face: drawtext (overlay, graphics, grid) draws every glyph from it (a missing glyph is its .notdef box, same md5 as a noncharacter), libass draws a glyph it lacks from another face (log: "Glyph 0x2605 not found, selecting one more font"), not recorded. Worded in _contract.py, docs/contract.md, README, CHANGELOG. B. A face with only Macintosh-platform name records now falls back to fontconfig's family (FreeType's); a DejaVu Sans copy with its Microsoft records renumbered draws through libass as the original and was reported as no match. C/D. The weight ranking is dropped. A file is named only for a face of the family in exactly the asked style (upright, 400 / 700 with --bold, as libass logs the request) or one asked by full/PostScript name; a folder holding the family in other styles only reports file null, substituted false, with a note. D reproduced: a Bold-only folder with DejaVu Sans installed, regular burn, libass selected the installed DejaVuSans.ttf. C not reproduced (no such installed pair); moot under the new rule. E. The test's fixture guard checks fc-match's family, weight and slant and skips otherwise. Test: test_a_fonts_dir_face_is_matched_as_libass_matches_it (exact faces checked against libass's fontselect log; Bold-only folder; Mac-names fixture; nested/PS/full-name cases kept). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: fit measures the picture with one filter thread (#543 CI) On FFmpeg 8.0.1 (ubuntu py3.13, once in three runs) the white frame's render differed from the black frame's in rows 128-217 of a 360x640 red-pad frame, one slice of rows above the picture, so picture reported 293 rows for the 203 the file held. The two renders are compared byte for byte; they run with -filter_threads 1. The cause in FFmpeg was not isolated and it did not recur on 6.1.1; the test pins the flag and that five runs give the same rows. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
feat: fit reports where the picture sits; a still where a clip is expected is refused, stabilize takes RGB input (#543) * fix: a still where a clip is expected is refused as an input, stabilize takes RGB input, and fit reports where the picture sits (#531, #519) #531: - A still picture where a length is needed is refused, kind input, naming the still and insert.py (_common.emit.refuse_still): cut.py (was kind output, or 0.04 s verified), fit.py --duration (0.00x; --method trim wrote one frame verified), freeze.py --at past 0, speedramp.py, loop.py (a length of 0.000) and metadata.py --chapters (kind ffmpeg, found by the sweep). freeze --at 0 and look --at 0 still show the one frame. - look.py --at past a still's frame, or past either clip under --compare, is refused as a clip past its end is (was kind output). A still is read without -ss, so --at 0 on a TGA writes its frame. - stabilize.py stabilizes a YUV 4:4:4 copy of an RGB source (vid.stab aborts on RGB); the output's pixel format is the encoder line's, with a note. - caption.py --fonts-dir reads the folder for the asked face; without one the font is reported as fontconfig resolves it, a substitution included. - Not changed, recorded as a product decision: export.py STILL and join.py --transition none with a still. #519: fit.py's result has `picture` after a reframe: the rectangle measured by ffmpeg running the frame's own filters on a white and a black frame, checked against the written file in the test. Swept: tests/sweep.py still.png/still.jpg against a clip, rgb24 qtrle and gbrp H.264 .mov against YUV .mov, plus join, render, verify, sync, multicam and the --compare and prores paths by hand. Siblings left are listed in docs/design-decisions.md "After 3.8.23". Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir reports the face libass matches, fit's cropped rows at the vertical scale Review of #543, seven findings: 1. font_used wording (docs/contract.md, _contract.py, CHANGELOG, references/scripts.md, design-decisions): a folder without the face is reported as a run without --fonts-dir; substituted only when fontconfig gives another family (an installed family: false; Windows: by_name, false). Test: an installed family missing from the folder reports resolved/false (added to the existing fonts-dir test). 2. fonts_dir_face matched any of family/full/PostScript names from fc-scan. Checked by drawing on libass 0.17.1 with fontconfig given no fonts: family is name ID 1 only (not ID 16: "Inter" drew nothing from Inter-Black.otf), full name for TrueType and PostScript name for CFF/Type 1, not the other way. Names are now read from the file's name table (_sfnt_names) and the outline type from fc-scan %{fontformat}. 3. Of a family's faces the one nearest the asked weight (400/700, upright) is reported, for both values of --bold (DejaVuSans + DejaVuSans-Bold: Bold was reported, Regular drawn). 4. fc-scan is given the folder's own files (no subfolders, no dot-files, chunks of 200, no shell); a face in DIR/nested matched before and libass never loads it. Test for 2-4: test_a_fonts_dir_face_is_matched_as_libass_matches_it (fails 6 subtests on the old code). 5. fit.py --fit crop converted top/bottom with the width's scale: 360x640 to 360x202 now reports 219 rows a side, not 220, checked against the written file (a 219-row black band is cut whole, a 220-row one is not). Case added to test_fit_reports_the_picture_rectangle_it_wrote. 6. Not changed: an RGB HDR/BT.2020 source gets BT.601 YUV under BT.2020 tags on every re-encoding line (stabilize, fit, denoise, crop measured alike), and a gbr-tagged one fails every re-encoding tool (sweep); the shared cause is the HDR line, recorded in design-decisions "After 3.8.23" with the claims of this PR narrowed to it. 7. The RGB stabilize test asserts three channels before comparing them. (a) design-decisions #519 entry: only fit.py reports picture; pad.py pads in time, crop.py's rectangle is the caller's, render's frame stage and export's reframing presets are not measured yet. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir names a file only where libass's pick is certain, font_used names one face Review round 2 of #543: A. font_used names one face: drawtext (overlay, graphics, grid) draws every glyph from it (a missing glyph is its .notdef box, same md5 as a noncharacter), libass draws a glyph it lacks from another face (log: "Glyph 0x2605 not found, selecting one more font"), not recorded. Worded in _contract.py, docs/contract.md, README, CHANGELOG. B. A face with only Macintosh-platform name records now falls back to fontconfig's family (FreeType's); a DejaVu Sans copy with its Microsoft records renumbered draws through libass as the original and was reported as no match. C/D. The weight ranking is dropped. A file is named only for a face of the family in exactly the asked style (upright, 400 / 700 with --bold, as libass logs the request) or one asked by full/PostScript name; a folder holding the family in other styles only reports file null, substituted false, with a note. D reproduced: a Bold-only folder with DejaVu Sans installed, regular burn, libass selected the installed DejaVuSans.ttf. C not reproduced (no such installed pair); moot under the new rule. E. The test's fixture guard checks fc-match's family, weight and slant and skips otherwise. Test: test_a_fonts_dir_face_is_matched_as_libass_matches_it (exact faces checked against libass's fontselect log; Bold-only folder; Mac-names fixture; nested/PS/full-name cases kept). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: fit measures the picture with one filter thread (#543 CI) On FFmpeg 8.0.1 (ubuntu py3.13, once in three runs) the white frame's render differed from the black frame's in rows 128-217 of a 360x640 red-pad frame, one slice of rows above the picture, so picture reported 293 rows for the 203 the file held. The two renders are compared byte for byte; they run with -filter_threads 1. The cause in FFmpeg was not isolated and it did not recur on 6.1.1; the test pins the flag and that five runs give the same rows. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: require execution evidence for result claims (#588) Derive verification from measured artifacts and lossless from successful stream-copy writes. Keep pending dry-run measurements out of confirmed results. これはCodexが処理しました | 17 hours ago | |
fix: a refused stream costs only itself, and an output is what its name says (#554) * fix: a refused stream costs only itself, and an output is what its name says (#539) The leftovers of #515, #517 and #520, each reproduced on FFmpeg 6.1.1 with lavfi fixtures and swept before the fix (tests/sweep.py and by hand). - run_keeping_subtitles() retried without every subtitle and data stream when one of them failed, so an MP4 with a mov_text track and a timecode track lost the subtitles (into .mkv, and on 6.1.1 into .mp4). It now has export.py's rungs: the data stream is retried away alone first. The fallback records `dropped`, and dropped_non_av_streams is read back. - metadata.py copies every stream, so a stream the output container does not take as a copy failed in the muxer after a dry run that completed. It is refused before ffmpeg runs (kind input). An MP4 with a timecode track, which 6.1.1 would not copy into .mp4, is written without the copied track; the muxer writes it again. - A command that encodes a picture into an audio-only extension is refused in run() (reason output_extension_audio_only), dry runs included; the tools completed with H.264 in .m4a and failed in the muxer for .wav. - audio.py and loudness.py write a late sound's lead as silence in a sound-only output, which played that much early before. - graphics.py's 0.3 s fades and the lower third's slide are cut to a window shorter than they are, on drawtext and on the ASS route. Not changed, recorded in docs/design-decisions.md: caption.py's burn and waveform.py dropping the subtitle track, cut.py -o x.mka, and color.py --retag's re-encode fallback (an existing test pins its field). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings and FFmpeg 9 CI failures of the #539 change (#554 review) Each finding was measured on FFmpeg 6.1.1 before it was changed. - waveform.py's refusal of an audio extension carries the reason output_extension_audio_only, as README says of the picture tools. - The contract's dropped_non_av_streams text states the case it holds for (a source with a subtitle and a data stream). - audio.py reads the lead of the track --audio-stream selects, not the first one's: a second track 0.476 s late lost its lead. - metadata.py judges .m4a as the MPEG-4 family (SubRip into .m4a failed in ffmpeg after a dry run that completed), and its timecode retry into .m4v passes -write_tmcd 1: the ipod muxer wrote no timecode track. - graphics.py: the lower third's slides meet on a centisecond, the instant the ASS route can write (the text led the bar by 11 px at 60 fps); a window the slides fill writes no rest line (its emoji showed for one frame); the sticker's pop is cut with its fade on drawtext too. From the CI log on FFmpeg 9.0.2, not run locally: - The padded lead is cut into 4096-sample frames (asetnsamples, p=0): the FLAC encoder refused the one large frame the pad produced. - The fade's alpha is held in 0..1. This removes a negative alpha on a frame just before --start; the bright first frame seen on 9.0.2 has an alpha of exactly 0, so this may not clear it (see design-decisions). - The new tests no longer assert how many data tracks a muxer writes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the short-window fade test says what its odd frame holds when it fails (#554 CI) The chapter chip's first frame came out far brighter than the rest on FFmpeg 9.0.2 only, which cannot be run locally. On a failure the test now reports each frame's brightest and darkest pixel, the frames before the window and the filter graph, so the next CI run shows what was drawn. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the fade test counts decoded pixels, and three more review findings (#554 review) - tests: the short-window fade test took each frame's brightness from signalstats' YAVG. On FFmpeg 9.0.2 (CI) that named three frames 6.94 above the rest while the same frames' YMAX was 1 above black, which no picture can be: the sum is the filter's, not the frame's. The level is now the mean of the decoded gray pixels; a frame that really flashes still fails, and a failure prints the raw signalstats lines and frame types around --start. - graphics.py: the sticker's pop factor is held in 0..1, as its fade is. - waveform.py: an audio-only output extension is refused before the level measurement, which runs ffmpeg over the whole input (a silent input with --on-silent fail stopped there first). - metadata.py: a planned sound input that is not written yet (a dry run of a later stage) has a placeholder picture, not a measured one, and is no longer refused for it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: audit #463, #464 and #467 open items across media tools, delivery and packaging (#521) * fix: frame size, colour tags, delivery limits and verify's outputs (#463, #464) The items of #463 and #464 still real on 3.8.8, each reproduced on FFmpeg 6.1.1, swept for its class, fixed at the shared cause and pinned by a test that measures the written file: - #463 A153: an untagged BT.2020 transfer was re-encoded as HLG (all 10-bit lines); now bt2020-10 (decision.source_transfer). - #463 A161: --to-sdr --force named every non-HDR input "untagged input as PQ"; sibling: a fully untagged input failed kind ffmpeg (zscale, 6.1). - #463 A154: straighten --fit crop wrote 2 px short on each axis. - #463 A157: crop rounded an odd --x/--y to the chroma grid; sibling: redact left the asked rectangle's last column/row unredacted. - #463 A158: fit --method trim from 0 added -ss 0 and dropped subtitles. - #463 A159: look --compare / the sheet used an extensionless -o verbatim; sibling (every media tool): an extensionless output is refused as input (output_extension_missing) before ffmpeg, dry runs included. - #463 A162: proxy upscaled small sources; tiny --scale became full size. - #463 A163: sphere's default v-fov now keeps pixels square. - #463 A156: stabilize sends optzoom=0, so --zoom is the only zoom. - #464 A115: youtube presets conform a source above fps_max to rate/n. - #464 A112: --max-mb --normalize re-encodes sound at the preset's bitrate. - #464 A123: probe --field prints JSON, not Python repr. - #464 A121, A119, A120: verify stems never collide, --overwrite/--fast reach its steps, --keep alone keeps the outputs and names out_dir. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the open items of audit #467 (memory, distribution, contract size, gallery, eval runner) - A217: decode_pcm_mono returns array('f') (about 8 bytes a sample at the peak, not ~81) - A218: verify.py walks _common's MEDIA_EXT less the stills, as batch.py does - A220: contract --json --tool NAME narrows `tools` (55-85 KB instead of ~950 KB) - A222/A27/A36: negated `files` entries keep __pycache__/*.pyc out of the npm tarball; release_check.sh fails a tarball that has one - A224: 39 unused imports, two dead variables and two uncalled functions removed; an ast test fails on an unused import in the shipped modules - A243/A244: the cropdetect_crop and scenes_highlights demos run the measured value and stop when it is not the one docs/demos.md prints - A271/A274: run_agents_25.sh installs the shipped skill with bin/install.js --dir and exits 1 when the checkout's git status changed during a run A221 was already fixed (#466 A195); A219 is decided in docs/design-decisions.md. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #463/#464/#467 sit under Unreleased after the rebase onto 3.8.13 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: redact blurs a region four pixels wide on FFmpeg 7.1; the dry-run alpha test patches the decode CI of #521: boxblur on FFmpeg 7.1 refuses a chroma radius of 1 on a 2 px chroma plane (6.1 took it), so --mode blur on a 4 px wide region failed. And test_accel patched export._run_analysis, an unused alias this PR removed: the check never guarded anything; it patches alpha_stats, the decode a dry run must not run. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on verify's outputs, probe --field and the eval runner - verify.py plans every file before any step runs and checks every name it will write (step outputs, caption's .srt/.ass, the cue file) against every collected input with refuse_output_is_input (a link or another case of the name included): FOLDER --out FOLDER --overwrite with clip.mp4 and clip_cut.mp4 replaced clip_cut.mp4. Refused as kind input, reason output_is_input, before anything is written. batch.py was swept: it already refuses this (#428), and delivers by rename. - apply_common sets args.json to the effective mode, so --json-brief alone prints the JSON document in probe --field, check, cropdetect, silence --list, report and verify (the sweep of the probe finding). - redact.py: the chroma radius is at least 1 wherever the chroma plane can hold it; --blur-strength 1 left colour-only detail unblurred. - evals/run_agents_25.sh: the checkout's state adds the tracked 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. - CHANGELOG and design-decisions: probe --field's missing field is an empty line without --json/--json-brief, null with either. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's checkout state carries HEAD, and a snapshot that fails stops the run Review of #521: a run that committed its edit left git status clean and the diff against the new HEAD empty; and two failed snapshots compared equal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the committed-edit case of the agent eval commits a file that was clean Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed listing of untracked files fails the agent eval's snapshot Review of #521: ls-files sat in front of a pipe, so its failure was the pipeline's success. The list goes through a file and both statuses count. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the agent eval's snapshot carries the index and hashes an untracked link without following it Review of #521: a run that only restaged a file left the status line and the diff against HEAD as they were; a dangling untracked link failed the snapshot (git hash-object opens the target). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a late-starting sound in a copied extraction is reported, subtitle tracks kept by crop and stabilize, an overlay's fade-out cut to its window (#542) * fix: a late-starting sound in a copied extraction is reported, subtitle tracks kept by crop and stabilize, an overlay's fade-out cut to its window (#515, #517, #520) #515: an audio-only stream copy (cut.py -o x.m4a without --accurate) of a source whose sound starts after --start begins at the sound's first packet, so the silence before it was left out and the result only said the file was short. The copy is kept (a start off by less than --tolerance is kept, as a keyframe snap is; a lead of --tolerance or more already re-encodes, and the re-encode pads it, A46) and an audio_lead_dropped warning {seconds, segments} and a note say how much went, in a dry run too. Whether the copy should re-encode instead is left open. #520: crop.py and stabilize.py left subtitles to FFmpeg's default stream selection, which on 6.1 drops an MP4's mov_text track; they now map the probed picture and sound and run through run_keeping_subtitles(), as fit.py and color.py do, and report dropped_non_av_streams. The helper converts a text track to the output container's own format (subtitle_codec_for(), the rule export.py and a crop of several tracks already used) instead of copying it and dropping it on a container change. test_delivery's source_carry test used crop.py as the tool with no field for the loss; it uses reverse.py now. #517 item 2: overlay.py --fade N with --end E <= N left the image layer's fade-out out; the fade-out is now min(N, E - start) for image and text layers. #517 item 1 (aif/aiff/caf/wma outputs) needs a product decision (support or refuse) and is not changed. Swept: tests/sweep.py with an MP4 holding a mov_text track and an MKV holding SubRip against a plain clip, before and after; a late-sound clip against a plain one (no row differs: the defect is in the written sound), so every tool that writes an audio output was run by hand to .m4a and .wav and its leading silence measured; the aif/aiff/caf/wma outputs through cut, audio, silence, join and loudness, dry run and real run, from a video and an audio source. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a dry run reports a dropped lead whatever --tolerance is, and the subtitle claims name the retry that drops them (#542 review) A run judges the written length, which is short by up to a packet less than the lead, so a --tolerance between the two kept the copy and warned while the dry run left the warning out for any lead of --tolerance or more. The dry run reports it and its note says a run re-encodes and pads instead when the copy comes out more than --tolerance short. CHANGELOG, design-decisions and references say that a data stream the container refuses beside a subtitle track makes the retry drop both. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: a copy that comes out --tolerance or more short is re-encoded, not only more than it (#542 review) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: a variable-frame-rate cut lays its grid from its start; a hold or pad at the start keeps its last frame (#549) * fix: a variable-frame-rate cut lays its grid from its start; a hold or pad at the start keeps its last frame (#538) Reproduced on FFmpeg 6.1.1 with make_vfr_flash_beep() (27 fps average) against make_flash_beep() (30 fps): - cut.py --start 0.45: flash minus beep 24 ms (plain and --accurate); --start 0.2: 33 ms. fit.py --method trim --from-center (same class): 24 ms at --duration 3.1, 37 ms at 3. Cause: the fps conform laid its grid from the first decoded frame (time base 1/27), then the output-side -ss dropping the seek margin landed between slots and -fps_mode cfr -r rounded again. Fix: cfr_filter(meta, seek=m) shifts the picture so the seek is 0, conforms with start_time=0, moves it back on AVTB. After: -13, -4 and -13 ms (each slot shows the frame nearest start + k/27); frame counts unchanged (96, 103, 84). - freeze.py --at 0 --hold 1: 134 frames of 135. Same in pad.py --start 1 (sibling, 134/135). Cause: tpad's start pad passes its input's end on without the pad, and the fps conform after it ended at the last frame's start (tpad,fps=30 on a 30 fps clip: 149 of 150). Fix: conform first, hold/pad round(T*R) frames of the delivered rate, sound delayed by n/R. After: 135, 122 (0.5 s), 109 (--hold 0.05); flash minus beep 18-19 ms (the grid's own). Sweep: tests/sweep.py VFR --baseline CFR printed no differing row before and after. Every output-side -ss and every tpad start pad in scripts/ read; measured pad --end, freeze at the end, fit trim from the start (already right). Left: fit.py --method trim --from-center --fps R on VFR (24 ms at --fps 27; no -r on its encoder line, so a trailing setpts is not safe on FFmpeg 7+). Tests: test_a_variable_rate_cut_lays_its_grid_from_its_start, test_a_variable_rate_hold_or_pad_at_the_start_keeps_every_frame (9 subtests fail before, pass after); pad rows of test_a_variable_rate_conform_keeps_each_picture_within_half_a_frame updated. Run on 6.1.1: test_editing 286 OK (1 skipped), test_cut_copy 176 OK, test_contract 420 OK (3 skipped). FFmpeg 7.1/8/9 not run here (CI). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the conform rounds each frame's time onto the grid; a slot in a gap keeps the frame before it (#549 review) Measured with numbered frames on FFmpeg 6.1.1: cut at 2.2, inside the VFR fixture's second gap, slots 0-2 show the frame of 2.067 and slot 3 the frame of 2.3, though 2.3 is nearer slot 2. The CHANGELOG, design decision and cfr_filter()'s docstring said each slot shows the nearest source frame; the test's model of the shown frame is the filter's rule now (same expectations on this fixture: the gap's neighbours are dark). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: require execution evidence for result claims (#588) Derive verification from measured artifacts and lossless from successful stream-copy writes. Keep pending dry-run measurements out of confirmed results. これはCodexが処理しました | 17 hours ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#534) * fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#522, #527, #529) #522 - sync.py --trim-second refuses an output that is REF (the default name SECOND_synced.mp4 or -o REF): the command reads SECOND only, so run()'s guard never saw REF. Swept: the same in cut --snap-source, metadata --chapters, graphics/overlay --brand and --font-file, silence --words/--filler-words and join --list; all now refused as output_is_input before any write, dry run included. - verify.py no longer replaces an existing <stem>_cues.txt without --overwrite (the caption step fails on it, output_exists); a link or folder there is never written through. - verify.py --report naming a file its steps write is refused (output_path_taken). #527 - The blur pad (fit --fit blur, --pad-fill blur, waveform --image-fit blur) gives boxblur a chroma radius measured on the chroma planes' own size; a 68x120 frame failed in ffmpeg. redact.py's 4:2:0-only chroma bound is the same helper now (4:1:1 and 4:1:0 failed). - fit --smooth interpolate makes its frames at --fps when given (they were repeated). - fit says "source looks variable-frame-rate" when a speed change conforms the picture. - Checked, not changed: speedramp with a 0 Hz sound (does not decode, no output), --smooth blend at --fps (no gain). #529 - sequence.py decodes every frame once (ffprobe over the concat list) and refuses frames that do not decode (still_unreadable, naming them), dry run included. - Checked, not changed: caption sidecars after a timed-out verify step (decided: a deliverable), the colour table (tested against each CI FFmpeg). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the side-file guards against a brand's font and logo, and the sequence frame check's wording Review of #534: - graphics.py --write-ass naming a brand file's font_file replaced the font with ASS text (the guard held --font-file only). Swept every side output against every side input, brand-named files included: caption.py --write-srt, --write-ass and --plan, and render.py --template --write-project (font_file and logo) did the same. graphics guards with the font it draws with, caption with the brand's font_file and logo, render's --write-project with project_input_files(), and load_brand() adds font_file and logo to the files a --plan binds. Every other guard probed held. Seen, not changed: waveform.py --brand fails on any brand file (bare hex 101418 refused by validate_color). - sequence.py's frame check: references/scripts.md, the contract's dry-run note and the still_unreadable reason (and docs/contract.md) say the refusal holds when the check completes, and that a check that cannot be made is a note and the encode runs. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
feat(mcp): advertise a named selection of tools (#604) Allow a smaller advertised catalogue without removing tools from direct calls. Document measured payload sizes. Closes #573. これはCodexが処理しました | 8 hours ago | |
fix: require execution evidence for result claims (#588) Derive verification from measured artifacts and lossless from successful stream-copy writes. Keep pending dry-run measurements out of confirmed results. これはCodexが処理しました | 17 hours ago | |
fix: variable-frame-rate sources in reverse and freeze, and atempo's early sound in speedramp, fit and audio (#541) * fix: variable-frame-rate sources in reverse and freeze, and atempo's early sound in speedramp, fit and audio (#523) #523, each item reproduced on FFmpeg 6.1.1 with make_vfr_flash_beep()'s clip (30 fps with two runs of six frames dropped, 27 fps on average): - reverse.py: every flash 0.19 s after its reversed beep (0.23 s late with --no-audio). The picture is conformed before the reversal with fps=R:round=down (the last frame that starts before each slot ends), so the reversed clip still opens on the source's last frame (test_reverse_of_a_variable_rate_clip_opens_on_its_last_frame, which a plain conform fails). Flashes now within half a 27 fps frame, picture 4.000 s. - speedramp.py's early sound is atempo's: it opens its output half its window into the input (1024 samples at 44.1/48 kHz), so every sound came 21 ms early at 1x, 40 ms at 0.5x, on constant-rate clips too. The same cause reaches fit.py --method speed (17-25 ms) and audio.py --tempo (10-31 ms, about 100 ms through two stages). decision.atempo_chain_with_lead() puts that many samples of silence ahead of the chain (one lead summed over the stages: an adelay between two stages failed speedramp's concat graph on 6.1.1). Beep offsets now within 5 ms on average. - freeze.py --at inside a gap: the picture after the hold 56 ms (--at 0.15) and 93 ms (--at 2.2) early, and short. A hold inside the clip now conforms the picture as it is decoded (cfr_at_input with an origin: the grid laid from the picture's start, on a microsecond time base, because trim rescales its half-frame bound to the fps filter's one-frame base and cut a frame). - test_speedramp_segments_hold_the_frames_between_their_retimed_ends now holds the sound to its place (4 ms) and the picture to its slot separately: a beep on a segment's first sample comes 2.7 ms late with the lead. #526 is not changed: it reproduces (frame 16:9 with fit 1:1 over two clips renders 120x120 and exports 160x90), and which of frame and fit wins is a product decision; the options and a recommendation are in docs/design-decisions.md. Swept: tests/sweep.py VFR --baseline CFR (no row differs, before and after), and every tool that moves time measured on both clips. Left: cut.py --start on a VFR source (24-33 ms, the conform grid on the source clock), freeze.py --at 0 on VFR one frame short (as before). Only FFmpeg 6.1.1 was run here. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a late-starting variable-rate picture is conformed the same on FFmpeg 7.1, and a one-frame hold and an attack under atempo are kept (#523, #541) - FFmpeg 7.1 (CI, debian job 3/3): freeze.py on a VFR picture that starts 50 ms late wrote 109 frames for 135. cfr_at_input(origin=) moved the picture to 0, conformed it, and moved it back with a setpts; from FFmpeg 7.0 setpts leaves its frames no duration and no rate, so the hold's tpad after it cloned nothing. The picture is no longer moved back: freeze.py counts the conformed picture's trims from its own start (6.1.1, 8.0.1 and 9.0.2 passed the same test). - #541 review, freeze.py: a one-frame hold (--hold 0.04 at 30 fps, 0.05 at 27) was a one-frame concat segment, which FFmpeg 6.1.1 gives no length: the frame was lost and the rest ran a frame before its sound (33 ms), on constant-rate sources too. The joined picture's timestamps are rebuilt from its frame count before the closing fps, as in speedramp.py. - #541 review, speedramp.py: the half-window lead put atempo's faded-in first window over the input's first 21 ms, so a 10 ms tone at a segment's start came out 7 dB down at 1x and was lost at 1.5x and 2x (also fit --method speed 2x and audio --tempo 1.5 at a track's start). The chain now gets a window more of silence that is cut off after it, and a 1x segment is not put through atempo at all. Tests: one-frame-hold cases (VFR, late VFR, CFR) in test_a_variable_rate_freeze_keeps_every_frame_of_its_grid, and test_a_tempo_change_keeps_an_attack_at_its_start; both fail on d15e695's code. A single beep in the tempo test is held to half a 30 fps frame (atempo's placement moved one to 12.6 ms at 0.91x). Only FFmpeg 6.1.1 was run here; 7.1, 8 and 9 are verified by CI only. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: a beep slowed through an atempo chain is measured from its first grain, and held to the window it can land in (#523) macOS's FFmpeg 9.0.2 (CI run 43687426111, the only failing job; 6.1.1, 7.1, 8.0.1 and Windows' 9.0.2 passed) failed test_a_tempo_change_keeps_each_sound_ where_it_belongs at audio --tempo 0.3: a 40 ms beep slowed to 0.3x by two atempo stages is laid down as several grains (its third beep read as four silencedetect onsets), the test paired each beep with the nearest onset, and the mean came to 7 ms against a 5 ms bound. The bound claimed more than a chain can hold: the first grain that carries a beep can sit anywhere in a window holding its start. - Each beep is now the first onset from 0.25 s before its place to a window after the stretched beep ends, "not measured" when there is none; a failure prints the onsets, the chain and ffmpeg -version. - A single stage (0.5x-3x) keeps its 5 ms mean and half-a-frame bound. A chain below 0.5x is held to one atempo window (2048 samples, 43 ms at 48 kHz) per beep: -34 to +26 ms on 6.1.1 and macOS 9.0.2, where origin's code is 70-155 ms early and fails. - CHANGELOG and design-decisions narrow the claim to the same. No script changes. Only FFmpeg 6.1.1 was run here (test_editing 280 OK, test_contract 420 OK); 7.1, 8 and 9 are verified by CI only. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: each beep is held to the timing bound of the segment it is in (#541 review) The 3x beeps after speedramp.py's 0.25x segment took the chain's one-window bound from the case; they are held to half a 30 fps frame like any single-stage beep, and the 5 ms mean is read where all four beeps go through one stage. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: the open items of audit #462 (cut, smart cut, silence) and #459 (audio, sync, multicam, scenes, waveform) (#516) * fix: the audit #462 items still real in cut.py, the smart cut and silence.py silence.py - A40: a silence at the start or the end of the sound keeps no margin on that side (FFmpeg 6.1 closes a trailing silence at EOF, so `inf` never caught it) - A41: --edl is written once the cut exists, like --timemap; the same for scenes.py --highlights --edl (after --sheet) and multicam.py --edl / --write-project (after the render) - A42: cut_points are measured on the file's clock (PCM padded from 0) - A43: a frame-snapped bound never lies before 0 (late picture) - A48: --list / --filler-list take the picture from the -o given - A49: --max-cuts counts removal ranges, does not refuse a list, and gives a reason that still holds cut.py / _smartcut.py - A44: --smart copies the last GOP of a cut that runs to the end of the video - A45: a .mkv smart cut is checked against Matroska's 1/1000 time base - A46: a re-encoded single-range audio extraction pads a late sound from 0 - A51: --remove leaves out a leftover with no frame of the picture, with a note Docs: A47 (the time map does not drift), design-decisions section "After 3.8.5 — audit #462", CHANGELOG. One regression test per class. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: audio, sync, multicam, scenes and waveform findings of the 2026-11 audit (#459, 14 items) audio.py: --duck-amount is the depth in dB (a sidechain gate keyed by the speech, A134); --limit and --voice strong's limiter run on the mix when a bed is added (A129); --music-fade-out ends at a short bed's end (A142); --bitrate (and loudness.py --audio-bitrate) reaches an audio output's codec, refused for .wav/.flac (A143); the dynamics help gives ffmpeg's own defaults in dB and a knee's full 0..18.06 dB range (A135). sync.py: the confidence's runner-up is the best other peak (A126); --help states the correction the code applies (A133). multicam.py: overlapping --switch ranges are refused (A132); --fix-drift retimes the picture with the sound (A130). scenes.py: a --highlights range lies inside its scene (A139); the beat grid keeps its refit on a --beat-range bound (A146). waveform.py: an audio input's default output is <name>_waveform.mp4 and an audio -o is refused (A137). Shared: a filter graph over 100,000 characters (8,000 on Windows) reaches ffmpeg in a file, so scenes.py --sheet of thousands of scenes runs (A138); decode_pcm_mono() returns an array('f'), 4 bytes a sample (A140). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #462 and #459 sit under Unreleased Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #462 and #459 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the cut/silence/audio audit items Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam's project states a cut between its clips; a sheet with no scene plans one tile Review of #516: --write-project named no transition, so render.py crossfaded 0.5 s at every switch and the project ran short of the direct render; and scenes.py --sheet --dry-run on a pending input recursed on an empty term list. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: temporary diagnostic for the empty mix on FFmpeg 9 (removed before merge) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam --switch refuses a range shorter than a millisecond The sibling of scenes.py --highlights' empty range: the cut list is written to the millisecond, so such a range was a 0.000-0.000 line cut.py --segments refuses. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: multicam folds a filled range under a millisecond into its neighbour Review of #516: the gap fill made a camera-0 range before --switch "0.0004-3:1" that the cut list wrote as 0.000-0.000. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the FFmpeg 9 diagnostic probes the written file (temporary, removed before merge) The graph writes 576078 bytes on 9.0.2 in every variant; it is ffprobe that cannot read the file. This round records the header, ffprobe's own error and whether the file decodes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a written WAV is probed as a WAV (FFmpeg 9.0.2 read some as MPEG-TS) FFmpeg 9.0.2's MPEG-TS probe outscores the WAV one when the PCM holds 0x47 at a transport packet's spacing, so ffprobe failed the output check of a file the run had written whole ("End of file"). probe() passes -f wav for a file that opens with a RIFF/WAVE header. The two changes made on the earlier reading of this failure (beds mixed in float, no -shortest on audio outputs) are reverted, and the temporary diagnostic test is removed. Inputs: #518. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a bed's fade-out ends where the bed is cut Review of #516: with --length-samples past the main track the fade sat after the bed's trim, and a bed whose sound starts late in its file was faded on its length instead of its timestamps. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the WAV header is read from a regular file only riff_wave_args() opened whatever path probe() was given, and a named pipe blocks on open: the contract test that passes a FIFO (refused before anything opens it) hung every POSIX CI job. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#534) * fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#522, #527, #529) #522 - sync.py --trim-second refuses an output that is REF (the default name SECOND_synced.mp4 or -o REF): the command reads SECOND only, so run()'s guard never saw REF. Swept: the same in cut --snap-source, metadata --chapters, graphics/overlay --brand and --font-file, silence --words/--filler-words and join --list; all now refused as output_is_input before any write, dry run included. - verify.py no longer replaces an existing <stem>_cues.txt without --overwrite (the caption step fails on it, output_exists); a link or folder there is never written through. - verify.py --report naming a file its steps write is refused (output_path_taken). #527 - The blur pad (fit --fit blur, --pad-fill blur, waveform --image-fit blur) gives boxblur a chroma radius measured on the chroma planes' own size; a 68x120 frame failed in ffmpeg. redact.py's 4:2:0-only chroma bound is the same helper now (4:1:1 and 4:1:0 failed). - fit --smooth interpolate makes its frames at --fps when given (they were repeated). - fit says "source looks variable-frame-rate" when a speed change conforms the picture. - Checked, not changed: speedramp with a 0 Hz sound (does not decode, no output), --smooth blend at --fps (no gain). #529 - sequence.py decodes every frame once (ffprobe over the concat list) and refuses frames that do not decode (still_unreadable, naming them), dry run included. - Checked, not changed: caption sidecars after a timed-out verify step (decided: a deliverable), the colour table (tested against each CI FFmpeg). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the side-file guards against a brand's font and logo, and the sequence frame check's wording Review of #534: - graphics.py --write-ass naming a brand file's font_file replaced the font with ASS text (the guard held --font-file only). Swept every side output against every side input, brand-named files included: caption.py --write-srt, --write-ass and --plan, and render.py --template --write-project (font_file and logo) did the same. graphics guards with the font it draws with, caption with the brand's font_file and logo, render's --write-project with project_input_files(), and load_brand() adds font_file and logo to the files a --plan binds. Every other guard probed held. Seen, not changed: waveform.py --brand fails on any brand file (bare hex 101418 refused by validate_color). - sequence.py's frame check: references/scripts.md, the contract's dry-run note and the still_unreadable reason (and docs/contract.md) say the refusal holds when the check completes, and that a check that cannot be made is a note and the encode runs. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: aif, aiff and caf outputs are written, a .wma encode is refused; a still given to export or join is refused (#550) * fix: aif, aiff and caf outputs are written; a .wma output is refused unless it is a stream copy of WMA sound; a still given to export or join is refused naming insert.py (#517, #531) The audio extensions the skill reads were wider than the ones it had a codec for: -o x.aiff/.aif/.caf failed inside ffmpeg after a dry run had planned it, and -o x.wma completed, verified, with H.264 and AAC in an ASF file. AIFF and CAF are now written as 16-bit PCM; a run that would encode into .wma is refused as kind input (output_extension_unsupported) before ffmpeg runs, dry runs included, while a stream copy of WMA sound into .wma is kept. export.py and join.py given a still wrote one frame with verified: true (or refused it for an unrelated reason); both now give the refusal the other tools give a still since 3.9.0, naming insert.py. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a failed WMA copy leaves the file --overwrite named; render refuses .wma only where it writes one (#550 review) - cut.py: the refusal of a re-encode into .wma removed the destination after a failed copy too. A failed copy wrote nothing there (run() replaces an existing path only on success), so that was the caller's earlier file. - render.py: a single uncut WMA clip no stage touches is delivered as it was, as before #517; a --stop-after preview is not refused for a delivery it does not write; a stage's file is refused at the delivery. - cut.py --accurate reports precision "sample" for pcm_s16be (.aif/.aiff/.caf). - test: the planned command is matched through its quoting (Windows paths are quoted). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
feat: fit reports where the picture sits; a still where a clip is expected is refused, stabilize takes RGB input (#543) * fix: a still where a clip is expected is refused as an input, stabilize takes RGB input, and fit reports where the picture sits (#531, #519) #531: - A still picture where a length is needed is refused, kind input, naming the still and insert.py (_common.emit.refuse_still): cut.py (was kind output, or 0.04 s verified), fit.py --duration (0.00x; --method trim wrote one frame verified), freeze.py --at past 0, speedramp.py, loop.py (a length of 0.000) and metadata.py --chapters (kind ffmpeg, found by the sweep). freeze --at 0 and look --at 0 still show the one frame. - look.py --at past a still's frame, or past either clip under --compare, is refused as a clip past its end is (was kind output). A still is read without -ss, so --at 0 on a TGA writes its frame. - stabilize.py stabilizes a YUV 4:4:4 copy of an RGB source (vid.stab aborts on RGB); the output's pixel format is the encoder line's, with a note. - caption.py --fonts-dir reads the folder for the asked face; without one the font is reported as fontconfig resolves it, a substitution included. - Not changed, recorded as a product decision: export.py STILL and join.py --transition none with a still. #519: fit.py's result has `picture` after a reframe: the rectangle measured by ffmpeg running the frame's own filters on a white and a black frame, checked against the written file in the test. Swept: tests/sweep.py still.png/still.jpg against a clip, rgb24 qtrle and gbrp H.264 .mov against YUV .mov, plus join, render, verify, sync, multicam and the --compare and prores paths by hand. Siblings left are listed in docs/design-decisions.md "After 3.8.23". Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir reports the face libass matches, fit's cropped rows at the vertical scale Review of #543, seven findings: 1. font_used wording (docs/contract.md, _contract.py, CHANGELOG, references/scripts.md, design-decisions): a folder without the face is reported as a run without --fonts-dir; substituted only when fontconfig gives another family (an installed family: false; Windows: by_name, false). Test: an installed family missing from the folder reports resolved/false (added to the existing fonts-dir test). 2. fonts_dir_face matched any of family/full/PostScript names from fc-scan. Checked by drawing on libass 0.17.1 with fontconfig given no fonts: family is name ID 1 only (not ID 16: "Inter" drew nothing from Inter-Black.otf), full name for TrueType and PostScript name for CFF/Type 1, not the other way. Names are now read from the file's name table (_sfnt_names) and the outline type from fc-scan %{fontformat}. 3. Of a family's faces the one nearest the asked weight (400/700, upright) is reported, for both values of --bold (DejaVuSans + DejaVuSans-Bold: Bold was reported, Regular drawn). 4. fc-scan is given the folder's own files (no subfolders, no dot-files, chunks of 200, no shell); a face in DIR/nested matched before and libass never loads it. Test for 2-4: test_a_fonts_dir_face_is_matched_as_libass_matches_it (fails 6 subtests on the old code). 5. fit.py --fit crop converted top/bottom with the width's scale: 360x640 to 360x202 now reports 219 rows a side, not 220, checked against the written file (a 219-row black band is cut whole, a 220-row one is not). Case added to test_fit_reports_the_picture_rectangle_it_wrote. 6. Not changed: an RGB HDR/BT.2020 source gets BT.601 YUV under BT.2020 tags on every re-encoding line (stabilize, fit, denoise, crop measured alike), and a gbr-tagged one fails every re-encoding tool (sweep); the shared cause is the HDR line, recorded in design-decisions "After 3.8.23" with the claims of this PR narrowed to it. 7. The RGB stabilize test asserts three channels before comparing them. (a) design-decisions #519 entry: only fit.py reports picture; pad.py pads in time, crop.py's rectangle is the caller's, render's frame stage and export's reframing presets are not measured yet. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: --fonts-dir names a file only where libass's pick is certain, font_used names one face Review round 2 of #543: A. font_used names one face: drawtext (overlay, graphics, grid) draws every glyph from it (a missing glyph is its .notdef box, same md5 as a noncharacter), libass draws a glyph it lacks from another face (log: "Glyph 0x2605 not found, selecting one more font"), not recorded. Worded in _contract.py, docs/contract.md, README, CHANGELOG. B. A face with only Macintosh-platform name records now falls back to fontconfig's family (FreeType's); a DejaVu Sans copy with its Microsoft records renumbered draws through libass as the original and was reported as no match. C/D. The weight ranking is dropped. A file is named only for a face of the family in exactly the asked style (upright, 400 / 700 with --bold, as libass logs the request) or one asked by full/PostScript name; a folder holding the family in other styles only reports file null, substituted false, with a note. D reproduced: a Bold-only folder with DejaVu Sans installed, regular burn, libass selected the installed DejaVuSans.ttf. C not reproduced (no such installed pair); moot under the new rule. E. The test's fixture guard checks fc-match's family, weight and slant and skips otherwise. Test: test_a_fonts_dir_face_is_matched_as_libass_matches_it (exact faces checked against libass's fontselect log; Bold-only folder; Mac-names fixture; nested/PS/full-name cases kept). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: fit measures the picture with one filter thread (#543 CI) On FFmpeg 8.0.1 (ubuntu py3.13, once in three runs) the white frame's render differed from the black frame's in rows 128-217 of a 360x640 red-pad frame, one slice of rows above the picture, so picture reported 293 rows for the 203 the file held. The two renders are compared byte for byte; they run with -filter_threads 1. The cause in FFmpeg was not isolated and it did not recur on 6.1.1; the test pins the flag and that five runs give the same rows. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: align HDR RGB conversion with its matrix tag (#560) Use the tagged BT.2020 matrix on the 10-bit HDR RGB path. The original implementation and regression fixtures were prepared in Claude Code. Codex verified the full CI and the combined tree on Windows with FFmpeg 9.0.2. Closes #544. Refs #559. これはCodexが処理しました | 8 hours ago | |
fix: a VFR conform keeps the picture with its sound; odd overlay and crop positions land where asked; the timeline takes the fit object's frame (#525) * fix: odd overlay/fit-crop positions, the VFR conform's A/V offset, and the timeline's fit aspect (#512, #506) - A VFR source is conformed by the fps filter in the picture's graph (cfr_filter, with_cfr, cfr_at_input, graph_with_cfr) instead of the encoder line's -fps_mode cfr -r alone, which let a frame run up to two frames late: on a 27 fps-average clip flashes came 18/56 ms after their beeps in every tool through cfr_args; now 18-19 ms (within half a frame). speedramp and freeze close on the delivered rate; freeze counts a VFR hold at the nominal rate; fit --method speed puts its retimed picture back on the grid; pad conforms right after tpad; a -vf that is a graph (drawn_chain) gets the conform first. reverse.py is left unchanged (see design-decisions). - overlay.py --image/--video lands at the luma position it names: an odd x/y gets a transparent column/row and the even position before it. - fit.py --fit crop sizes the scaled frame so its margin is a multiple of 4 and computes an even offset, so --crop-x/-y 0, 0.5 and 1 land exactly (straighten.py's rule). - render.py --export-timeline sizes a one-clip sequence by the fit object first, as the render's fit stage does. - fit.py --method speed keeps atempo's own length and the millisecond -t for a retimed track: the sample conform after atempo in -af (from #470) never ended on FFmpeg 8.0.1, so it is taken back out as on the #470 branch (8371ed3, dfc0210) and its test rows go with it; only the picture's fps regrid above is new there. - export.py's platform rate divisor (120 -> 60) is reached by the same fps filter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: changelog entries of #512/#506 sit under Unreleased after the rebase Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the three VFR leftovers of #512 are filed as #523 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the VFR conform's placement, speedramp's segment grid, fit's speed rate and an image layer's size - speedramp.py: each segment holds the frames (and samples) between its retimed ends on the whole output's grid, and the picture and the sound are joined by separate concats. Eight 0.5 s 1x segments of the 27 fps average clip ran 4.148 s and put the last flash (and its beep) at 3.630 s; now 4.000 s, flash at 3.481 s, flash minus beep 18/-19 ms. - fit.py --method speed --fps: the retimed picture goes straight to the rate asked (fps=30,fps=60 kept 60 of a 2x speed-up's 120 pictures). - fit.py, export.py: the rate conform comes before a -vf that is a split/overlay or maskedmerge graph (chain_with_rate, as with_cfr), after any retiming: reels --pad-fill blur wrote 119 of 120 frames, fit --fit blur on the VFR clip 107 of 108, fit --fit blur --fps 60 238 of 240. - overlay.py --image: a still with non-square pixels is made square at its display aspect first, as a --video layer is (x=120 for 220). - references/gotchas.md names the tools that conform with fps, says reverse.py is unchanged and what cfr_args() does to a near-whole rate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: a one-frame speedramp segment keeps its length; fit says when a crop anchor is moved to the grid - speedramp.py: the joined picture's timestamps are rebuilt from its running frame count after the video concat (setpts=N*den/(num*TB) and a closing fps on the delivered rate). FFmpeg 6.1.1's video-only concat gave a one-frame segment no length: three in a row (0-0.1:3, 0.1-0.2:3, 0.2-0.3:3 on a 30 fps clip) put the flash at 0.267 s for 0.3 (113 frames for 114); now 0.300 s, 114 frames. - fit.py --fit crop: an anchor whose offset is not even keeps the even offset nearest it and says so in a notes line naming the axis, the offset asked and the one applied (--crop-x 0.25 over 36 px: 10, not 9). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#534) * fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#522, #527, #529) #522 - sync.py --trim-second refuses an output that is REF (the default name SECOND_synced.mp4 or -o REF): the command reads SECOND only, so run()'s guard never saw REF. Swept: the same in cut --snap-source, metadata --chapters, graphics/overlay --brand and --font-file, silence --words/--filler-words and join --list; all now refused as output_is_input before any write, dry run included. - verify.py no longer replaces an existing <stem>_cues.txt without --overwrite (the caption step fails on it, output_exists); a link or folder there is never written through. - verify.py --report naming a file its steps write is refused (output_path_taken). #527 - The blur pad (fit --fit blur, --pad-fill blur, waveform --image-fit blur) gives boxblur a chroma radius measured on the chroma planes' own size; a 68x120 frame failed in ffmpeg. redact.py's 4:2:0-only chroma bound is the same helper now (4:1:1 and 4:1:0 failed). - fit --smooth interpolate makes its frames at --fps when given (they were repeated). - fit says "source looks variable-frame-rate" when a speed change conforms the picture. - Checked, not changed: speedramp with a 0 Hz sound (does not decode, no output), --smooth blend at --fps (no gain). #529 - sequence.py decodes every frame once (ffprobe over the concat list) and refuses frames that do not decode (still_unreadable, naming them), dry run included. - Checked, not changed: caption sidecars after a timed-out verify step (decided: a deliverable), the colour table (tested against each CI FFmpeg). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the side-file guards against a brand's font and logo, and the sequence frame check's wording Review of #534: - graphics.py --write-ass naming a brand file's font_file replaced the font with ASS text (the guard held --font-file only). Swept every side output against every side input, brand-named files included: caption.py --write-srt, --write-ass and --plan, and render.py --template --write-project (font_file and logo) did the same. graphics guards with the font it draws with, caption with the brand's font_file and logo, render's --write-project with project_input_files(), and load_brand() adds font_file and logo to the files a --plan binds. Every other guard probed held. Seen, not changed: waveform.py --brand fails on any brand file (bare hex 101418 refused by validate_color). - sequence.py's frame check: references/scripts.md, the contract's dry-run note and the still_unreadable reason (and docs/contract.md) say the refusal holds when the check completes, and that a check that cannot be made is a note and the encode runs. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#534) * fix: sync --trim-second and other side inputs as outputs, the blurred pad on a small frame and sequence's bad frames (#522, #527, #529) #522 - sync.py --trim-second refuses an output that is REF (the default name SECOND_synced.mp4 or -o REF): the command reads SECOND only, so run()'s guard never saw REF. Swept: the same in cut --snap-source, metadata --chapters, graphics/overlay --brand and --font-file, silence --words/--filler-words and join --list; all now refused as output_is_input before any write, dry run included. - verify.py no longer replaces an existing <stem>_cues.txt without --overwrite (the caption step fails on it, output_exists); a link or folder there is never written through. - verify.py --report naming a file its steps write is refused (output_path_taken). #527 - The blur pad (fit --fit blur, --pad-fill blur, waveform --image-fit blur) gives boxblur a chroma radius measured on the chroma planes' own size; a 68x120 frame failed in ffmpeg. redact.py's 4:2:0-only chroma bound is the same helper now (4:1:1 and 4:1:0 failed). - fit --smooth interpolate makes its frames at --fps when given (they were repeated). - fit says "source looks variable-frame-rate" when a speed change conforms the picture. - Checked, not changed: speedramp with a 0 Hz sound (does not decode, no output), --smooth blend at --fps (no gain). #529 - sequence.py decodes every frame once (ffprobe over the concat list) and refuses frames that do not decode (still_unreadable, naming them), dry run included. - Checked, not changed: caption sidecars after a timed-out verify step (decided: a deliverable), the colour table (tested against each CI FFmpeg). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings on the side-file guards against a brand's font and logo, and the sequence frame check's wording Review of #534: - graphics.py --write-ass naming a brand file's font_file replaced the font with ASS text (the guard held --font-file only). Swept every side output against every side input, brand-named files included: caption.py --write-srt, --write-ass and --plan, and render.py --template --write-project (font_file and logo) did the same. graphics guards with the font it draws with, caption with the brand's font_file and logo, render's --write-project with project_input_files(), and load_brand() adds font_file and logo to the files a --plan binds. Every other guard probed held. Seen, not changed: waveform.py --brand fails on any brand file (bare hex 101418 refused by validate_color). - sequence.py's frame check: references/scripts.md, the contract's dry-run note and the still_unreadable reason (and docs/contract.md) say the refusal holds when the check completes, and that a check that cannot be made is a note and the encode runs. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 4 days ago | |
fix: a refused stream costs only itself, and an output is what its name says (#554) * fix: a refused stream costs only itself, and an output is what its name says (#539) The leftovers of #515, #517 and #520, each reproduced on FFmpeg 6.1.1 with lavfi fixtures and swept before the fix (tests/sweep.py and by hand). - run_keeping_subtitles() retried without every subtitle and data stream when one of them failed, so an MP4 with a mov_text track and a timecode track lost the subtitles (into .mkv, and on 6.1.1 into .mp4). It now has export.py's rungs: the data stream is retried away alone first. The fallback records `dropped`, and dropped_non_av_streams is read back. - metadata.py copies every stream, so a stream the output container does not take as a copy failed in the muxer after a dry run that completed. It is refused before ffmpeg runs (kind input). An MP4 with a timecode track, which 6.1.1 would not copy into .mp4, is written without the copied track; the muxer writes it again. - A command that encodes a picture into an audio-only extension is refused in run() (reason output_extension_audio_only), dry runs included; the tools completed with H.264 in .m4a and failed in the muxer for .wav. - audio.py and loudness.py write a late sound's lead as silence in a sound-only output, which played that much early before. - graphics.py's 0.3 s fades and the lower third's slide are cut to a window shorter than they are, on drawtext and on the ASS route. Not changed, recorded in docs/design-decisions.md: caption.py's burn and waveform.py dropping the subtitle track, cut.py -o x.mka, and color.py --retag's re-encode fallback (an existing test pins its field). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: review findings and FFmpeg 9 CI failures of the #539 change (#554 review) Each finding was measured on FFmpeg 6.1.1 before it was changed. - waveform.py's refusal of an audio extension carries the reason output_extension_audio_only, as README says of the picture tools. - The contract's dropped_non_av_streams text states the case it holds for (a source with a subtitle and a data stream). - audio.py reads the lead of the track --audio-stream selects, not the first one's: a second track 0.476 s late lost its lead. - metadata.py judges .m4a as the MPEG-4 family (SubRip into .m4a failed in ffmpeg after a dry run that completed), and its timecode retry into .m4v passes -write_tmcd 1: the ipod muxer wrote no timecode track. - graphics.py: the lower third's slides meet on a centisecond, the instant the ASS route can write (the text led the bar by 11 px at 60 fps); a window the slides fill writes no rest line (its emoji showed for one frame); the sticker's pop is cut with its fade on drawtext too. From the CI log on FFmpeg 9.0.2, not run locally: - The padded lead is cut into 4096-sample frames (asetnsamples, p=0): the FLAC encoder refused the one large frame the pad produced. - The fade's alpha is held in 0..1. This removes a negative alpha on a frame just before --start; the bright first frame seen on 9.0.2 has an alpha of exactly 0, so this may not clear it (see design-decisions). - The new tests no longer assert how many data tracks a muxer writes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the short-window fade test says what its odd frame holds when it fails (#554 CI) The chapter chip's first frame came out far brighter than the rest on FFmpeg 9.0.2 only, which cannot be run locally. On a failure the test now reports each frame's brightest and darkest pixel, the frames before the window and the filter graph, so the next CI run shows what was drawn. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix: the fade test counts decoded pixels, and three more review findings (#554 review) - tests: the short-window fade test took each frame's brightness from signalstats' YAVG. On FFmpeg 9.0.2 (CI) that named three frames 6.94 above the rest while the same frames' YMAX was 1 above black, which no picture can be: the sum is the filter's, not the frame's. The level is now the mean of the decoded gray pixels; a frame that really flashes still fails, and a failure prints the raw signalstats lines and frame types around --start. - graphics.py: the sticker's pop factor is held in 0..1, as its fade is. - waveform.py: an audio-only output extension is refused before the level measurement, which runs ffmpeg over the whole input (a silent input with --on-silent fail stopped there first). - metadata.py: a planned sound input that is not written yet (a dry run of a later stage) has a placeholder picture, not a measured one, and is no longer refused for it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Claude <noreply@anthropic.com> | 3 days ago |
| Files | Last commit | Last update |
|---|---|---|
| 8 hours ago | ||
| 5 days ago | ||
| 8 hours ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 17 hours ago | ||
| 4 days ago | ||
| 8 hours ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 17 hours ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 8 hours ago | ||
| 8 hours ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 8 hours ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 17 hours ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 17 hours ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 8 hours ago | ||
| 17 hours ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 3 days ago | ||
| 3 days ago | ||
| 4 days ago | ||
| 8 hours ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 4 days ago | ||
| 3 days ago |