| Files | Last commit | Last update |
|---|---|---|
fix: one reader for caller text and JSON files; SRT cues keyed on timing lines (#483) Caller text files lose their BOM and end lines at \n; SRT cues are grouped by timing line and read only the first two times; caller JSON is refused as kind input when malformed, nested past 64 levels or of the wrong type (projects, plans, recipes, brands); a null brand block keeps its defaults. Findings A12-A20, A66, A73, A262, A268-A270 (#457, #465), plus siblings in join --list, cut --snap-source, metadata, report and the whisper readers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs | 6 days ago | |
fix: the audit's high findings (captions, timing, audio, colour, batch), MCP argument safety, and Scorecard hardening (#469) * Add OpenSSF Scorecard and MCP registry manifests Scorecard is pinned to the commits in ossf/scorecard's own workflow, and it publishes results and uploads SARIF. server.json and glama.json follow the current official schemas. npx ffmpeg-skill mcp starts the stdio server, which is the command the registry package argument names. Co-authored-by: Shogo Kajiyama <kajisho5@gmail.com> * Add join.py --copy (stream-copy join of matching clips) and cut.py --remove join.py --copy --transition none joins clips by the concat demuxer with -c copy when their first video and audio streams match (cut.py's stream signature, now in _common.probe), no clip hides edit-list pre-roll, and each clip's picture and sound end within 50 ms. A mismatch is refused by name (join_copy_mismatch). The join is measured beside the output -- video packet count and length against the clips' -- before it is renamed into place. cut.py --remove A-B,... is the inverse of --segments: the ranges are dropped and the rest is cut exactly as the equivalent --segments would be. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add output_identity, check.py --expect-identity and check.py --compare Every writing tool reports output_identity {size, mtime_ns} for the file it wrote; check.py --expect-identity SIZE:MTIME_NS FAILs an identity row when the file at the path is another one (an older file a failed run left, or a later run's), and check.py reports input_identity. SKILL.md's delivery check passes it. check.py --compare REF measures SSIM (mean and worst frame) and PSNR against a reference in one pass, WARN under a mean SSIM of 0.98. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add export.py --max-mb: a two-pass x264 encode that fits a size, measured The preset's x264 line takes the video bitrate the running time leaves after the audio (97% of the budget) in place of its CRF and runs in two passes; the written size is a file_size verification step, with a size_over_limit warning when the encoder overshot. Refused for non-x264 presets and --hw, with --crf, without a duration, and under 100 kb/s for the picture. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add export.py --preset copy --repair: remux a damaged file and count the damage Timestamps regenerated, corrupt packets dropped, demuxing past errors, negative timestamps shifted to zero; the input is only read. Both files are decoded once at -v warning and the timestamp warnings and other errors counted (repair: {before, after}); the repair step fails if either grew, and errors a remux cannot fix are a repair_incomplete warning. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --repair: judge the remux on timestamp warnings Decode errors are log lines whose count changes with the container (one full run logged 51 before and 53 after), so they are reported before and after but no longer fail the verification step; timestamp warnings, the part a remux rewrites, do. The test source is encoded single-threaded so its damage is the same every run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Label the join copy, identity, compare, max-mb and repair additions 2.24 2.23.0 shipped with check.py's flashes and first-frame rows; these land in the next minor. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address review: cover art, timeline times, copy-join guards, --remove edges, --max-mb - check.py --compare reads V:0 (not cover art), times the worst frame by its presentation time, counts frames from packets when nb_frames is missing, prints that note in plain text, runs under --dry-run when both files exist, and starts each main() with no leftover notes. - join.py --copy compares the first non-cover video stream, judges A/V alignment on each stream's end (start + length), includes colour matrix and range in the signature, refuses --sample-rate/--channels, fails the packet check when nothing was counted, and reports short and silent clips as the re-encoding join does. - cut.py --remove refuses any explicit --start and --snap beats, and clamps a range that starts before 0. - export.py --max-mb refuses nan and inf. - contract: join is conditional on audio; check notes name --compare. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add a cut.py --remove demo; damage only video payload in the repair test The repair test overwrote bytes anywhere in the middle of the TS, which on ffmpeg 5.1.1 hit data its ffprobe could not open the file past; it now flips one byte in 40 video payload packets, never a TS header, the PAT/PMT or audio. The demo shows --remove 2-4 on the gallery's clip. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: refuse a different stream layout and a start offset; --compare pairs by index - The concat demuxer maps each clip's packets by stream index, so a clip whose streams are laid out differently (a cover picture between video and sound) is refused before the signatures are compared. - Picture and sound must start, as well as end, within 50 ms of each other in each clip. - check.py --compare re-times both pictures to their frame index before ssim/psnr, which otherwise pair frames by timestamp. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --expect-identity: grade an existing file under --dry-run too The identity was skipped for every dry run; it is skipped only for a pending placeholder now, as the content rows are. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy checks starts without stream lengths; --expect-identity reads twice - Matroska/WebM state no stream length, so the A/V guard compared nothing there; the starts are now compared whenever both streams exist and the ends only where both lengths are known. - check.py reads the input's identity before measuring and again after: a file replaced mid-check FAILs, and a path that holds no regular file is a FAIL row rather than no row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py: read the identity after --compare; note a mid-check change without --expect-identity The second identity read now follows every measurement, --compare's pass included, so a replacement during that pass is caught. Without --expect-identity no row is added (the row set is unchanged), but a file that changed while it was being checked is named in notes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Route the 2.24 features from SKILL.md; make the repair test open on every ffprobe SKILL.md: routing rows for export --max-mb, cut --remove, join --copy, export --preset copy --repair and check --compare; the report format moved above the routing table (inside the first ~5000 tokens kept after compaction); the --dry-run and VFR-phone rows dropped (covered by the shared-flag line and a gotcha). Checked with a blind behaviour test, original vs trimmed, on 10 prompts: trimmed 60/60, original 56/60. test_delivery: the repair source is one TS appended to itself byte for byte, not bytes flipped inside packets; ffprobe 5.1.1 could not open the latter. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Repair test: fall back to MPEG-PS where ffprobe cannot read MPEG-TS The static ffprobe 5.1.1 build segfaults on every MPEG-TS file (the same reason test_contract gates its TS cut test), so the appended-recording source is written as MPEG-PS there; both containers are what a recorder appends to, and the assertions run unchanged on every runner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Repair test: give the MPEG-PS fallback damage every ffmpeg reports On the static 5.1.1 build an appended program stream decoded without a timestamp warning, so nothing was counted before the repair. The PS copy now also has bytes flipped inside packs of its middle half; the MPEG-2 decoder reports those as decode errors on every version. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Docs: multicam energy switch is a loudness proxy; why 啊 is not a default filler; link ffmpeg-full PATH setup - README / gotchas: `--switch energy` picks the loudest camera, not a recognised speaker; it needs each speaker on their own camera's mic. - design-decisions / scripts.md: Chinese 啊 is out of the zh list for the same reason as `like` (a sentence-final particle far more often than a hesitation); `--filler-extra 啊` adds it. - README HDR section: the zscale note links to Install for putting the keg-only ffmpeg-full on PATH. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Drop color.py --tonemap bt2390 (never worked); bring README in line with the tree color.py: FFmpeg's CPU tonemap filter has no BT.2390 curve in any version (only tonemap_opencl and libplacebo do), so --tonemap bt2390 always failed. The choices are now the filter's own curves; a test checks every choice against the ffmpeg under test and that bt2390 is refused. README, checked against the code: - CI runs on pull requests and pushes to main, not every push; the FFmpeg list includes the Ubuntu 26.04 / 8.0.1 job. - templates include audiogram; --template all renders the seven video destinations, not every template. - doctor's fonts.scripts covers bn, ta, lo and latin too. - --crf was removed except on export.py; --codec/--quality are not on export.py, render.py or batch.py. - output_schema names output_identity (every tool but probe). tests: the Windows skip for the POSIX shim was on the wrong test since #269; it is back on test_dry_run_never_runs_ffmpeg_and_writes_nothing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --compare: pair only frames both files have; RGB and rotated inputs; empty --expect-identity - ssim/psnr ran with eof_action=repeat, so a shorter output's last frame was measured against the rest of the reference (a mean of 0.90 for 52 identical frames, the worst frame timed past the file's end): shortest=1. - An RGB file printed "SSIM R:"/"PSNR r:", which the parser missed ("could not measure"), and the value depended on argument order: both inputs go through format=yuv444p and the parser takes any channel letter. - A clip stored sideways was compared by its stored size although ffmpeg auto-rotates: the size check uses the displayed size. - --expect-identity "" (an unset shell variable) skipped the identity row; it is now malformed (kind input), parsed before any measurement. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: measure stream spans from packets; VP9; --codec; dropped tracks. cut.py --remove outside the file - The A/V end check read stated stream lengths: Matroska/WebM state none, so a clip whose sound stopped 1.5 s early joined as verified; MPEG-TS states an estimate, so every AAC .ts was refused with wrong numbers. Each stream's span now comes from its first and last packet. - VP8/VP9 carry no extradata in WebM: identical clips were refused as "codec configuration differs" (and cut.py --remove/--segments fell back to an x264 re-encode into .webm, which ffmpeg refuses). - --copy with --codec, --quality or --hw ignored them silently; refused now. - A copy join drops subtitle/data tracks: dropped_non_av_streams says so. - cut.py --remove 15-20 on a 4 s file kept a range past the end and removed nothing; a range outside the file is refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --max-mb with gif / huge limits; --repair damage counting; render.py --stop-after keeps its output - --preset gif returned before --max-mb was applied, so a 0.7 MB GIF passed a 1 KB limit as verified: refused. A limit so large it constrains nothing is refused as input instead of failing inside libx264. - --repair counted "Packet corrupt (... dts = ...)" as a timestamp warning and "Last message repeated N times" as a decode error, so a remux that made timestamps worse could pass: repeat+warning prints each occurrence and corrupt-packet lines count as damage. - render.py --stop-after (before export) without --work removed the work directory at exit, deleting the file the result named; it is kept. - docs/contract.md lists size_over_limit and repair_incomplete. - CHANGELOG: the fixes found reviewing 2.24.0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: refuse when a stream span cannot be measured; refuse any --quality - _stream_span skipped a packet without a time and measured the rest, which could invent (or hide) an A/V gap; it returns None, and join_copy refuses the copy when either span is unmeasured instead of skipping the clip. - --quality was refused only when it differed from the default, so --quality 18 passed silently; any --quality given is refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: read packet spans as JSON with a dts fallback; rotation fixture for FFmpeg 5.x The ffmpeg 5.1.1 job refused every copy join as "could not measure": a row of the csv packet listing did not parse as a time there. The spans are read from ffprobe's JSON (only packet entries), and a packet without a pts falls back to its dts; a packet with neither still makes the span unmeasured. -display_rotation is FFmpeg 6.0+; the rotated --compare fixture falls back to the rotate tag, and the test asserts the rotation was written. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: read a long clip's packet span at its head and tail only; test the dts fallback _stream_span listed every packet of the first video and audio stream of each clip to keep two numbers. A clip longer than 30 s is now read at its first 64 packets (past B-frame reordering) and its last 10 s; a test checks that this gives the same span as reading every packet. A packet with no pts is placed by its dts, now pinned by a test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: seek a long clip's tail from its own start time -read_intervals takes timestamps, so the tail read started at duration - 10 s on the absolute clock. A clip whose clock starts late (an MPEG-TS hours in) read from its first packet instead, and one starting below -10 s read nothing and was refused. The seek is start_time + duration - 10 s; a test checks a clip starting at 100 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Restore CHANGELOG.md to main after the 2.24.1 release bump The merge kept the 2.24.1 Fixed entries under Unreleased as well as in the released section. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0: require FFmpeg 6.1+ and Python 3.10+; apply the cut.py defaults decided for 3.0 Work in progress (docs, the MCP snapshot and the tests that pin the 2.x cut defaults follow). - FFmpeg 5.x (deprecated in 2.14.0) and Python 3.9 (2.10.9) move from `deprecated` to `removed`; requirements are >=6.1 and >=3.10. Every tool refuses an ffmpeg/ffprobe it reads as older than 6.1 (kind missing_tool, reason ffmpeg_unsupported); doctor lists `unsupported` and reports ok false; the installer skips Python < 3.10 and prints the FFmpeg line. drawtext's 5.x/6.0 spellings and CI's 5.1.1 static job are gone; CI's Python floor is 3.10. - cut.py: edit list kept on a single .mp4/.mov copy (--no-edit-list for 2.x), SDR HEVC stays HEVC (--no-keep-hevc), --vfr-guard sampled; precision is the least exact segment's and keyframe_snapped equals start_snapped. contract_version 2.0. - package.json / plugin.json / roadmap 3.0.0 (a major is bumped by hand), CHANGELOG 3.0.0, the 2.14.0 release date and the 3.0.0 window waiver in the deprecation-window test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * v3.0: tests for the new cut defaults, sampled guard on seeked TS windows, structured ffmpeg_unsupported refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * design-decisions: point stale test names at their 3.0 successors Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: report the source's start timecode (video stream, tmcd, then format) Requested by nle-bridge-skill (roadmap step 12) for OTIO/EDL/FCP7 source timecodes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test_asr: pick the engine call out of the fake's calls (the 6.1 floor reads ffprobe -version first) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: timecode minutes/seconds 00-59; ffprobe floor in doctor; git builds refused only below 6; README edit-list wording; sampled guard's keyframe start in the contract Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: doctor marks every tool unusable below the floor; an old ffprobe is named even when ffmpeg is old too Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * ci: spread the suite over three machines per platform (run_parallel.py --part I/N) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * run_parallel --part: bound N by the work items; test that every shard lands in exactly one part Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Failures say why: ffmpeg error reasons and hints, no orphan ffmpeg after SIGKILL, no blind re-encode, silent template renders, ASR missing_tool Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: one encoder_missing entry, asr_engine_missing under error.reason, render dry run reads sources and counts audio.replace Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: a talking-head reframe demo from AI-generated footage, on the landing page Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * silence --filler --transcribe: prime Whisper with the filler list; filler_en demo from AI-generated English speech Whisper leaves uh/um out of its text unless its context shows them, so --filler found none on clean speech. The filler list is passed as the initial prompt to whisper.cpp, faster-whisper and openai-whisper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pages: the two AI-footage demos as MP4 with sound, built with grid/graphics/join/look Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * demos: the gallery intro names the two AI-footage demos Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs/demos.md: keep render_project's --export-timeline line (a partial --docs rebuild dropped it) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pages: the reframe label and aria-label give the source's real size; the lede names pages_videos.py Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: SKILL.md restructured, no behavior change SKILL.md keeps Quick start, Workflow 0-8, ask/assume, the does-not-decide line, the report template, the request table (one command per row) and the one-line gotchas. The long forms moved unchanged to references/flags.md, workflow.md, report.md, requests.md and timestamps.md, shipped via package.json files. Description cut to three sentences (924 -> 399 chars); trigger set 55/55 before and after. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: restore three SKILL.md rules the restructure cut; fix moved-text references - step 6 verifies from the writer's --json probe or probe.py (a routing check on agent_prompts_24 lost the verify step on ml1 without it) - step 7 allows --overwrite for a file the user asked to replace - the 60 s highlight row names cut.py --segments as the step that renders - references/flags.md points at SKILL.md's request table, not "the table below" - references/gotchas.md index links are same-file anchors - evals/results/routing-skill-md-restructure.json: old 93/99, new 94/99 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: step 5 keeps --expect-identity; highlight row and --from wording - step 5 runs check.py with --expect-identity S:M (the writer's output_identity), as before the restructure - the highlight row says picks.txt's lines join with commas for --segments - timestamps.md: only broll.py --from takes a time; render.py --from takes a stage name Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix #393-#397: lean builds, tofu emoji, SRT as --text, dropped audio tracks - #393 caption.py / graphics.py ask the build for subtitles/ass first (dry run included) and refuse with reason filter_missing; the path is a named option (filename=) so ffmpeg itself says "No such filter" - #394 look.py without drawtext draws the sheet without timecodes (fallback look_no_drawtext) - #395 doctor fonts.emoji.libass_glyph tells a glyph from the .notdef box; caption.py refuses to burn empty boxes - #396 an SRT given to --text / render --cues is refused with a pointer to --srt - #397 cut.py reports audio tracks it did not write (dropped_audio_streams) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * #395: an auto caption leaves undrawable emoji out with a warning; CodeRabbit fixes - caption.py --emoji auto with no glyph (libass_glyph false) takes mode none, as on a machine with no emoji face, and warns emoji_not_drawn instead of refusing (CI's stock Ubuntu and macOS runners are such machines); --emoji mono still refuses, with or without assets (mono never uses them) - the glyph probe searches the caller's --fonts-dir / --font-file dir and is cached per fonts dir - graphics.py checks the ass filter before writing its ASS sidecar - caption.py's early check asks for ass under --animate / --karaoke - a ';' drop-frame cue line in --text is refused instead of burned as text - docs: the assets exception for libass_glyph, cut.py's default audio pick Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: qualify the emoji none mode and the one-audio-track gotcha CodeRabbit follow-ups on #399: mode none applies when no emoji assets resolve (png with assets); the one-track rule names the editing tools and notes that color.py --strip-dovi copies every track. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Escape line breaks and control characters from arguments in messages (#400) die(), refuse() and every diagnostic record pass their message through neutralize(): a command-line argument carrying a line break or a C0/C1 control, line separator or bidi mark is written escaped wherever the message prints it (as given, resolved, or as a file name), and any such character left is escaped too. Clean paths and a tool's own line breaks are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Report audio tracks every tool leaves out; grid/scenes sheets without drawtext - emit() compares the first input's audio tracks with the written file's and warns audio_tracks_dropped (with a notes line) when the output kept sound but fewer tracks; --no-audio / picture-only outputs report nothing; cut.py's own report is not doubled - grid.py --label auto and scenes.py --sheet on a build without drawtext draw the picture without the text (fallbacks grid_no_drawtext, scenes_sheet_no_drawtext), as look.py does since #394 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #405: escape --opt=value and notes; no track report across several sound sources - neutralize() also matches the value of --opt=value, which tools print back alone - a tool's own notes (join.py's duplicate inputs) are escaped as messages are - audio_tracks_dropped is skipped when two or more inputs with sound were probed (grid --audio-from, a mix): which source's tracks the output keeps is the tool's choice - scenes.py --json-brief goes through emit(), so scenes_sheet_no_drawtext reaches it - test shims no longer write a file named after `ffmpeg -filters` Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: README and the Pages docs layout carry the v3 brand - README opens with the v3 horizontal logo (light and dark) instead of the old app icon and wordmark, links the website, and shows the v3 card (the same image as the social preview, brand/og.png) linking to it - the docs layout (README, Demos, Contract pages) uses the v3 favicon, apple-touch icon and og:image instead of the old app icon and assets/social-preview.png, and puts the symbol in its nav Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: drop the 3.1.3 entries the merge left under Unreleased Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Side files and paths get the output guards (#408) - waveform.py's vis/titled intermediates are hidden, pid-named siblings: the fixed <out>_vis / <out>_titled names were removed whoever had written them (the input included) - --plan FILE, graphics' ASS file, caption --write-srt/--write-ass, verify --report and render's <stem>_pack.md go through guard_sidecar (never the input, never an existing file without --overwrite, a missing directory refused before anything is written) - single-image outputs pass -update 1 and still inputs -pattern_type none, so a % in the name is the file itself, not an image2 sequence - on POSIX a backslash is escaped in filter paths and kept in concat lists (Windows-only conversion to /); edge whitespace in a filter path is escaped - a one-letter x:name.mp4 off Windows is refused as a URL with the ./name hint Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * HDR10 static metadata reaches libx265 and SVT-AV1, and its loss is measured (#401) - probe reads the mastering display and content light level (stream, else the first frame) into video.hdr10_static - the HEVC and AV1 HDR lines pass them as master-display/max-cll and mastering-display/content-light: FFmpeg before 7.0 does not hand them to either encoder - emit() warns hdr10_metadata_dropped when a PQ output still lacks what a PQ source had (prores_ks, libaom) - note() escapes its line as die()/warn() do; batch.py's unverified names are shown() - the audio_tracks_dropped contract text states the several-sources exception Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #412: zero luminance, per-part frame read, V:0, SVT precision, several pictures - a mastering display's 0 cd/m2 minimum is a value (_ratio), not a missing field - the first frame is read when either part is missing from the stream; each part is taken from wherever it is found first; the frame read selects V:0 (not a cover picture) - SVT-AV1 values keep five decimals - hdr10_metadata_dropped is not compared when two or more inputs have a picture (sync --trim-second, a join): the first input need not be the output's source - tests compare the values the re-encodes carry and compute the ProRes test's expected loss Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #412: a still is not a picture source; read the frame when a stream record is incomplete Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix the audit G colour paths (#410) - color.py --lut: lut3d between explicit BT.709 legs through gbrpf32le (an identity LUT shifted every colour); the --lut-strength blend is made in RGB. - join/multicam/grid/broll/overlay --video: a full-range clip in a plain format (VP9/AV1 full range, 10-bit pc) is taken to limited range before it is mixed (range_chain); a converted RGB layer's main likewise (layer_base_chain: graphics boxes/bars, emoji, overlay --image). - drawn_chain on FFmpeg < 7.1: such a picture is taken to limited range before ff_draw / rgb_island draw on it (pad black 101010, white EBEBEB, 10-bit chroma oversaturated). 7.1+ plans are unchanged. - look.py, scenes.py --sheet, export.py --preset gif: an untagged source is labelled BT.709 before the conversion to RGB. - color.py --correct: no colortemperature term at the default 6500 K. - grid.py: a 1x1 grid is refused up front. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Square pixels everywhere that scales; wrap graphics titles and overlay text; fix look/scenes sheets, grid labels, freeze --at, broll odd sizes, EXIF stills; SMPTE start-timecode hint (#411) - probe reports sample_aspect_ratio; display_size()/square_pixels() in _common.probe are the one place a non-square-pixel picture is squared at its display aspect (height kept), used by every tool that sizes, crops, places or draws (fit, export, crop, join, broll, multicam, insert, sequence, grid, look, scenes --sheet, caption, overlay, graphics, redact, proxy, cropdetect, straighten, waveform --image, the render timeline frame). Re-encode-only tools keep the stored size and SAR. Recorded in docs/design-decisions.md. - graphics title/lower-third/chapter/bug wrap with the shared wrapped() helper; the lower-third bar widens to the widest line. overlay --text wraps to the frame and says so in notes. - look's contact sheet stamps from -copyts -start_at_zero (MPEG-TS was a seek late); scenes --sheet selects each scene start by time (VFR); look --compare letterboxes. - grid labels use a font that covers their script, and a long label is shortened from the end. - freeze --at holds the frame shown at T. - An hh:mm:ss:ff time past the end of a timecoded source gets a hint naming the start timecode. - broll scales an odd-sized A-roll to even up front (odd_dimensions_scaled_even). - probe reads an EXIF-rotated still's rotation off its decoded frame. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: match main after the 3.1.4 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(timing): keep picture and sound on one clock in silence, sync, multicam, freeze, speedramp, join, broll, reverse and loop (#409) Every tool in the audit rebased a stream to its own first sample, or cut each stream on its own grid, and lost the A/V offset while reporting verified: true. Each now does what cut.py --segments --accurate does: both streams cut on the same frame-time bounds, shifted by one constant, a late sound padded with silence. Shared helpers in _common/decision.py: stream_starts, frame_rate, grid_at_or_after (cut.py now uses it), av_trim, av_clip_length. - silence.py renders through cut.join_from_source (no declick); bounds snap to the nearest frame time and --timemap is built from them. - decode_pcm_mono pads a late sound (aresample first_pts=0); sync.py and multicam.py place the second's sound on its own clock before delaying it. - freeze.py, speedramp.py cut on frame times; join.py (#338) and broll.py shift the sound by the picture's start. - reverse.py turns both streams about one frame-time end. - loop.py repeats through the concat demuxer instead of -stream_loop. Tests: OneClockTests (test_editing), test_a_late_sound_is_measured_and_placed_on_its_files_clock (test_analysis); two existing tests updated for the new bounds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * render pack guard only where the folder exists; pin the pack table refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG for #410 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: check the MP4 header instead of calling file(1), which the Debian container lacks Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #413: guard every input and output the side files can collide with - write_plan refuses a plan path that is any bound input (side inputs included) or the promised output, default names too - caption's claim() records which artifact holds each path: a different artifact claiming it is refused (output_path_taken); every --srt track is a guarded input - graphics' ASS guard covers --brand/--font-file; verify --report the collected files; render's pack table every existing file argument, and a missing folder at write time is a structured failure - look's frame reads pass still_input_args(); the % check looks at the whole path - waveform removes the title stage's _gfx.ass when this run created it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse hard-linked side files and allow a report inside a new --out (#408) - caption.py claim(): a second name for an already-claimed file (a hard link) is the same path, so --write-srt/--write-ass naming two links to one file are refused before either is written, even under --overwrite. - verify.py: create --out before guarding --report, so a report inside a new output folder is accepted. - tests: plan input compared byte for byte; hard-link and new --out cases. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Protect caption's --brand file from its side-file writes (#408) A --write-srt/--write-ass naming the --brand JSON (or a hard link to it) is refused as output_is_input, --overwrite included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pass -pattern_type only to image2 stills; never reuse a leftover temp name (#408) - still_input_args(): `-pattern_type none` only for an image2 extension. A video or GIF with a % in its name was given it too, and ffmpeg refused the input ("Option pattern_type not found"), so look failed on `take%d.mp4`. probe uses the same helper. - sibling_temp(): a hidden pid-named file that already exists (a leftover of an earlier run with the same pid) is skipped, not reused and then removed. - test: look on a video named with a %. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge origin/main (v3.1.5) into the branch Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * freeze, reverse: clone frames before setpts so the hold survives FFmpeg 7.x From FFmpeg 7.0, setpts sets its output's frame rate to unknown (1/0) and every frame's duration to 0. tpad derives a clone's timestamp step from the link's frame rate, so a tpad placed after av_trim's setpts gave every clone the same timestamp. Those clones were then dropped: - freeze --at: the tpad start clone of the tail added nothing (3 s came out 1.97 s, and the sound ran a second past the picture) - reverse: the stop clone that holds a short picture's last frame added nothing (408 ms off for a sound 0.6 s longer than its picture) The fix uses one graph on every FFmpeg version: - freeze: the held frame is its own concat segment. It is one frame cut on the input's clock, cloned with tpad's stop clone, then shifted. - reverse: tpad runs before av_trim's trim/setpts, as cut.py already does. - Both video chains end in frame_durations(), an fps filter at the source's r_frame_rate. It moves no on-grid frame and gives frames back the durations setpts cleared. Without it, 7.x wrote the last frame with no duration and the picture read one frame short (3.967 s for 120 frames). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address CodeRabbit's review of #416 freeze: hold the last frame when no frame lies at or after --at (no --at, the clip's end), and take the held frame as the first frame after the cut (trim=end_frame=1) so a variable-frame-rate gap still yields one. silence: refuse -o equal to the input before the chunked join, and report keep, kept_duration, removed_seconds and --edl from the rendered (frame-snapped) ranges, EDL frame times truncated to the millisecond. insert runs the square-pixel scale it noted; cropdetect, sequence and waveform --image note theirs; the still-rotation probe passes -pattern_type none. straighten --fit crop scales by |cos| + max(W/H, H/W)|sin|. speedramp refuses a segment holding no frame. graphics places PNG emoji per wrapped line, notes a title taller than the frame and caps the lower-third bar at the margins; overlay --text notes text taller than the frame or too near an edge to wrap; grid shortens a label in a one-em cell; broll scales a 1-px A-roll to 2; the SMPTE hint uses the timecode of the input a time is in. Each with a regression test that fails before the fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * broll: B's sound lasts the whole cutaway when --from falls between frames The sound started at B's first frame at or after --from but ended at --from + length, so a --from between frames left the end of the cutaway without B's sound (23 ms at 30 fps with --audio b). It now ends one cutaway length after its own start. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop: refuse an output that is its input before the concat list The concat list is the only -i of the non-boomerang path, so run()'s own input/output check never saw the source: with --overwrite, -o naming the input replaced it with its own loop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * color --lut: keep a BT.601 source's matrix on the ProRes path --codec prores keeps a non-BT.709 source's matrix and tags instead of converting it, but the LUT legs were pinned to BT.709, so a BT.601 source's samples were read and written with the wrong coefficients. The legs now use the source's matrix on that path; every other path is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Subtitles, transcription, timeline export, Windows fonts and contract drift (#414) - silence.py: --margin kept around silences only, not filler spans - asr: transcripts on the file's clock; no_speech for silent word timings; caption.py --transcribe refuses a source without audio - render --export-timeline: source start timecode, per-clip EDL rate, out clamped to the media end, bad chapters refused - caption.py --mode mux converts existing tracks per container; ffmpeg gets the cues parse_srt read - fonts: an explicit Windows --font resolves through the family table first - contract: publishes the output keys and ASR error reasons the tools emit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Dry-run honesty, process handling, faststart and batch/render orchestration (#415) - render: a file a stage of this run wrote counts as the run's own - metadata.py: MP4/MOV output keeps +faststart - .webm sources default to an .mp4 output on re-encoding tools - batch.py: per-item work dirs, steps without {out}, write-nothing recipes report output null, case-insensitive name collisions - the first failure is reported instead of re-encoding up to three times - timeouts/interrupts after the deliverable name the kept file - probes honour a --timeout below 120 s Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop.py --crossfade: seamless repeats, and every loop measures its seam (#422) --crossfade S fades the clip's last S seconds into its first S (xfade for the picture; afade + amix for the sound, since acrossfade writes nothing when its first input is exactly the fade long). One repeat is the clip minus S, written losslessly to a temporary file and repeated through the concat list, so the output keeps one lossy encode. --times 1 writes one seamless loop file. Every non-boomerang run reports `seam` (picture step across the first seam against the typical frame-to-frame step, sound level step) and checks the written length against the request (`expected_duration`, 1.5 frames) and its streams. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py: DNxHR editing-master presets, and masters read their codec back (#419) dnxhr-lb / -sq / -hq (8-bit 4:2:2), dnxhr-hqx (10-bit 4:2:2) and dnxhr-444 (10-bit 4:4:4), .mov with PCM audio at the source's size and rate, handled like prores (source colour tags kept, no BT.709 matrix conversion, no HDR-to-SDR note). prores and the dnxhr presets now add a `codec` verification step comparing the written codec, profile and pixel format with the preset's; a VideoToolbox line under --hw is not compared. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loudness.py --versions: one master, one measured file per loudness target (#421) Named targets streaming (-14/-1), ebu-r128 (-23/-1) and broadcast-jp (-24/-1, ARIB TR-B32), or LUFS/TP pairs. Each version is a full two-pass normalisation from the master (pass 1 per target, since loudnorm's target_offset depends on it), measured after; the document lists every version, its verification steps carry a `version` key, and `verified` holds only when every version met its own target. The single-target path is the same code, moved into _normalize(). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py: alpha masters that read their alpha back (#417) prores4444 / prores4444xq (16-bit alpha), qtrle, vp9-alpha (WebM, Opus, decoded with libvpx-vp9 to read the alpha) and hap-alpha (needs snappy; sizes in multiples of 4). A source without alpha is refused; a missing encoder is kind missing_tool / encoder_missing. --alpha-mode straight|premultiplied (premultiply filter) is reported as alpha_mode. The `alpha` verification step compares the mean alpha of the first written frame with the source's. PNG/EXR sequences are not included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --preset key-fill: a switcher's fill and key from one alpha source (#418) One decode split into the fill (straight, or premultiplied with --alpha-mode) and the key (alphaextract: alpha 255 is video white), both ProRes 422 HQ with the source's start timecode, written by one ffmpeg run so they hold the same frames. Verified after: key_fill_pair (size, frame count, rate, length equal) and key_level (the key's mean level against the source's alpha); both files' SHA-256 are reported. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * look.py --guides: broadcast safe-area outlines on the preview (#423) action (93 %) and title (90 %) safe per SMPTE ST 2046-1, and 4:3 (the centre 4:3 window of a wider frame, with SD's 90 % action / 80 % title inside it), drawn as outlines in fractions of the frame so a contact sheet's tiles get them too. look.py writes PNGs only, so a guide never reaches a delivery; the result's `guides` lists what was drawn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the filler keep range is the rendered one since #416 (#414) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: refresh the design-decisions contents list (#414) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the #414 decisions ship in 3.1.7 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge origin/main (v3.1.6) into the branch; CHANGELOG from main Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the fake whisper-cli fixture runs on Windows too (#414) No symlinks or shebangs there: the real ffmpeg/ffprobe folders go on PATH and the fake engine is a .cmd wrapper, as test_accel's fake ffmpeg is. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the #415 decisions ship in 3.1.8 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #427 review: margin input, filler message, SRT grammar, mux languages, EIA-608, timecodes - silence.py: --margin nan/inf/negative refused; "output would equal the input" only when the render removes nothing (filler included) - caption.py: the shared-grammar check reads timing lines only; the .webm mux check reads the --audio-stream it copies; tracks report the language/title the output holds - runner: eia_608 is a text subtitle (mov_text in .mp4, copied in .mov) - asr: a whisper segment without `text` is unreadable output, not no_speech - timeline: a start timecode that cannot exist at the source rate is not read; EDL M2 speed at each clip's own rate - contract: dropped_non_av_streams description covers both cases; contract.md drops a wrong "since 3.1.6" Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: skip a start timecode whose frames field the picture's rate cannot reach (#427 review) A 25 fps video tag of 01:00:00:29 was picked before a tmcd track's valid 01:00:00:00, and the export then read the file as having no timecode. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: validate a timecode's frames field at avg_frame_rate first, as probe()'s fps does (#427 review) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG from main (v3.1.7) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #428 review: stopped-child exit codes, batch delivery, copy-preset extension - render.py: an executed plan and a pack destination that were stopped pass the child's exit code (124, 130/143) and kind on; an export stage stopped in --normalize hands the delivery it kept to render's `output`. - export.py: --preset copy keeps the source's extension (a .webm stays .webm); the export counts as delivered as soon as it is written, so a stop in the output probe or loudness read says it is kept. - batch.py: a delivered intermediate is named by its own container (.mp4); delivery holds the output lock (_OutputLock gains on_locked so a worker can raise instead of exit) and copies across filesystems through a sibling temp; the item directory and step files no longer carry the stem, so a stem near the 255-byte limit runs. - Docs: the --timeout 0 and probe-ceiling statements name the reads that keep 120 s; contract.md lists plans and pack destinations among stopped children. - The hung-ffmpeg shim writes a partial file so the test checks its removal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #428 review: dry-run output check, signal exit codes, batch contract - batch.py --dry-run refuses an existing final path that a recipe ending in a read-only step would deliver to, as the real run does. - check.py --dry-run on a file an earlier dry-run step would write grades nothing (checks: [], a note) instead of crashing on a 0 height. - runner.child_exit_code(): a child killed by a signal (returncode -9) is passed on as 128 + the signal by export.py --normalize, render.py's executed plan and its pack, not as a wrapped 247. - The contract's batch outputs and docs/contract.md say a recipe in which no step names {out} writes no artifact (rows carry output null). - Tests compare kept-output paths resolved (macOS /var is /private/var). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse a batch output_dir that is the input folder (#428 review) An output_dir like "sub/.." named the input folder under another name: the input scan did not exclude it, and with suffix "" each final path was the source, which --overwrite replaced. The output dir is now resolved and a batch whose output_dir is the input folder is refused before anything runs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Contract: a render-project batch recipe always writes its artifact (#428 review) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Keep batch rows' output a string; mark a write-free recipe with written: false (#428 review) A recipe in which no step names {out} reported output: null, changing the type of a --json key within 3.x, which the stability table forbids. output stays the final-path string; such rows add written: false, and the cache recognises those entries as claiming no file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * batch: key the cache on the files a step names; fix the QC note (#428 review) A QC pass of check.py --compare REF was served from the cache after REF changed or was removed: only the source and the recipe text were keyed. Every step argument that names a file now contributes that file's content (or its absence) to the key, read on each pass. The note for a recipe that writes nothing said output was null; it says written: false now. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Take main's CHANGELOG after the v3.1.8 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: an encoder capability name may contain a hyphen (libvpx-vp9) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop.py --crossfade: restate the frame rate before xfade (FFmpeg 7.x) FFmpeg 7.1's setpts marks the frame rate unknown (1/0) and xfade then refuses the graph; each trimmed branch now restates its rate. Found by the debian-trixie (ffmpeg 7.1) CI job. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #430 review: key-fill outputs, alpha checks, --versions edges, docs and a demo export.py - key-fill no longer inherits the -map 0:V:0/0:a:0 added for sources with subtitle/data streams (the fill got the unfiltered picture and the sound twice); the dropped tracks are reported in non_av_streams (reason key_fill). - key-fill guards the fill like the key (input under another spelling, an existing fill without --overwrite), stages an existing fill through a temp file, compares the video streams' lengths, and refuses --max-mb. - Alpha presets: a VP9 source is decoded with libvpx-vp9 (the native decoder drops the alpha); a dry run decides on the pixel format and the VP9 alpha_mode tag and decodes nothing; alpha (and key level) is compared over every frame (mean and most transparent frame), not the first; --hw leaves prores4444/xq and key-fill on prores_ks with a note. loudness.py: an explicitly empty --versions is refused; --versions/-o are checked before measuring. write_plan refuses a --plan path that any planned command writes (a version file). look.py: --compare --guides 4:3 checks the second clip's aspect too. loop.py: --crossfade counts frames after the fps= conform and keeps the unit's rate for a variable-rate source (the file missed its own length). Docs: README, SKILL.md (rows extended, no lines added), docs/contract.md, references/scripts.md, design-decisions; loop_crossfade demo. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #430 review: loudness, look, demos, loop, key-fill, alpha - loudness --versions skips the extra default-target measurement - look --compare draws tone map and guide boxes before pad - demos page recorder no longer drops commands after side-file checks - loop refuses nan/inf crossfade; conformed frame count uses container duration - export key-fill stages every fill and removes it on a failed encode - export compares alpha and key level frame by frame when counts match Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix the demo page test off Linux; address the #430 follow-up review - demos/build.py: resolve DEMO_OUT before shortening paths (a Windows 8.3 temp name and macOS's /var symlink left absolute paths in the page) - tests: the --compare guides test checks the stamp only where drawtext exists - docs: the alpha check compares the mean of a 64x36 copy of each frame Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge main (v3.2.0) into the working branch Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release: find a squash commit's PR by its SHA when the subject does not end with (#N) 3.2.3 shipped three features as a patch: the squash subject ended "(#433, carries #432 and #434)", so the resolver matched no PR and never read the feature label. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * source_carry: every result says what the output lost from its source (#401) emit() compares the first input's subtitle tracks, chapters, cover picture, surround channels, creation_time, audio language and start timecode with the written file's. A lost subtitle track, chapter list, cover picture or surround layout is a source_items_dropped warning unless the tool already reports it; tags are listed only. A report, not a check: verified and the written file are unchanged, and a source with none of the items gets no key. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix caption.py --srt --position (#436) and odd-sized JPEGs through the BT.709 conversion (#437) - caption.py: libass force_style reads Alignment in the legacy SSA numbering; the SRT burn passed the numpad value, so only the bottom row landed right. - _common.color: the colorspace filter refuses an odd frame; crop the last row/column before it, and let the -vf merge recognise the prefix. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: the mcp subcommand, the registry manifest and the Scorecard workflow Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review of #435: keep an odd frame's own size through the BT.709 conversion; stop the release on a failed PR lookup - bt709_matrix_filter: pad to even, convert, crop back (exact). The first version cropped the edge off, which shifted a caller's rectangle that ends on the last column and lost a row/column without saying so. - resolve_version: a failed commit->PR request raises instead of returning 'no PR', which was the fallback that turned a feature into a patch. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sweep a reported defect across the tools before fixing it: tests/sweep.py, the sibling-sweep skill, AGENTS.md and the PR template The class behind #437 (an odd frame reaching a step that refuses it) also breaks caption.py and stabilize.py on an odd-sized video; filed as #438. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix odd-sized clips in caption.py and stabilize.py (#438) The runner's even-size retry appended nothing the encoder could see: its scale went at the head of -vf and a later filter made the frame odd again. It now goes at the end. stabilize.py scales an odd frame to even before both vid.stab passes and records odd_dimensions_scaled_even. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin installs in Codex and Gemini CLI find the skill: skills/ffmpeg-skill/SKILL.md (generated), gemini-extension.json, .cursor-plugin/plugin.json (#406) Measured: Codex 0.167 installs the plugin from .claude-plugin/marketplace.json but lists no skill for a SKILL.md at the plugin root, with or without a manifest skills field; skills/<name>/SKILL.md is listed. Gemini CLI 0.68 validates the extension and discovers the same path. Claude Code counts one skill with both files present. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review of #435: bound each sweep child, refuse a failed probe, write the caption cue in its time format, and say what matching rows do not prove Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CONTRIBUTING: Smithery's publish commands no longer exist (its API and CLI stopped on 2026-10-31) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Publish a plugin branch with only the runtime set on each release A plugin install copies what the marketplace entry points at; at the repository root that is 28 MB including tests and evals. release.yml now commits the npm tarball's contents plus the plugin manifest and skills/ (2.6 MB, 88 files) to the plugin branch. The marketplace entry stays at ./ until a release has published the branch. Measured with Claude Code 2.1.322 and Codex 0.167: a git source with ref installs with git alone; an npm source needs npm on the user's machine. design-decisions also records why the plugin does not register the MCP server (12 tools, about 17,000 tokens of definitions per session, for a second route to the same scripts). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * sweep.py: --list needs no input; pin its probe and timeout handling; say where plain paths are in the plugin copy Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Record the questions an outside reviewer kept asking, as a pre-push list 353 inline findings on PRs #92 to #435, grouped into twenty questions in .claude/skills/code-review/references/reviewer-findings.md, with the rule in AGENTS.md to ask them before a push and to add a question when a finding fits none. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin branch step: do not hold the npm publish back; tell a missing branch from a failed lookup Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * reviewer-findings: path guards hold in dry runs; side-file order depends on who reads the file Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --compare --expect-changed: fail on a picture change outside the declared scope (#439) A change_scope row measures the output against the reference tile by tile (32x18, 8-bit RGB difference) and grades it against spans the caller declares: a change elsewhere is FAIL with its time and place, a declared span that did not change is WARN, and two files that cannot be compared are FAIL. The floor (24) is set from phone and screen-recording footage. The plugin branch now holds the plugin under ffmpeg-skill/ for a git-subdir marketplace source: by ref alone Codex kept the whole clone (39 MB); as a subdirectory both Codex and Claude Code copy 2.6 MB. The publish script's remote answers are tested against real temporary remotes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: grade overlapping declarations independently; refuse a rectangle past the edge; unreadable frame times are not compared Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse nan and inf for every float flag of every tool, and for every time string argparse's float reads nan, inf and 1e999; 232 of 232 runs across 25 tools accepted them (tracebacks, ffmpeg 'Invalid argument' as kind ffmpeg, or silently carried). add_common() now registers finite_float for type=float on every parser, and parse_time() refuses a non-finite time. A contract walk tests every numeric flag of all 42 tools. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: --change-floor 254 is asserted by what the document reports, not by a status that depends on the FFmpeg build Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Non-finite flag values: a failure document with --json, as the flags that checked themselves gave; keep 'not a time' in the time parser's refusal Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: rectangles are in the displayed picture (non-square pixels squared before the difference); unreadable frame times are not compared for none either Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: equal displayed sizes are compared whatever the stored size; a different stored size is a note Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: the stored-size note reads the file's own width and height, before rotation Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync CHANGELOG with main after the 3.4.0 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Marketplace entry names the plugin branch's ffmpeg-skill/ directory A plugin install copied the whole repository (28 MB). The entry is now a git-subdir source on the plugin branch, which 3.4.0 published in that layout. README marketplace commands gain --sparse .claude-plugin. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the plugin route's minimum Claude Code version for a git-subdir source Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: a table of the install routes that work today, and the directories not listed yet Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin tree: ship the listing icon, leave the npx installer out The plugin directory's validation of the 3.4.0 plugin branch reported no icon and flagged bin/install.js. The tree now carries .claude-plugin/icon.png and no bin/; the npm package is unchanged. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the one-line plugin install needs 2.1.275; older versions use the two commands Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin tree: a short README of its own; workflow.md says add --overwrite The plugin directory held six findings as credential use; none reads one. Two came from the repository README and one from the word 'pass' beside a URL. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin README: transcription engines may download a model and use a logged-in token Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.1 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin icon: picture chunks only, pixels recompressed; drop mcpName from the plugin's package.json The plugin directory read the 3.4.1 icon as a download-and-run command: a C2PA provenance chunk's text sat beside compressed pixel bytes. The copy on the plugin branch keeps the same pixels and nothing else. assets/icon.png is unchanged. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.2 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin listing: displayName, documentation, support and privacy links; PRIVACY.md The Claude plugin directory showed 'Ffmpeg Skill' with four of six links unset. The plugin branch's package.json also drops scripts. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.3 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: publish the MCP Registry entry after npm, by OIDC server.json was kept on the package version but never sent. The workflow now asks npm and the registry and publishes when npm has the version and the registry does not. mcp-publisher is pinned and its Sigstore bundle is verified. A failure never blocks a release. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: an npm that cannot be asked is not 'not on npm' Only npm's E404 is a quiet skip of the MCP Registry step; any other npm view failure stops the step and reaches the failure reporter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: a tagged version npm does not have fails the registry step A re-run after a failed npm publish skips npm publish because the tag exists; the job would be green with the release on neither registry. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: report both publish failures; recovery text without a flawed recipe The plugin-branch failure step gains always(), since the MCP Registry's failure step now sits before it. The npm-missing message no longer gives a local publish command, and the re-run claim says which version it lists. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: ask the MCP Registry for the exact name and version The list endpoint is paginated; a version on a later page would read as absent and be published twice. 200 and 404 are the only answers. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the MCP Registry entry is live Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the registry row says what the release workflow does when its step fails Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: read the icon word list from the publish script and bound the PNG chunk walk The plugin-branch test kept its own copy of the words the publish script refuses in the icon, and walked PNG chunks without checking that a chunk fits in the file. It now reads the list from the script, walks chunks through one bounds-checked iterator, and pins documentationUrl and supportUrl in the plugin manifest. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: keep the seven known icon words as a fixed minimum Reading the list only from the publish script meant that removing a word there also removed it from the test. The test now requires the seven known words in the script's list and checks the icon against the union. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * feat(silence): --margin-lead and --margin-tail keep different amounts on each side of a silence --margin kept the same amount of each silence on both sides. --margin-tail is kept after speech ends and --margin-lead before it starts again; each defaults to --margin. The result's margins gives the two values used (absent under --filler alone, which removes no silences), and a render.py project takes silence.margin_lead and silence.margin_tail. First part of #449. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * feat(silence): --sentence-min-silence and --sentence-margin-tail for the pause after a sentence end With word timings (--words or --transcribe), a silence that follows a word ending in sentence-final punctuation gets its own floor and its own tail; every other silence keeps --min-silence and --margin-tail. The result lists silence_positions and sentence, and a transcript with no such punctuation changes nothing and warns sentence_ends_not_found. Refused with --speech-aware and without a transcript. Second part of #449. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(changelog): keep the sentence-flags entry under Unreleased after merging the 3.5.0 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): tell the word before a pause by word midpoints, and test the sentence flags on recorded speech The word before a silence was the last one that started before the silence's midpoint. On the repository's one speech clip the engine timed three word starts 0.45 to 0.5 s before the sound, inside the pause before them; a start stretched a little further would have made the next word the one before the pause. A word's own midpoint stays on its side of the pause whether the engine stretched the previous word's end or the next word's start. Adds a test on demos/footage/ai_talk_en.mp4 with its recorded word timings: the three pauses after sentences are told from the others, each keeps its tail, and no cut lands near sound. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): measure an open-ended silence, refuse unusable word timings, add the pauses_en demo Review of #452: - A silence that runs to the end of the file was kept whatever its length, so a 0.5 s pause after the last sentence could be cut under a 3 s floor. Its length is what is left of the file. - Words with a non-finite time or no length are left out of the sentence classification, and a transcript with none left is refused. - The sentence_ends_not_found warning and the docs said the flags changed nothing; with --filler they still turn silence removal on. They now say what happens: every silence uses --min-silence and --margin-tail. - demos: pauses_en runs the lead, tail and sentence flags on the recorded speech clip (CONTRIBUTING asks a feat PR for a before/after demo). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): a silence that runs to the end of the file ends at the file's end when placing the word before it The last word's timing can reach into the trailing silence; with the silence's start as the reference its midpoint fell after it and the pause was read as inside a sentence. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test(silence): the sentence test makes its own clip OUT/gappy.mp4 is written by two fixtures with different layouts (MediaFixtures._gappy and test_silence_removal), so which seconds are silent depended on test order: the test passed locally and failed on CI, where the other layout was on disk. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(silence): margins is present under --filler when a sentence flag turns silence removal on Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * chore: sync the branch with main after #452 and the 3.6.0 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): refuse DASH/IMF manifests by content, subfile URLs and line breaks in concat paths; never write temps through a planted link From the full audit: - A DASH manifest saved as .mp4 made every tool follow its BaseURL to another local file or over HTTP (FFmpeg 6.1's -protocol_whitelist file did not stop it). DASH and IMF are now told by the content ffmpeg's own probes look for. - looks_like_url() stopped at the comma of subfile,<options>:<url>. - concat_list_line() wrote a path with a line break as is, adding directives. - The staging temp and place_output()'s temp had fixed names opened through whatever was there, a planted symlink included; they are created with O_EXCL, and a lock's default audit log is opened with O_NOFOLLOW. - render.py's output-is-input guard covered clips and overlay videos only. - color_hex() checked the length only, so 'yellow' drew black, verified. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): export.py's fill temp and batch.py's delivery temp are reserved exclusively too Sibling sweep of the fixed hidden temp names: both built .<stem>.ffskill-... names and opened them through whatever was there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): copy through the reserved descriptor, refuse the 3GPP DASH profile, guard overlay fonts Review of #456: - place_output() and batch.py's cross-filesystem delivery closed the exclusively created temp and reopened it by name with copyfile(); they now write the descriptor that created it (copy_to_sibling()). - dash_probe() also accepts 3GPP:PSS:profile:DASH1 in place of dash:profile. - project_input_files() takes overlays[].font_file, and an overlays value that is not a list is left to the overlay stage instead of a traceback. - audiogram.image reaches waveform.py resolved against the project file, the path the output guard checks. - Without O_NOFOLLOW (Windows), a symlink at the default audit-log name is refused by looking first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): guard the project file, resolve overlay fonts, fail closed on an unreserved temp Second review of #456: - render.py's output guard includes the project file itself - an overlay's font_file reaches overlay.py resolved against the project file - reserve_sibling() refuses (output_dir_not_writable) when the exclusive create fails instead of returning a name nobody reserved - copy_to_sibling() closes the descriptor if wrapping it fails Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): verify the default audit log after opening it where O_NOFOLLOW is missing Without O_NOFOLLOW a link could appear between the islink() look and the open; the descriptor is now compared with the name (lstat) before anything is written, and closed unwritten if they differ. The render guard test compares forwarded paths by file identity, not spelling (Windows 8.3 short names). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): guard the brand's own logo and font, skip the output guard on previews Third review of #456: a brand file's font_file and logo (resolved against the brand file, as load_brand() does) are inputs the output may not replace; a --stop-after preview never writes the output, so it is not refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): refuse a hard-linked default audit log where O_NOFOLLOW is missing A link swapped for a hard link to its target before the lstat passed the identity check; the opened file must also be a regular file with one name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: require the --stop-after preview to succeed, not only to avoid one refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * chore: sync with main after #456 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(batch): project side files beside the project; --watch waits for a source to settle A batch project recipe wrote each item's project copy into its work directory with relative paths unchanged, and render.py reads them against the project file's directory, so "bed.wav", a brand, captions or a chapters file beside the project failed every item. The copy now carries them absolute through one list of a project's path keys (_common.decision.PROJECT_PATH_KEYS / rebase_project_paths), which render.py applies to its own project too (audiogram.image and overlays[].font_file were read against the working directory) and multicam.py --write-project applies before writing (cwd-relative clip paths in a file written elsewhere). batch.py --watch processed a source still being copied, reported the truncated output ok (verified for MKV/TS), and once the copy finished refused every later pass with output_exists. A source is now processed once its size and mtime are unchanged across one poll (first seen: last modified at least one interval ago); until then its row is skipped: "settling", not cached. An output an earlier pass of the same watch delivered and nobody changed since is replaced; any other existing output is still refused without --overwrite. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit fb471e619fab764305dd6ec616dcd377ab7791b7) (cherry picked from commit b32fe8bcdc97f83cc1ad61905f0077072afb351d) * fix(captions): keep text symbols, spaces at moved breaks and SRT markup Three ways the burned caption differed from the words handed in, each fixed at its shared cause: - emoji.py counted the whole Misc Symbols / Dingbats / Misc Symbols and Arrows blocks (and the playing cards) as emoji, so a mode-none run cut out SDH music notes and check marks. In those blocks only Emoji_Presentation characters start an emoji; a text-default symbol does with VS16 or a skin tone. Reaches caption, graphics and overlay. - wrap.py's orphan, weak-line and rebalance passes guessed from the script whether a space stood at a break, dropping it for Korean, spaced Japanese/Chinese, emoji neighbours and Latin/CJK pairs. The greedy fill records each break's space and every pass carries it; Korean is wrapped as a spaced script. - caption.py's own ASS route escaped SRT markup and drew it as text. <i> <b> <u> <s>, <font color> and {\anN} become the overrides ffmpeg writes; wrap, fit, emoji placement and the blank-cue count measure the words only; split cues carry their tags; other {\...} blocks stay escaped (design-decisions entry). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit f6fc0e390c5b17ddce708727fa7ed3cae7db78ca) (cherry picked from commit bd48428c2cfe34e93c86052f9ae62d689d02a0bd) * fix: speedramp frame durations, sync --trim-second copy, upper-case MPEG-4 suffixes, moved start timecode - speedramp: restore each segment's frame durations after setpts (frame_durations), so a sped-up segment no longer pushes the rest of the picture two frames late and adds frames. - sync --trim-second: never stream-copy a video second file (the copy started at the keyframe before the trim point); re-encode from the frame. A sound-only file is still copied. - cut: lower-case the output suffix in the join paths, so .MOV/.MP4 keep hvc1 and +faststart; signatures_match recognises an upper-case .TS. - Start timecode: cut (one range), fit --method trim and sync --trim-second write the label of the frame the output starts on (render clips follow via cut); a join and an MXF copy write none. source_carry compares the timecode's value. Swept: suffix comparisons across scripts/ (only cut.py and signatures_match missed .lower()); setpts rescales feeding a concat or ending a graph (fit --method speed measured exact; freeze and reverse already restore durations); stream copies behind an input -ss (only sync's); tests/sweep.py on a timecoded MOV against a plain one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 1902b948f54d99f9d8e6758c6ffe1e717f3357dd) (cherry picked from commit 6eae59b1b0e4f377f743bdd3fd660011b6730cfb) * fix: downmix gain, HDR tone-map peak, PiP opacity colour and 10-bit overlays - audio.py --downmix: ITU-R BS.775 weights (centre and surrounds -3 dB, LFE dropped), renormalised with pan's '<' so in-phase channels cannot clip; pan's '=' summed 2.7x per side and a -6 dBFS 5.1 came out hard-clipped while the run said verified. - color.py --to-sdr and look.py's HDR sheets: linearise at npl=100 (the unit the tonemap filter reads) and pass the source peak as tonemap's peak, so a PQ pixel at --peak nits and HLG 100 % reach SDR white (they stopped at Y 151); --peak is a PQ setting, HLG peaks at 1000 nits. The chains move to _common/color.py; look.py converts a BT.2020 SDR source's gamut only. - overlay.py --video --opacity scales the alpha with lut in YUV instead of the RGB-only colorchannelmixer, which made the whole main picture composite in RGB and come back BT.601 under its BT.709 tag. - overlay_format(): broll, redact (blur/pixelate), graphics layers and boxes, emoji (graphics, caption), overlay --video and --image composite a 10-bit HDR/ProRes picture at 10 bits instead of overlay's 8-bit default (or 8-bit RGB for --image); a 10-bit PiP stays 10-bit. Regression tests: DownmixLevelTests, HdrLevelsAndDepthTests (fail on the previous tree, pass here). Design decisions, README, references/scripts.md, demos and CHANGELOG updated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit fb2123404d2880ae39ca9c3e797b8294a60cf950) (cherry picked from commit 2a54302d7f3388d152c368a4b0d2dc17d7026cf8) * fix(mcp): truthful tool annotations; a structured value is never read as an option A9: probe, check and cropdetect were annotated readOnlyHint, yet plan + overwrite let them replace any existing file, and every tool said destructiveHint: false. The server now runs the read-only tools with FFMPEG_SKILL_READ_ONLY=1, and apply_common refuses --plan under it after argparse (kind input, error.reason read_only_plan_refused), so neither the structured form nor any raw-argv spelling writes a file. The plan/overwrite property names stay in their inputSchema (3.x removes no property), with descriptions that say they are refused. Every writing tool is now destructiveHint: true. The CLI is unchanged. A10: build_argv emitted [flag, value] and bare positionals, so "-50%" was refused and an input "--plan=x" became an option. A value starting with '-' is now one --long-flag=value token (_common.option_args), and positionals starting with '-' follow '--'. render.py's stage commands use the same helper, so a project overlay or title starting with '-' works. Swept: every string option and positional of all 42 tools through build_argv into each tool's own parser (new test), and render's stage argv builders; batch.py recipes are raw argv and stay as given. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 1bb99302bcba9c71ebfd2c24209316b7cae7cb22) (cherry picked from commit 7924aca092d0ae5330b5e8c8f96ca2b9a378f9cc) * fix(silence): name the sentence-flag transcription path everywhere, and look the word up by binary search The audit found that #452's silence.py --transcribe with a sentence flag runs the speech engine (under --dry-run too) while the contract's capabilities, refusal reasons, model-download statement, PRIVACY.md, the plugin README and README named only --filler --transcribe; README's silence row and SKILL.md did not mention the 3.5.0/3.6.0 flags. silence_positions() also rescanned the whole transcript for every silence (11 s for three hours of speech). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit ad14bb441660fd12119151aff265f31225816db4) (cherry picked from commit 04e2bb549be62d2d349bd079aa11faa6e7c05c78) * test: name the render.py paths the output guard reads in the rebase list check Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 4e4461da09495fa7f78d8176839d9b3ea1a1b53a) * ci(security): pin every action to a commit, grant writes per job, fuzz the parsers Scorecard Pinned-Dependencies, Token-Permissions, Fuzzing and Security-Policy: - every uses: names the commit of the tag it replaced (Dependabot keeps both) - no workflow grants a write scope at the top level; checkouts that never push drop the token (persist-credentials: false) - the landing page loads the one Phosphor stylesheet it uses with SRI instead of the unpkg script that injected six at run time - ClusterFuzzLite with atheris targets for parse_time, parse_srt and the timeline XML reader; the contract suite runs each target on its seeds - SECURITY.md links the advisory form by its full URL Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit a2d33054c0479b0b7dba3c6e04f4ba7b9ae524e0) * docs: the Unreleased entries for the merged fixes; why ffmpeg's temps stay beside the output Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: compare the landing page's asset hosts by hostname, not by substring CodeQL (incomplete URL substring sanitization) on the SRI check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(readme): why the Claude plugin directory listing is on hold Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(overlay): label an untagged PiP beside a tagged main (FFmpeg 8.0 converted the main) The new PiP test failed on ubuntu-latest's FFmpeg 8.0.1: a BT.709-tagged main under an untagged PiP came back 88/103/192 instead of 80/110/192 at every opacity, because 8.0's overlay negotiates one colourspace and took the untagged side as BT.601. The untagged picture is now labelled whenever the other one is tagged, not only when one is converted; two untagged inputs are left as they are (the decided join/grid/broll/multicam plans are unchanged). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(fuzz): a negative time is a valid parse; say what the SRI and fuzz jobs cover parse_time is the grammar for offsets too (sync --offset -0.5), so the parse_time target no longer treats a negative result as a crash; seed added. CHANGELOG: SRI pins the icon stylesheet (its fonts load from it unpinned), and the PR fuzz job runs on the workflow's path filters. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(caption, look): whitespace-only markup, a split cue's alignment, --compare's own display conversion Three CodeRabbit findings on #469, each verified against the tree first: - reinsert_srt_markup anchored a tag pair that encloses no visible character like any other tag: closing after the character before, opening before the one after, so `A<font color=red> </font>B` coloured B. In a gap holding such a pair the tags now keep their source order (opening before the whitespace, closing after; whole at the next line's start when the whitespace became the break). Other tags are anchored as before. - carry_srt_markup took the cue's {\anN} from the first chunk that had one, so chunks before it were left unplaced. The alignment is now read from all chunks first. - look.py --compare built the display conversion from the first input's tags and ran it on both pictures: an SDR clip beside a PQ one was read as PQ, and an HDR clip given second was not tone-mapped. Each input now gets display_conversion() of its own probe; BT.709 SDR gets none. Swept: look.py is the only tool that converts two inputs for display (color.py has one input; grid.py's no-tone-map is decided); reinsert/carry_srt_markup have one caller (caption.py layout_cues). Tests: test_markup_around_whitespace_only_keeps_its_place, test_a_split_cue_takes_its_alignment_from_any_piece, test_look_compare_converts_each_input_by_its_own_tags (all fail on the parent commit). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 95b6ee8997b69dd09a874c14dbdaf01dee86bfcf) * fix(timecode): count the start label at the delivered rate, from a late picture's first frame; sync seeks early Three CodeRabbit findings on #469: - cut.first_frame_from (re-encode) and fit --method trim clamp the start to the picture's first frame before the grid lookup: a source whose video starts after the file's origin was stamped a frame before its picture (00:59:59:21 for a cut from 0.2 s of a picture starting at 0.5 s). sync's padded head still counts back. - output_timecode_args takes the delivered rate: the source label is read at the source's rate and written at the output's. fit passes --fps (or the rate cfr_args forces) and relabels whenever the rate changes, without copying the source's tmcd track (run_keeping_subtitles(relabel_timecode=True)); cut's re-encode passes the rate cfr_args forces on a VFR source; sync does not change the rate. render.py reaches frame.fps through fit, so it is covered. - sync --trim-second seeks SEEK_MARGIN early and drops the margin on the output side (on the drift-corrected clock), as cut.py's re-encode does. SEEK_MARGIN moves to _common.decision so both import it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit c54f59fbd527e31ebd5418fd14e4732bc97418dd) * fix(batch): a watch pass replaces only the files it delivered, checked per path where each is written The replace permission was a Boolean taken when the item started and handed to every step as --overwrite, so a hand-edited sidecar (.srt) or a file swapped in meanwhile was replaced. Now each step gets a snapshot of the files this item delivered (size, mtime, inode) and its own output check passes only those, still unchanged; deliver() compares the final path again under the output lock. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(timecode): round the start label once at the delivered rate; fit trim seeks early The source label and the offset to the first frame were rounded apart, so a rate change could land one frame off (4.8 frames written as 4). A delivered rate below 1 fps divided by a zero-length day; it now writes no label. fit --method trim seeked straight to its start before -i, which can lose the frames before a keyframe on a B-frame HEVC source; it now seeks SEEK_MARGIN early and drops the margin on the output side, as cut.py does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test(sync): read the first frame count FFmpeg 8's ffprobe prints for the stream On macOS and Windows (FFmpeg 8) csv output repeats the stream section, so the count read back as '92\n\n92'. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(batch): record each delivered file as it was when written, not when the item ends caption.py writes its SRT before the video finishes encoding. batch stated every reported path after the whole item completed, so an SRT corrected while the video encoded was recorded as batch's own and the next watch pass replaced it. Each writer now reports the identity its file has right after the write (runner.report_committed, from run()'s place step, place_output, write_srt and remember_written), and deliver() records the final path under the output lock. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Claude <noreply@anthropic.com> | 6 days ago | |
fix: one reader for caller text and JSON files; SRT cues keyed on timing lines (#483) Caller text files lose their BOM and end lines at \n; SRT cues are grouped by timing line and read only the first two times; caller JSON is refused as kind input when malformed, nested past 64 levels or of the wrong type (projects, plans, recipes, brands); a null brand block keeps its defaults. Findings A12-A20, A66, A73, A262, A268-A270 (#457, #465), plus siblings in join --list, cut --snap-source, metadata, report and the whisper readers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs | 6 days ago | |
fix: the audit's high findings (captions, timing, audio, colour, batch), MCP argument safety, and Scorecard hardening (#469) * Add OpenSSF Scorecard and MCP registry manifests Scorecard is pinned to the commits in ossf/scorecard's own workflow, and it publishes results and uploads SARIF. server.json and glama.json follow the current official schemas. npx ffmpeg-skill mcp starts the stdio server, which is the command the registry package argument names. Co-authored-by: Shogo Kajiyama <kajisho5@gmail.com> * Add join.py --copy (stream-copy join of matching clips) and cut.py --remove join.py --copy --transition none joins clips by the concat demuxer with -c copy when their first video and audio streams match (cut.py's stream signature, now in _common.probe), no clip hides edit-list pre-roll, and each clip's picture and sound end within 50 ms. A mismatch is refused by name (join_copy_mismatch). The join is measured beside the output -- video packet count and length against the clips' -- before it is renamed into place. cut.py --remove A-B,... is the inverse of --segments: the ranges are dropped and the rest is cut exactly as the equivalent --segments would be. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add output_identity, check.py --expect-identity and check.py --compare Every writing tool reports output_identity {size, mtime_ns} for the file it wrote; check.py --expect-identity SIZE:MTIME_NS FAILs an identity row when the file at the path is another one (an older file a failed run left, or a later run's), and check.py reports input_identity. SKILL.md's delivery check passes it. check.py --compare REF measures SSIM (mean and worst frame) and PSNR against a reference in one pass, WARN under a mean SSIM of 0.98. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add export.py --max-mb: a two-pass x264 encode that fits a size, measured The preset's x264 line takes the video bitrate the running time leaves after the audio (97% of the budget) in place of its CRF and runs in two passes; the written size is a file_size verification step, with a size_over_limit warning when the encoder overshot. Refused for non-x264 presets and --hw, with --crf, without a duration, and under 100 kb/s for the picture. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add export.py --preset copy --repair: remux a damaged file and count the damage Timestamps regenerated, corrupt packets dropped, demuxing past errors, negative timestamps shifted to zero; the input is only read. Both files are decoded once at -v warning and the timestamp warnings and other errors counted (repair: {before, after}); the repair step fails if either grew, and errors a remux cannot fix are a repair_incomplete warning. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --repair: judge the remux on timestamp warnings Decode errors are log lines whose count changes with the container (one full run logged 51 before and 53 after), so they are reported before and after but no longer fail the verification step; timestamp warnings, the part a remux rewrites, do. The test source is encoded single-threaded so its damage is the same every run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Label the join copy, identity, compare, max-mb and repair additions 2.24 2.23.0 shipped with check.py's flashes and first-frame rows; these land in the next minor. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address review: cover art, timeline times, copy-join guards, --remove edges, --max-mb - check.py --compare reads V:0 (not cover art), times the worst frame by its presentation time, counts frames from packets when nb_frames is missing, prints that note in plain text, runs under --dry-run when both files exist, and starts each main() with no leftover notes. - join.py --copy compares the first non-cover video stream, judges A/V alignment on each stream's end (start + length), includes colour matrix and range in the signature, refuses --sample-rate/--channels, fails the packet check when nothing was counted, and reports short and silent clips as the re-encoding join does. - cut.py --remove refuses any explicit --start and --snap beats, and clamps a range that starts before 0. - export.py --max-mb refuses nan and inf. - contract: join is conditional on audio; check notes name --compare. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Add a cut.py --remove demo; damage only video payload in the repair test The repair test overwrote bytes anywhere in the middle of the TS, which on ffmpeg 5.1.1 hit data its ffprobe could not open the file past; it now flips one byte in 40 video payload packets, never a TS header, the PAT/PMT or audio. The demo shows --remove 2-4 on the gallery's clip. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: refuse a different stream layout and a start offset; --compare pairs by index - The concat demuxer maps each clip's packets by stream index, so a clip whose streams are laid out differently (a cover picture between video and sound) is refused before the signatures are compared. - Picture and sound must start, as well as end, within 50 ms of each other in each clip. - check.py --compare re-times both pictures to their frame index before ssim/psnr, which otherwise pair frames by timestamp. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --expect-identity: grade an existing file under --dry-run too The identity was skipped for every dry run; it is skipped only for a pending placeholder now, as the content rows are. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy checks starts without stream lengths; --expect-identity reads twice - Matroska/WebM state no stream length, so the A/V guard compared nothing there; the starts are now compared whenever both streams exist and the ends only where both lengths are known. - check.py reads the input's identity before measuring and again after: a file replaced mid-check FAILs, and a path that holds no regular file is a FAIL row rather than no row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py: read the identity after --compare; note a mid-check change without --expect-identity The second identity read now follows every measurement, --compare's pass included, so a replacement during that pass is caught. Without --expect-identity no row is added (the row set is unchanged), but a file that changed while it was being checked is named in notes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Route the 2.24 features from SKILL.md; make the repair test open on every ffprobe SKILL.md: routing rows for export --max-mb, cut --remove, join --copy, export --preset copy --repair and check --compare; the report format moved above the routing table (inside the first ~5000 tokens kept after compaction); the --dry-run and VFR-phone rows dropped (covered by the shared-flag line and a gotcha). Checked with a blind behaviour test, original vs trimmed, on 10 prompts: trimmed 60/60, original 56/60. test_delivery: the repair source is one TS appended to itself byte for byte, not bytes flipped inside packets; ffprobe 5.1.1 could not open the latter. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Repair test: fall back to MPEG-PS where ffprobe cannot read MPEG-TS The static ffprobe 5.1.1 build segfaults on every MPEG-TS file (the same reason test_contract gates its TS cut test), so the appended-recording source is written as MPEG-PS there; both containers are what a recorder appends to, and the assertions run unchanged on every runner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Repair test: give the MPEG-PS fallback damage every ffmpeg reports On the static 5.1.1 build an appended program stream decoded without a timestamp warning, so nothing was counted before the repair. The PS copy now also has bytes flipped inside packs of its middle half; the MPEG-2 decoder reports those as decode errors on every version. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Docs: multicam energy switch is a loudness proxy; why 啊 is not a default filler; link ffmpeg-full PATH setup - README / gotchas: `--switch energy` picks the loudest camera, not a recognised speaker; it needs each speaker on their own camera's mic. - design-decisions / scripts.md: Chinese 啊 is out of the zh list for the same reason as `like` (a sentence-final particle far more often than a hesitation); `--filler-extra 啊` adds it. - README HDR section: the zscale note links to Install for putting the keg-only ffmpeg-full on PATH. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Drop color.py --tonemap bt2390 (never worked); bring README in line with the tree color.py: FFmpeg's CPU tonemap filter has no BT.2390 curve in any version (only tonemap_opencl and libplacebo do), so --tonemap bt2390 always failed. The choices are now the filter's own curves; a test checks every choice against the ffmpeg under test and that bt2390 is refused. README, checked against the code: - CI runs on pull requests and pushes to main, not every push; the FFmpeg list includes the Ubuntu 26.04 / 8.0.1 job. - templates include audiogram; --template all renders the seven video destinations, not every template. - doctor's fonts.scripts covers bn, ta, lo and latin too. - --crf was removed except on export.py; --codec/--quality are not on export.py, render.py or batch.py. - output_schema names output_identity (every tool but probe). tests: the Windows skip for the POSIX shim was on the wrong test since #269; it is back on test_dry_run_never_runs_ffmpeg_and_writes_nothing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --compare: pair only frames both files have; RGB and rotated inputs; empty --expect-identity - ssim/psnr ran with eof_action=repeat, so a shorter output's last frame was measured against the rest of the reference (a mean of 0.90 for 52 identical frames, the worst frame timed past the file's end): shortest=1. - An RGB file printed "SSIM R:"/"PSNR r:", which the parser missed ("could not measure"), and the value depended on argument order: both inputs go through format=yuv444p and the parser takes any channel letter. - A clip stored sideways was compared by its stored size although ffmpeg auto-rotates: the size check uses the displayed size. - --expect-identity "" (an unset shell variable) skipped the identity row; it is now malformed (kind input), parsed before any measurement. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: measure stream spans from packets; VP9; --codec; dropped tracks. cut.py --remove outside the file - The A/V end check read stated stream lengths: Matroska/WebM state none, so a clip whose sound stopped 1.5 s early joined as verified; MPEG-TS states an estimate, so every AAC .ts was refused with wrong numbers. Each stream's span now comes from its first and last packet. - VP8/VP9 carry no extradata in WebM: identical clips were refused as "codec configuration differs" (and cut.py --remove/--segments fell back to an x264 re-encode into .webm, which ffmpeg refuses). - --copy with --codec, --quality or --hw ignored them silently; refused now. - A copy join drops subtitle/data tracks: dropped_non_av_streams says so. - cut.py --remove 15-20 on a 4 s file kept a range past the end and removed nothing; a range outside the file is refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --max-mb with gif / huge limits; --repair damage counting; render.py --stop-after keeps its output - --preset gif returned before --max-mb was applied, so a 0.7 MB GIF passed a 1 KB limit as verified: refused. A limit so large it constrains nothing is refused as input instead of failing inside libx264. - --repair counted "Packet corrupt (... dts = ...)" as a timestamp warning and "Last message repeated N times" as a decode error, so a remux that made timestamps worse could pass: repeat+warning prints each occurrence and corrupt-packet lines count as damage. - render.py --stop-after (before export) without --work removed the work directory at exit, deleting the file the result named; it is kept. - docs/contract.md lists size_over_limit and repair_incomplete. - CHANGELOG: the fixes found reviewing 2.24.0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: refuse when a stream span cannot be measured; refuse any --quality - _stream_span skipped a packet without a time and measured the rest, which could invent (or hide) an A/V gap; it returns None, and join_copy refuses the copy when either span is unmeasured instead of skipping the clip. - --quality was refused only when it differed from the default, so --quality 18 passed silently; any --quality given is refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: read packet spans as JSON with a dts fallback; rotation fixture for FFmpeg 5.x The ffmpeg 5.1.1 job refused every copy join as "could not measure": a row of the csv packet listing did not parse as a time there. The spans are read from ffprobe's JSON (only packet entries), and a packet without a pts falls back to its dts; a packet with neither still makes the span unmeasured. -display_rotation is FFmpeg 6.0+; the rotated --compare fixture falls back to the rotate tag, and the test asserts the rotation was written. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: read a long clip's packet span at its head and tail only; test the dts fallback _stream_span listed every packet of the first video and audio stream of each clip to keep two numbers. A clip longer than 30 s is now read at its first 64 packets (past B-frame reordering) and its last 10 s; a test checks that this gives the same span as reading every packet. A packet with no pts is placed by its dts, now pinned by a test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * join.py --copy: seek a long clip's tail from its own start time -read_intervals takes timestamps, so the tail read started at duration - 10 s on the absolute clock. A clip whose clock starts late (an MPEG-TS hours in) read from its first packet instead, and one starting below -10 s read nothing and was refused. The seek is start_time + duration - 10 s; a test checks a clip starting at 100 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Restore CHANGELOG.md to main after the 2.24.1 release bump The merge kept the 2.24.1 Fixed entries under Unreleased as well as in the released section. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0: require FFmpeg 6.1+ and Python 3.10+; apply the cut.py defaults decided for 3.0 Work in progress (docs, the MCP snapshot and the tests that pin the 2.x cut defaults follow). - FFmpeg 5.x (deprecated in 2.14.0) and Python 3.9 (2.10.9) move from `deprecated` to `removed`; requirements are >=6.1 and >=3.10. Every tool refuses an ffmpeg/ffprobe it reads as older than 6.1 (kind missing_tool, reason ffmpeg_unsupported); doctor lists `unsupported` and reports ok false; the installer skips Python < 3.10 and prints the FFmpeg line. drawtext's 5.x/6.0 spellings and CI's 5.1.1 static job are gone; CI's Python floor is 3.10. - cut.py: edit list kept on a single .mp4/.mov copy (--no-edit-list for 2.x), SDR HEVC stays HEVC (--no-keep-hevc), --vfr-guard sampled; precision is the least exact segment's and keyframe_snapped equals start_snapped. contract_version 2.0. - package.json / plugin.json / roadmap 3.0.0 (a major is bumped by hand), CHANGELOG 3.0.0, the 2.14.0 release date and the 3.0.0 window waiver in the deprecation-window test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * 3.0.0 docs: work in progress Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * v3.0: tests for the new cut defaults, sampled guard on seeked TS windows, structured ffmpeg_unsupported refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * design-decisions: point stale test names at their 3.0 successors Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: report the source's start timecode (video stream, tmcd, then format) Requested by nle-bridge-skill (roadmap step 12) for OTIO/EDL/FCP7 source timecodes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test_asr: pick the engine call out of the fake's calls (the 6.1 floor reads ffprobe -version first) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: timecode minutes/seconds 00-59; ffprobe floor in doctor; git builds refused only below 6; README edit-list wording; sampled guard's keyframe start in the contract Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: doctor marks every tool unusable below the floor; an old ffprobe is named even when ffmpeg is old too Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * ci: spread the suite over three machines per platform (run_parallel.py --part I/N) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * run_parallel --part: bound N by the work items; test that every shard lands in exactly one part Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Failures say why: ffmpeg error reasons and hints, no orphan ffmpeg after SIGKILL, no blind re-encode, silent template renders, ASR missing_tool Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review: one encoder_missing entry, asr_engine_missing under error.reason, render dry run reads sources and counts audio.replace Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: a talking-head reframe demo from AI-generated footage, on the landing page Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * silence --filler --transcribe: prime Whisper with the filler list; filler_en demo from AI-generated English speech Whisper leaves uh/um out of its text unless its context shows them, so --filler found none on clean speech. The filler list is passed as the initial prompt to whisper.cpp, faster-whisper and openai-whisper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pages: the two AI-footage demos as MP4 with sound, built with grid/graphics/join/look Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * demos: the gallery intro names the two AI-footage demos Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs/demos.md: keep render_project's --export-timeline line (a partial --docs rebuild dropped it) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pages: the reframe label and aria-label give the source's real size; the lede names pages_videos.py Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: SKILL.md restructured, no behavior change SKILL.md keeps Quick start, Workflow 0-8, ask/assume, the does-not-decide line, the report template, the request table (one command per row) and the one-line gotchas. The long forms moved unchanged to references/flags.md, workflow.md, report.md, requests.md and timestamps.md, shipped via package.json files. Description cut to three sentences (924 -> 399 chars); trigger set 55/55 before and after. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: restore three SKILL.md rules the restructure cut; fix moved-text references - step 6 verifies from the writer's --json probe or probe.py (a routing check on agent_prompts_24 lost the verify step on ml1 without it) - step 7 allows --overwrite for a file the user asked to replace - the 60 s highlight row names cut.py --segments as the step that renders - references/flags.md points at SKILL.md's request table, not "the table below" - references/gotchas.md index links are same-file anchors - evals/results/routing-skill-md-restructure.json: old 93/99, new 94/99 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: step 5 keeps --expect-identity; highlight row and --from wording - step 5 runs check.py with --expect-identity S:M (the writer's output_identity), as before the restructure - the highlight row says picks.txt's lines join with commas for --segments - timestamps.md: only broll.py --from takes a time; render.py --from takes a stage name Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix #393-#397: lean builds, tofu emoji, SRT as --text, dropped audio tracks - #393 caption.py / graphics.py ask the build for subtitles/ass first (dry run included) and refuse with reason filter_missing; the path is a named option (filename=) so ffmpeg itself says "No such filter" - #394 look.py without drawtext draws the sheet without timecodes (fallback look_no_drawtext) - #395 doctor fonts.emoji.libass_glyph tells a glyph from the .notdef box; caption.py refuses to burn empty boxes - #396 an SRT given to --text / render --cues is refused with a pointer to --srt - #397 cut.py reports audio tracks it did not write (dropped_audio_streams) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * #395: an auto caption leaves undrawable emoji out with a warning; CodeRabbit fixes - caption.py --emoji auto with no glyph (libass_glyph false) takes mode none, as on a machine with no emoji face, and warns emoji_not_drawn instead of refusing (CI's stock Ubuntu and macOS runners are such machines); --emoji mono still refuses, with or without assets (mono never uses them) - the glyph probe searches the caller's --fonts-dir / --font-file dir and is cached per fonts dir - graphics.py checks the ass filter before writing its ASS sidecar - caption.py's early check asks for ass under --animate / --karaoke - a ';' drop-frame cue line in --text is refused instead of burned as text - docs: the assets exception for libass_glyph, cut.py's default audio pick Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: qualify the emoji none mode and the one-audio-track gotcha CodeRabbit follow-ups on #399: mode none applies when no emoji assets resolve (png with assets); the one-track rule names the editing tools and notes that color.py --strip-dovi copies every track. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Escape line breaks and control characters from arguments in messages (#400) die(), refuse() and every diagnostic record pass their message through neutralize(): a command-line argument carrying a line break or a C0/C1 control, line separator or bidi mark is written escaped wherever the message prints it (as given, resolved, or as a file name), and any such character left is escaped too. Clean paths and a tool's own line breaks are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Report audio tracks every tool leaves out; grid/scenes sheets without drawtext - emit() compares the first input's audio tracks with the written file's and warns audio_tracks_dropped (with a notes line) when the output kept sound but fewer tracks; --no-audio / picture-only outputs report nothing; cut.py's own report is not doubled - grid.py --label auto and scenes.py --sheet on a build without drawtext draw the picture without the text (fallbacks grid_no_drawtext, scenes_sheet_no_drawtext), as look.py does since #394 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #405: escape --opt=value and notes; no track report across several sound sources - neutralize() also matches the value of --opt=value, which tools print back alone - a tool's own notes (join.py's duplicate inputs) are escaped as messages are - audio_tracks_dropped is skipped when two or more inputs with sound were probed (grid --audio-from, a mix): which source's tracks the output keeps is the tool's choice - scenes.py --json-brief goes through emit(), so scenes_sheet_no_drawtext reaches it - test shims no longer write a file named after `ffmpeg -filters` Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: README and the Pages docs layout carry the v3 brand - README opens with the v3 horizontal logo (light and dark) instead of the old app icon and wordmark, links the website, and shows the v3 card (the same image as the social preview, brand/og.png) linking to it - the docs layout (README, Demos, Contract pages) uses the v3 favicon, apple-touch icon and og:image instead of the old app icon and assets/social-preview.png, and puts the symbol in its nav Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: drop the 3.1.3 entries the merge left under Unreleased Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Side files and paths get the output guards (#408) - waveform.py's vis/titled intermediates are hidden, pid-named siblings: the fixed <out>_vis / <out>_titled names were removed whoever had written them (the input included) - --plan FILE, graphics' ASS file, caption --write-srt/--write-ass, verify --report and render's <stem>_pack.md go through guard_sidecar (never the input, never an existing file without --overwrite, a missing directory refused before anything is written) - single-image outputs pass -update 1 and still inputs -pattern_type none, so a % in the name is the file itself, not an image2 sequence - on POSIX a backslash is escaped in filter paths and kept in concat lists (Windows-only conversion to /); edge whitespace in a filter path is escaped - a one-letter x:name.mp4 off Windows is refused as a URL with the ./name hint Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * HDR10 static metadata reaches libx265 and SVT-AV1, and its loss is measured (#401) - probe reads the mastering display and content light level (stream, else the first frame) into video.hdr10_static - the HEVC and AV1 HDR lines pass them as master-display/max-cll and mastering-display/content-light: FFmpeg before 7.0 does not hand them to either encoder - emit() warns hdr10_metadata_dropped when a PQ output still lacks what a PQ source had (prores_ks, libaom) - note() escapes its line as die()/warn() do; batch.py's unverified names are shown() - the audio_tracks_dropped contract text states the several-sources exception Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #412: zero luminance, per-part frame read, V:0, SVT precision, several pictures - a mastering display's 0 cd/m2 minimum is a value (_ratio), not a missing field - the first frame is read when either part is missing from the stream; each part is taken from wherever it is found first; the frame read selects V:0 (not a cover picture) - SVT-AV1 values keep five decimals - hdr10_metadata_dropped is not compared when two or more inputs have a picture (sync --trim-second, a join): the first input need not be the output's source - tests compare the values the re-encodes carry and compute the ProRes test's expected loss Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #412: a still is not a picture source; read the frame when a stream record is incomplete Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix the audit G colour paths (#410) - color.py --lut: lut3d between explicit BT.709 legs through gbrpf32le (an identity LUT shifted every colour); the --lut-strength blend is made in RGB. - join/multicam/grid/broll/overlay --video: a full-range clip in a plain format (VP9/AV1 full range, 10-bit pc) is taken to limited range before it is mixed (range_chain); a converted RGB layer's main likewise (layer_base_chain: graphics boxes/bars, emoji, overlay --image). - drawn_chain on FFmpeg < 7.1: such a picture is taken to limited range before ff_draw / rgb_island draw on it (pad black 101010, white EBEBEB, 10-bit chroma oversaturated). 7.1+ plans are unchanged. - look.py, scenes.py --sheet, export.py --preset gif: an untagged source is labelled BT.709 before the conversion to RGB. - color.py --correct: no colortemperature term at the default 6500 K. - grid.py: a 1x1 grid is refused up front. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Square pixels everywhere that scales; wrap graphics titles and overlay text; fix look/scenes sheets, grid labels, freeze --at, broll odd sizes, EXIF stills; SMPTE start-timecode hint (#411) - probe reports sample_aspect_ratio; display_size()/square_pixels() in _common.probe are the one place a non-square-pixel picture is squared at its display aspect (height kept), used by every tool that sizes, crops, places or draws (fit, export, crop, join, broll, multicam, insert, sequence, grid, look, scenes --sheet, caption, overlay, graphics, redact, proxy, cropdetect, straighten, waveform --image, the render timeline frame). Re-encode-only tools keep the stored size and SAR. Recorded in docs/design-decisions.md. - graphics title/lower-third/chapter/bug wrap with the shared wrapped() helper; the lower-third bar widens to the widest line. overlay --text wraps to the frame and says so in notes. - look's contact sheet stamps from -copyts -start_at_zero (MPEG-TS was a seek late); scenes --sheet selects each scene start by time (VFR); look --compare letterboxes. - grid labels use a font that covers their script, and a long label is shortened from the end. - freeze --at holds the frame shown at T. - An hh:mm:ss:ff time past the end of a timecoded source gets a hint naming the start timecode. - broll scales an odd-sized A-roll to even up front (odd_dimensions_scaled_even). - probe reads an EXIF-rotated still's rotation off its decoded frame. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: match main after the 3.1.4 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(timing): keep picture and sound on one clock in silence, sync, multicam, freeze, speedramp, join, broll, reverse and loop (#409) Every tool in the audit rebased a stream to its own first sample, or cut each stream on its own grid, and lost the A/V offset while reporting verified: true. Each now does what cut.py --segments --accurate does: both streams cut on the same frame-time bounds, shifted by one constant, a late sound padded with silence. Shared helpers in _common/decision.py: stream_starts, frame_rate, grid_at_or_after (cut.py now uses it), av_trim, av_clip_length. - silence.py renders through cut.join_from_source (no declick); bounds snap to the nearest frame time and --timemap is built from them. - decode_pcm_mono pads a late sound (aresample first_pts=0); sync.py and multicam.py place the second's sound on its own clock before delaying it. - freeze.py, speedramp.py cut on frame times; join.py (#338) and broll.py shift the sound by the picture's start. - reverse.py turns both streams about one frame-time end. - loop.py repeats through the concat demuxer instead of -stream_loop. Tests: OneClockTests (test_editing), test_a_late_sound_is_measured_and_placed_on_its_files_clock (test_analysis); two existing tests updated for the new bounds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * render pack guard only where the folder exists; pin the pack table refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG for #410 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: check the MP4 header instead of calling file(1), which the Debian container lacks Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CodeRabbit on #413: guard every input and output the side files can collide with - write_plan refuses a plan path that is any bound input (side inputs included) or the promised output, default names too - caption's claim() records which artifact holds each path: a different artifact claiming it is refused (output_path_taken); every --srt track is a guarded input - graphics' ASS guard covers --brand/--font-file; verify --report the collected files; render's pack table every existing file argument, and a missing folder at write time is a structured failure - look's frame reads pass still_input_args(); the % check looks at the whole path - waveform removes the title stage's _gfx.ass when this run created it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse hard-linked side files and allow a report inside a new --out (#408) - caption.py claim(): a second name for an already-claimed file (a hard link) is the same path, so --write-srt/--write-ass naming two links to one file are refused before either is written, even under --overwrite. - verify.py: create --out before guarding --report, so a report inside a new output folder is accepted. - tests: plan input compared byte for byte; hard-link and new --out cases. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Protect caption's --brand file from its side-file writes (#408) A --write-srt/--write-ass naming the --brand JSON (or a hard link to it) is refused as output_is_input, --overwrite included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Pass -pattern_type only to image2 stills; never reuse a leftover temp name (#408) - still_input_args(): `-pattern_type none` only for an image2 extension. A video or GIF with a % in its name was given it too, and ffmpeg refused the input ("Option pattern_type not found"), so look failed on `take%d.mp4`. probe uses the same helper. - sibling_temp(): a hidden pid-named file that already exists (a leftover of an earlier run with the same pid) is skipped, not reused and then removed. - test: look on a video named with a %. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge origin/main (v3.1.5) into the branch Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * freeze, reverse: clone frames before setpts so the hold survives FFmpeg 7.x From FFmpeg 7.0, setpts sets its output's frame rate to unknown (1/0) and every frame's duration to 0. tpad derives a clone's timestamp step from the link's frame rate, so a tpad placed after av_trim's setpts gave every clone the same timestamp. Those clones were then dropped: - freeze --at: the tpad start clone of the tail added nothing (3 s came out 1.97 s, and the sound ran a second past the picture) - reverse: the stop clone that holds a short picture's last frame added nothing (408 ms off for a sound 0.6 s longer than its picture) The fix uses one graph on every FFmpeg version: - freeze: the held frame is its own concat segment. It is one frame cut on the input's clock, cloned with tpad's stop clone, then shifted. - reverse: tpad runs before av_trim's trim/setpts, as cut.py already does. - Both video chains end in frame_durations(), an fps filter at the source's r_frame_rate. It moves no on-grid frame and gives frames back the durations setpts cleared. Without it, 7.x wrote the last frame with no duration and the picture read one frame short (3.967 s for 120 frames). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address CodeRabbit's review of #416 freeze: hold the last frame when no frame lies at or after --at (no --at, the clip's end), and take the held frame as the first frame after the cut (trim=end_frame=1) so a variable-frame-rate gap still yields one. silence: refuse -o equal to the input before the chunked join, and report keep, kept_duration, removed_seconds and --edl from the rendered (frame-snapped) ranges, EDL frame times truncated to the millisecond. insert runs the square-pixel scale it noted; cropdetect, sequence and waveform --image note theirs; the still-rotation probe passes -pattern_type none. straighten --fit crop scales by |cos| + max(W/H, H/W)|sin|. speedramp refuses a segment holding no frame. graphics places PNG emoji per wrapped line, notes a title taller than the frame and caps the lower-third bar at the margins; overlay --text notes text taller than the frame or too near an edge to wrap; grid shortens a label in a one-em cell; broll scales a 1-px A-roll to 2; the SMPTE hint uses the timecode of the input a time is in. Each with a regression test that fails before the fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * broll: B's sound lasts the whole cutaway when --from falls between frames The sound started at B's first frame at or after --from but ended at --from + length, so a --from between frames left the end of the cutaway without B's sound (23 ms at 30 fps with --audio b). It now ends one cutaway length after its own start. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop: refuse an output that is its input before the concat list The concat list is the only -i of the non-boomerang path, so run()'s own input/output check never saw the source: with --overwrite, -o naming the input replaced it with its own loop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * color --lut: keep a BT.601 source's matrix on the ProRes path --codec prores keeps a non-BT.709 source's matrix and tags instead of converting it, but the LUT legs were pinned to BT.709, so a BT.601 source's samples were read and written with the wrong coefficients. The legs now use the source's matrix on that path; every other path is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Subtitles, transcription, timeline export, Windows fonts and contract drift (#414) - silence.py: --margin kept around silences only, not filler spans - asr: transcripts on the file's clock; no_speech for silent word timings; caption.py --transcribe refuses a source without audio - render --export-timeline: source start timecode, per-clip EDL rate, out clamped to the media end, bad chapters refused - caption.py --mode mux converts existing tracks per container; ffmpeg gets the cues parse_srt read - fonts: an explicit Windows --font resolves through the family table first - contract: publishes the output keys and ASR error reasons the tools emit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Dry-run honesty, process handling, faststart and batch/render orchestration (#415) - render: a file a stage of this run wrote counts as the run's own - metadata.py: MP4/MOV output keeps +faststart - .webm sources default to an .mp4 output on re-encoding tools - batch.py: per-item work dirs, steps without {out}, write-nothing recipes report output null, case-insensitive name collisions - the first failure is reported instead of re-encoding up to three times - timeouts/interrupts after the deliverable name the kept file - probes honour a --timeout below 120 s Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop.py --crossfade: seamless repeats, and every loop measures its seam (#422) --crossfade S fades the clip's last S seconds into its first S (xfade for the picture; afade + amix for the sound, since acrossfade writes nothing when its first input is exactly the fade long). One repeat is the clip minus S, written losslessly to a temporary file and repeated through the concat list, so the output keeps one lossy encode. --times 1 writes one seamless loop file. Every non-boomerang run reports `seam` (picture step across the first seam against the typical frame-to-frame step, sound level step) and checks the written length against the request (`expected_duration`, 1.5 frames) and its streams. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py: DNxHR editing-master presets, and masters read their codec back (#419) dnxhr-lb / -sq / -hq (8-bit 4:2:2), dnxhr-hqx (10-bit 4:2:2) and dnxhr-444 (10-bit 4:4:4), .mov with PCM audio at the source's size and rate, handled like prores (source colour tags kept, no BT.709 matrix conversion, no HDR-to-SDR note). prores and the dnxhr presets now add a `codec` verification step comparing the written codec, profile and pixel format with the preset's; a VideoToolbox line under --hw is not compared. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loudness.py --versions: one master, one measured file per loudness target (#421) Named targets streaming (-14/-1), ebu-r128 (-23/-1) and broadcast-jp (-24/-1, ARIB TR-B32), or LUFS/TP pairs. Each version is a full two-pass normalisation from the master (pass 1 per target, since loudnorm's target_offset depends on it), measured after; the document lists every version, its verification steps carry a `version` key, and `verified` holds only when every version met its own target. The single-target path is the same code, moved into _normalize(). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py: alpha masters that read their alpha back (#417) prores4444 / prores4444xq (16-bit alpha), qtrle, vp9-alpha (WebM, Opus, decoded with libvpx-vp9 to read the alpha) and hap-alpha (needs snappy; sizes in multiples of 4). A source without alpha is refused; a missing encoder is kind missing_tool / encoder_missing. --alpha-mode straight|premultiplied (premultiply filter) is reported as alpha_mode. The `alpha` verification step compares the mean alpha of the first written frame with the source's. PNG/EXR sequences are not included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * export.py --preset key-fill: a switcher's fill and key from one alpha source (#418) One decode split into the fill (straight, or premultiplied with --alpha-mode) and the key (alphaextract: alpha 255 is video white), both ProRes 422 HQ with the source's start timecode, written by one ffmpeg run so they hold the same frames. Verified after: key_fill_pair (size, frame count, rate, length equal) and key_level (the key's mean level against the source's alpha); both files' SHA-256 are reported. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * look.py --guides: broadcast safe-area outlines on the preview (#423) action (93 %) and title (90 %) safe per SMPTE ST 2046-1, and 4:3 (the centre 4:3 window of a wider frame, with SD's 90 % action / 80 % title inside it), drawn as outlines in fractions of the frame so a contact sheet's tiles get them too. look.py writes PNGs only, so a guide never reaches a delivery; the result's `guides` lists what was drawn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the filler keep range is the rendered one since #416 (#414) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: refresh the design-decisions contents list (#414) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the #414 decisions ship in 3.1.7 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge origin/main (v3.1.6) into the branch; CHANGELOG from main Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: the fake whisper-cli fixture runs on Windows too (#414) No symlinks or shebangs there: the real ffmpeg/ffprobe folders go on PATH and the fake engine is a .cmd wrapper, as test_accel's fake ffmpeg is. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs: the #415 decisions ship in 3.1.8 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #427 review: margin input, filler message, SRT grammar, mux languages, EIA-608, timecodes - silence.py: --margin nan/inf/negative refused; "output would equal the input" only when the render removes nothing (filler included) - caption.py: the shared-grammar check reads timing lines only; the .webm mux check reads the --audio-stream it copies; tracks report the language/title the output holds - runner: eia_608 is a text subtitle (mov_text in .mp4, copied in .mov) - asr: a whisper segment without `text` is unreadable output, not no_speech - timeline: a start timecode that cannot exist at the source rate is not read; EDL M2 speed at each clip's own rate - contract: dropped_non_av_streams description covers both cases; contract.md drops a wrong "since 3.1.6" Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: skip a start timecode whose frames field the picture's rate cannot reach (#427 review) A 25 fps video tag of 01:00:00:29 was picked before a tmcd track's valid 01:00:00:00, and the export then read the file as having no timecode. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * probe: validate a timecode's frames field at avg_frame_rate first, as probe()'s fps does (#427 review) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG from main (v3.1.7) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #428 review: stopped-child exit codes, batch delivery, copy-preset extension - render.py: an executed plan and a pack destination that were stopped pass the child's exit code (124, 130/143) and kind on; an export stage stopped in --normalize hands the delivery it kept to render's `output`. - export.py: --preset copy keeps the source's extension (a .webm stays .webm); the export counts as delivered as soon as it is written, so a stop in the output probe or loudness read says it is kept. - batch.py: a delivered intermediate is named by its own container (.mp4); delivery holds the output lock (_OutputLock gains on_locked so a worker can raise instead of exit) and copies across filesystems through a sibling temp; the item directory and step files no longer carry the stem, so a stem near the 255-byte limit runs. - Docs: the --timeout 0 and probe-ceiling statements name the reads that keep 120 s; contract.md lists plans and pack destinations among stopped children. - The hung-ffmpeg shim writes a partial file so the test checks its removal. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #428 review: dry-run output check, signal exit codes, batch contract - batch.py --dry-run refuses an existing final path that a recipe ending in a read-only step would deliver to, as the real run does. - check.py --dry-run on a file an earlier dry-run step would write grades nothing (checks: [], a note) instead of crashing on a 0 height. - runner.child_exit_code(): a child killed by a signal (returncode -9) is passed on as 128 + the signal by export.py --normalize, render.py's executed plan and its pack, not as a wrapped 247. - The contract's batch outputs and docs/contract.md say a recipe in which no step names {out} writes no artifact (rows carry output null). - Tests compare kept-output paths resolved (macOS /var is /private/var). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse a batch output_dir that is the input folder (#428 review) An output_dir like "sub/.." named the input folder under another name: the input scan did not exclude it, and with suffix "" each final path was the source, which --overwrite replaced. The output dir is now resolved and a batch whose output_dir is the input folder is refused before anything runs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Contract: a render-project batch recipe always writes its artifact (#428 review) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Keep batch rows' output a string; mark a write-free recipe with written: false (#428 review) A recipe in which no step names {out} reported output: null, changing the type of a --json key within 3.x, which the stability table forbids. output stays the final-path string; such rows add written: false, and the cache recognises those entries as claiming no file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * batch: key the cache on the files a step names; fix the QC note (#428 review) A QC pass of check.py --compare REF was served from the cache after REF changed or was removed: only the source and the recipe text were keyed. Every step argument that names a file now contributes that file's content (or its absence) to the key, read on each pass. The note for a recipe that writes nothing said output was null; it says written: false now. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Take main's CHANGELOG after the v3.1.8 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: an encoder capability name may contain a hyphen (libvpx-vp9) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * loop.py --crossfade: restate the frame rate before xfade (FFmpeg 7.x) FFmpeg 7.1's setpts marks the frame rate unknown (1/0) and xfade then refuses the graph; each trimmed branch now restates its rate. Found by the debian-trixie (ffmpeg 7.1) CI job. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #430 review: key-fill outputs, alpha checks, --versions edges, docs and a demo export.py - key-fill no longer inherits the -map 0:V:0/0:a:0 added for sources with subtitle/data streams (the fill got the unfiltered picture and the sound twice); the dropped tracks are reported in non_av_streams (reason key_fill). - key-fill guards the fill like the key (input under another spelling, an existing fill without --overwrite), stages an existing fill through a temp file, compares the video streams' lengths, and refuses --max-mb. - Alpha presets: a VP9 source is decoded with libvpx-vp9 (the native decoder drops the alpha); a dry run decides on the pixel format and the VP9 alpha_mode tag and decodes nothing; alpha (and key level) is compared over every frame (mean and most transparent frame), not the first; --hw leaves prores4444/xq and key-fill on prores_ks with a note. loudness.py: an explicitly empty --versions is refused; --versions/-o are checked before measuring. write_plan refuses a --plan path that any planned command writes (a version file). look.py: --compare --guides 4:3 checks the second clip's aspect too. loop.py: --crossfade counts frames after the fps= conform and keeps the unit's rate for a variable-rate source (the file missed its own length). Docs: README, SKILL.md (rows extended, no lines added), docs/contract.md, references/scripts.md, design-decisions; loop_crossfade demo. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Address the #430 review: loudness, look, demos, loop, key-fill, alpha - loudness --versions skips the extra default-target measurement - look --compare draws tone map and guide boxes before pad - demos page recorder no longer drops commands after side-file checks - loop refuses nan/inf crossfade; conformed frame count uses container duration - export key-fill stages every fill and removes it on a failed encode - export compares alpha and key level frame by frame when counts match Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix the demo page test off Linux; address the #430 follow-up review - demos/build.py: resolve DEMO_OUT before shortening paths (a Windows 8.3 temp name and macOS's /var symlink left absolute paths in the page) - tests: the --compare guides test checks the stamp only where drawtext exists - docs: the alpha check compares the mean of a 64x36 copy of each frame Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Merge main (v3.2.0) into the working branch Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release: find a squash commit's PR by its SHA when the subject does not end with (#N) 3.2.3 shipped three features as a patch: the squash subject ended "(#433, carries #432 and #434)", so the resolver matched no PR and never read the feature label. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * source_carry: every result says what the output lost from its source (#401) emit() compares the first input's subtitle tracks, chapters, cover picture, surround channels, creation_time, audio language and start timecode with the written file's. A lost subtitle track, chapter list, cover picture or surround layout is a source_items_dropped warning unless the tool already reports it; tags are listed only. A report, not a check: verified and the written file are unchanged, and a source with none of the items gets no key. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix caption.py --srt --position (#436) and odd-sized JPEGs through the BT.709 conversion (#437) - caption.py: libass force_style reads Alignment in the legacy SSA numbering; the SRT burn passed the numpad value, so only the bottom row landed right. - _common.color: the colorspace filter refuses an odd frame; crop the last row/column before it, and let the -vf merge recognise the prefix. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CHANGELOG: the mcp subcommand, the registry manifest and the Scorecard workflow Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review of #435: keep an odd frame's own size through the BT.709 conversion; stop the release on a failed PR lookup - bt709_matrix_filter: pad to even, convert, crop back (exact). The first version cropped the edge off, which shifted a caller's rectangle that ends on the last column and lost a row/column without saying so. - resolve_version: a failed commit->PR request raises instead of returning 'no PR', which was the fallback that turned a feature into a patch. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sweep a reported defect across the tools before fixing it: tests/sweep.py, the sibling-sweep skill, AGENTS.md and the PR template The class behind #437 (an odd frame reaching a step that refuses it) also breaks caption.py and stabilize.py on an odd-sized video; filed as #438. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Fix odd-sized clips in caption.py and stabilize.py (#438) The runner's even-size retry appended nothing the encoder could see: its scale went at the head of -vf and a later filter made the frame odd again. It now goes at the end. stabilize.py scales an odd frame to even before both vid.stab passes and records odd_dimensions_scaled_even. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin installs in Codex and Gemini CLI find the skill: skills/ffmpeg-skill/SKILL.md (generated), gemini-extension.json, .cursor-plugin/plugin.json (#406) Measured: Codex 0.167 installs the plugin from .claude-plugin/marketplace.json but lists no skill for a SKILL.md at the plugin root, with or without a manifest skills field; skills/<name>/SKILL.md is listed. Gemini CLI 0.68 validates the extension and discovers the same path. Claude Code counts one skill with both files present. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Review of #435: bound each sweep child, refuse a failed probe, write the caption cue in its time format, and say what matching rows do not prove Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * CONTRIBUTING: Smithery's publish commands no longer exist (its API and CLI stopped on 2026-10-31) Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Publish a plugin branch with only the runtime set on each release A plugin install copies what the marketplace entry points at; at the repository root that is 28 MB including tests and evals. release.yml now commits the npm tarball's contents plus the plugin manifest and skills/ (2.6 MB, 88 files) to the plugin branch. The marketplace entry stays at ./ until a release has published the branch. Measured with Claude Code 2.1.322 and Codex 0.167: a git source with ref installs with git alone; an npm source needs npm on the user's machine. design-decisions also records why the plugin does not register the MCP server (12 tools, about 17,000 tokens of definitions per session, for a second route to the same scripts). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * sweep.py: --list needs no input; pin its probe and timeout handling; say where plain paths are in the plugin copy Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Record the questions an outside reviewer kept asking, as a pre-push list 353 inline findings on PRs #92 to #435, grouped into twenty questions in .claude/skills/code-review/references/reviewer-findings.md, with the rule in AGENTS.md to ask them before a push and to add a question when a finding fits none. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin branch step: do not hold the npm publish back; tell a missing branch from a failed lookup Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * reviewer-findings: path guards hold in dry runs; side-file order depends on who reads the file Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * check.py --compare --expect-changed: fail on a picture change outside the declared scope (#439) A change_scope row measures the output against the reference tile by tile (32x18, 8-bit RGB difference) and grades it against spans the caller declares: a change elsewhere is FAIL with its time and place, a declared span that did not change is WARN, and two files that cannot be compared are FAIL. The floor (24) is set from phone and screen-recording footage. The plugin branch now holds the plugin under ffmpeg-skill/ for a git-subdir marketplace source: by ref alone Codex kept the whole clone (39 MB); as a subdirectory both Codex and Claude Code copy 2.6 MB. The publish script's remote answers are tested against real temporary remotes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: grade overlapping declarations independently; refuse a rectangle past the edge; unreadable frame times are not compared Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Refuse nan and inf for every float flag of every tool, and for every time string argparse's float reads nan, inf and 1e999; 232 of 232 runs across 25 tools accepted them (tracebacks, ffmpeg 'Invalid argument' as kind ffmpeg, or silently carried). add_common() now registers finite_float for type=float on every parser, and parse_time() refuses a non-finite time. A contract walk tests every numeric flag of all 42 tools. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: --change-floor 254 is asserted by what the document reports, not by a status that depends on the FFmpeg build Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Non-finite flag values: a failure document with --json, as the flags that checked themselves gave; keep 'not a time' in the time parser's refusal Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: rectangles are in the displayed picture (non-square pixels squared before the difference); unreadable frame times are not compared for none either Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: equal displayed sizes are compared whatever the stored size; a different stored size is a note Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * change_scope: the stored-size note reads the file's own width and height, before rotation Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync CHANGELOG with main after the 3.4.0 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Marketplace entry names the plugin branch's ffmpeg-skill/ directory A plugin install copied the whole repository (28 MB). The entry is now a git-subdir source on the plugin branch, which 3.4.0 published in that layout. README marketplace commands gain --sparse .claude-plugin. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the plugin route's minimum Claude Code version for a git-subdir source Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: a table of the install routes that work today, and the directories not listed yet Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin tree: ship the listing icon, leave the npx installer out The plugin directory's validation of the 3.4.0 plugin branch reported no icon and flagged bin/install.js. The tree now carries .claude-plugin/icon.png and no bin/; the npm package is unchanged. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the one-line plugin install needs 2.1.275; older versions use the two commands Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin tree: a short README of its own; workflow.md says add --overwrite The plugin directory held six findings as credential use; none reads one. Two came from the repository README and one from the word 'pass' beside a URL. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin README: transcription engines may download a model and use a logged-in token Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.1 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin icon: picture chunks only, pixels recompressed; drop mcpName from the plugin's package.json The plugin directory read the 3.4.1 icon as a download-and-run command: a C2PA provenance chunk's text sat beside compressed pixel bytes. The copy on the plugin branch keeps the same pixels and nothing else. assets/icon.png is unchanged. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.2 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Plugin listing: displayName, documentation, support and privacy links; PRIVACY.md The Claude plugin directory showed 'Ffmpeg Skill' with four of six links unset. The plugin branch's package.json also drops scripts. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * Sync with main after the 3.4.3 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: publish the MCP Registry entry after npm, by OIDC server.json was kept on the package version but never sent. The workflow now asks npm and the registry and publishes when npm has the version and the registry does not. mcp-publisher is pinned and its Sigstore bundle is verified. A failure never blocks a release. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: an npm that cannot be asked is not 'not on npm' Only npm's E404 is a quiet skip of the MCP Registry step; any other npm view failure stops the step and reaches the failure reporter. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: a tagged version npm does not have fails the registry step A re-run after a failed npm publish skips npm publish because the tag exists; the job would be green with the release on neither registry. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: report both publish failures; recovery text without a flawed recipe The plugin-branch failure step gains always(), since the MCP Registry's failure step now sits before it. The npm-missing message no longer gives a local publish command, and the re-run claim says which version it lists. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * release.yml: ask the MCP Registry for the exact name and version The list endpoint is paginated; a version on a later page would read as absent and be published twice. 200 and 404 are the only answers. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the MCP Registry entry is live Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * README: the registry row says what the release workflow does when its step fails Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: read the icon word list from the publish script and bound the PNG chunk walk The plugin-branch test kept its own copy of the words the publish script refuses in the icon, and walked PNG chunks without checking that a chunk fits in the file. It now reads the list from the script, walks chunks through one bounds-checked iterator, and pins documentationUrl and supportUrl in the plugin manifest. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: keep the seven known icon words as a fixed minimum Reading the list only from the publish script meant that removing a word there also removed it from the test. The test now requires the seven known words in the script's list and checks the icon against the union. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * feat(silence): --margin-lead and --margin-tail keep different amounts on each side of a silence --margin kept the same amount of each silence on both sides. --margin-tail is kept after speech ends and --margin-lead before it starts again; each defaults to --margin. The result's margins gives the two values used (absent under --filler alone, which removes no silences), and a render.py project takes silence.margin_lead and silence.margin_tail. First part of #449. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * feat(silence): --sentence-min-silence and --sentence-margin-tail for the pause after a sentence end With word timings (--words or --transcribe), a silence that follows a word ending in sentence-final punctuation gets its own floor and its own tail; every other silence keeps --min-silence and --margin-tail. The result lists silence_positions and sentence, and a transcript with no such punctuation changes nothing and warns sentence_ends_not_found. Refused with --speech-aware and without a transcript. Second part of #449. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(changelog): keep the sentence-flags entry under Unreleased after merging the 3.5.0 release Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): tell the word before a pause by word midpoints, and test the sentence flags on recorded speech The word before a silence was the last one that started before the silence's midpoint. On the repository's one speech clip the engine timed three word starts 0.45 to 0.5 s before the sound, inside the pause before them; a start stretched a little further would have made the next word the one before the pause. A word's own midpoint stays on its side of the pause whether the engine stretched the previous word's end or the next word's start. Adds a test on demos/footage/ai_talk_en.mp4 with its recorded word timings: the three pauses after sentences are told from the others, each keeps its tail, and no cut lands near sound. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): measure an open-ended silence, refuse unusable word timings, add the pauses_en demo Review of #452: - A silence that runs to the end of the file was kept whatever its length, so a 0.5 s pause after the last sentence could be cut under a 3 s floor. Its length is what is left of the file. - Words with a non-finite time or no length are left out of the sentence classification, and a transcript with none left is refused. - The sentence_ends_not_found warning and the docs said the flags changed nothing; with --filler they still turn silence removal on. They now say what happens: every silence uses --min-silence and --margin-tail. - demos: pauses_en runs the lead, tail and sentence flags on the recorded speech clip (CONTRIBUTING asks a feat PR for a before/after demo). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(silence): a silence that runs to the end of the file ends at the file's end when placing the word before it The last word's timing can reach into the trailing silence; with the silence's start as the reference its midpoint fell after it and the pause was read as inside a sentence. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test(silence): the sentence test makes its own clip OUT/gappy.mp4 is written by two fixtures with different layouts (MediaFixtures._gappy and test_silence_removal), so which seconds are silent depended on test order: the test passed locally and failed on CI, where the other layout was on disk. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(silence): margins is present under --filler when a sentence flag turns silence removal on Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * chore: sync the branch with main after #452 and the 3.6.0 release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): refuse DASH/IMF manifests by content, subfile URLs and line breaks in concat paths; never write temps through a planted link From the full audit: - A DASH manifest saved as .mp4 made every tool follow its BaseURL to another local file or over HTTP (FFmpeg 6.1's -protocol_whitelist file did not stop it). DASH and IMF are now told by the content ffmpeg's own probes look for. - looks_like_url() stopped at the comma of subfile,<options>:<url>. - concat_list_line() wrote a path with a line break as is, adding directives. - The staging temp and place_output()'s temp had fixed names opened through whatever was there, a planted symlink included; they are created with O_EXCL, and a lock's default audit log is opened with O_NOFOLLOW. - render.py's output-is-input guard covered clips and overlay videos only. - color_hex() checked the length only, so 'yellow' drew black, verified. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): export.py's fill temp and batch.py's delivery temp are reserved exclusively too Sibling sweep of the fixed hidden temp names: both built .<stem>.ffskill-... names and opened them through whatever was there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): copy through the reserved descriptor, refuse the 3GPP DASH profile, guard overlay fonts Review of #456: - place_output() and batch.py's cross-filesystem delivery closed the exclusively created temp and reopened it by name with copyfile(); they now write the descriptor that created it (copy_to_sibling()). - dash_probe() also accepts 3GPP:PSS:profile:DASH1 in place of dash:profile. - project_input_files() takes overlays[].font_file, and an overlays value that is not a list is left to the overlay stage instead of a traceback. - audiogram.image reaches waveform.py resolved against the project file, the path the output guard checks. - Without O_NOFOLLOW (Windows), a symlink at the default audit-log name is refused by looking first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): guard the project file, resolve overlay fonts, fail closed on an unreserved temp Second review of #456: - render.py's output guard includes the project file itself - an overlay's font_file reaches overlay.py resolved against the project file - reserve_sibling() refuses (output_dir_not_writable) when the exclusive create fails instead of returning a name nobody reserved - copy_to_sibling() closes the descriptor if wrapping it fails Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): verify the default audit log after opening it where O_NOFOLLOW is missing Without O_NOFOLLOW a link could appear between the islink() look and the open; the descriptor is now compared with the name (lstat) before anything is written, and closed unwritten if they differ. The render guard test compares forwarded paths by file identity, not spelling (Windows 8.3 short names). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): guard the brand's own logo and font, skip the output guard on previews Third review of #456: a brand file's font_file and logo (resolved against the brand file, as load_brand() does) are inputs the output may not replace; a --stop-after preview never writes the output, so it is not refused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(security): refuse a hard-linked default audit log where O_NOFOLLOW is missing A link swapped for a hard link to its target before the lstat passed the identity check; the opened file must also be a regular file with one name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: require the --stop-after preview to succeed, not only to avoid one refusal Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * chore: sync with main after #456 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(batch): project side files beside the project; --watch waits for a source to settle A batch project recipe wrote each item's project copy into its work directory with relative paths unchanged, and render.py reads them against the project file's directory, so "bed.wav", a brand, captions or a chapters file beside the project failed every item. The copy now carries them absolute through one list of a project's path keys (_common.decision.PROJECT_PATH_KEYS / rebase_project_paths), which render.py applies to its own project too (audiogram.image and overlays[].font_file were read against the working directory) and multicam.py --write-project applies before writing (cwd-relative clip paths in a file written elsewhere). batch.py --watch processed a source still being copied, reported the truncated output ok (verified for MKV/TS), and once the copy finished refused every later pass with output_exists. A source is now processed once its size and mtime are unchanged across one poll (first seen: last modified at least one interval ago); until then its row is skipped: "settling", not cached. An output an earlier pass of the same watch delivered and nobody changed since is replaced; any other existing output is still refused without --overwrite. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit fb471e619fab764305dd6ec616dcd377ab7791b7) (cherry picked from commit b32fe8bcdc97f83cc1ad61905f0077072afb351d) * fix(captions): keep text symbols, spaces at moved breaks and SRT markup Three ways the burned caption differed from the words handed in, each fixed at its shared cause: - emoji.py counted the whole Misc Symbols / Dingbats / Misc Symbols and Arrows blocks (and the playing cards) as emoji, so a mode-none run cut out SDH music notes and check marks. In those blocks only Emoji_Presentation characters start an emoji; a text-default symbol does with VS16 or a skin tone. Reaches caption, graphics and overlay. - wrap.py's orphan, weak-line and rebalance passes guessed from the script whether a space stood at a break, dropping it for Korean, spaced Japanese/Chinese, emoji neighbours and Latin/CJK pairs. The greedy fill records each break's space and every pass carries it; Korean is wrapped as a spaced script. - caption.py's own ASS route escaped SRT markup and drew it as text. <i> <b> <u> <s>, <font color> and {\anN} become the overrides ffmpeg writes; wrap, fit, emoji placement and the blank-cue count measure the words only; split cues carry their tags; other {\...} blocks stay escaped (design-decisions entry). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit f6fc0e390c5b17ddce708727fa7ed3cae7db78ca) (cherry picked from commit bd48428c2cfe34e93c86052f9ae62d689d02a0bd) * fix: speedramp frame durations, sync --trim-second copy, upper-case MPEG-4 suffixes, moved start timecode - speedramp: restore each segment's frame durations after setpts (frame_durations), so a sped-up segment no longer pushes the rest of the picture two frames late and adds frames. - sync --trim-second: never stream-copy a video second file (the copy started at the keyframe before the trim point); re-encode from the frame. A sound-only file is still copied. - cut: lower-case the output suffix in the join paths, so .MOV/.MP4 keep hvc1 and +faststart; signatures_match recognises an upper-case .TS. - Start timecode: cut (one range), fit --method trim and sync --trim-second write the label of the frame the output starts on (render clips follow via cut); a join and an MXF copy write none. source_carry compares the timecode's value. Swept: suffix comparisons across scripts/ (only cut.py and signatures_match missed .lower()); setpts rescales feeding a concat or ending a graph (fit --method speed measured exact; freeze and reverse already restore durations); stream copies behind an input -ss (only sync's); tests/sweep.py on a timecoded MOV against a plain one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 1902b948f54d99f9d8e6758c6ffe1e717f3357dd) (cherry picked from commit 6eae59b1b0e4f377f743bdd3fd660011b6730cfb) * fix: downmix gain, HDR tone-map peak, PiP opacity colour and 10-bit overlays - audio.py --downmix: ITU-R BS.775 weights (centre and surrounds -3 dB, LFE dropped), renormalised with pan's '<' so in-phase channels cannot clip; pan's '=' summed 2.7x per side and a -6 dBFS 5.1 came out hard-clipped while the run said verified. - color.py --to-sdr and look.py's HDR sheets: linearise at npl=100 (the unit the tonemap filter reads) and pass the source peak as tonemap's peak, so a PQ pixel at --peak nits and HLG 100 % reach SDR white (they stopped at Y 151); --peak is a PQ setting, HLG peaks at 1000 nits. The chains move to _common/color.py; look.py converts a BT.2020 SDR source's gamut only. - overlay.py --video --opacity scales the alpha with lut in YUV instead of the RGB-only colorchannelmixer, which made the whole main picture composite in RGB and come back BT.601 under its BT.709 tag. - overlay_format(): broll, redact (blur/pixelate), graphics layers and boxes, emoji (graphics, caption), overlay --video and --image composite a 10-bit HDR/ProRes picture at 10 bits instead of overlay's 8-bit default (or 8-bit RGB for --image); a 10-bit PiP stays 10-bit. Regression tests: DownmixLevelTests, HdrLevelsAndDepthTests (fail on the previous tree, pass here). Design decisions, README, references/scripts.md, demos and CHANGELOG updated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit fb2123404d2880ae39ca9c3e797b8294a60cf950) (cherry picked from commit 2a54302d7f3388d152c368a4b0d2dc17d7026cf8) * fix(mcp): truthful tool annotations; a structured value is never read as an option A9: probe, check and cropdetect were annotated readOnlyHint, yet plan + overwrite let them replace any existing file, and every tool said destructiveHint: false. The server now runs the read-only tools with FFMPEG_SKILL_READ_ONLY=1, and apply_common refuses --plan under it after argparse (kind input, error.reason read_only_plan_refused), so neither the structured form nor any raw-argv spelling writes a file. The plan/overwrite property names stay in their inputSchema (3.x removes no property), with descriptions that say they are refused. Every writing tool is now destructiveHint: true. The CLI is unchanged. A10: build_argv emitted [flag, value] and bare positionals, so "-50%" was refused and an input "--plan=x" became an option. A value starting with '-' is now one --long-flag=value token (_common.option_args), and positionals starting with '-' follow '--'. render.py's stage commands use the same helper, so a project overlay or title starting with '-' works. Swept: every string option and positional of all 42 tools through build_argv into each tool's own parser (new test), and render's stage argv builders; batch.py recipes are raw argv and stay as given. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 1bb99302bcba9c71ebfd2c24209316b7cae7cb22) (cherry picked from commit 7924aca092d0ae5330b5e8c8f96ca2b9a378f9cc) * fix(silence): name the sentence-flag transcription path everywhere, and look the word up by binary search The audit found that #452's silence.py --transcribe with a sentence flag runs the speech engine (under --dry-run too) while the contract's capabilities, refusal reasons, model-download statement, PRIVACY.md, the plugin README and README named only --filler --transcribe; README's silence row and SKILL.md did not mention the 3.5.0/3.6.0 flags. silence_positions() also rescanned the whole transcript for every silence (11 s for three hours of speech). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit ad14bb441660fd12119151aff265f31225816db4) (cherry picked from commit 04e2bb549be62d2d349bd079aa11faa6e7c05c78) * test: name the render.py paths the output guard reads in the rebase list check Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 4e4461da09495fa7f78d8176839d9b3ea1a1b53a) * ci(security): pin every action to a commit, grant writes per job, fuzz the parsers Scorecard Pinned-Dependencies, Token-Permissions, Fuzzing and Security-Policy: - every uses: names the commit of the tag it replaced (Dependabot keeps both) - no workflow grants a write scope at the top level; checkouts that never push drop the token (persist-credentials: false) - the landing page loads the one Phosphor stylesheet it uses with SRI instead of the unpkg script that injected six at run time - ClusterFuzzLite with atheris targets for parse_time, parse_srt and the timeline XML reader; the contract suite runs each target on its seeds - SECURITY.md links the advisory form by its full URL Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit a2d33054c0479b0b7dba3c6e04f4ba7b9ae524e0) * docs: the Unreleased entries for the merged fixes; why ffmpeg's temps stay beside the output Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test: compare the landing page's asset hosts by hostname, not by substring CodeQL (incomplete URL substring sanitization) on the SRI check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * docs(readme): why the Claude plugin directory listing is on hold Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(overlay): label an untagged PiP beside a tagged main (FFmpeg 8.0 converted the main) The new PiP test failed on ubuntu-latest's FFmpeg 8.0.1: a BT.709-tagged main under an untagged PiP came back 88/103/192 instead of 80/110/192 at every opacity, because 8.0's overlay negotiates one colourspace and took the untagged side as BT.601. The untagged picture is now labelled whenever the other one is tagged, not only when one is converted; two untagged inputs are left as they are (the decided join/grid/broll/multicam plans are unchanged). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(fuzz): a negative time is a valid parse; say what the SRI and fuzz jobs cover parse_time is the grammar for offsets too (sync --offset -0.5), so the parse_time target no longer treats a negative result as a crash; seed added. CHANGELOG: SRI pins the icon stylesheet (its fonts load from it unpinned), and the PR fuzz job runs on the workflow's path filters. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(caption, look): whitespace-only markup, a split cue's alignment, --compare's own display conversion Three CodeRabbit findings on #469, each verified against the tree first: - reinsert_srt_markup anchored a tag pair that encloses no visible character like any other tag: closing after the character before, opening before the one after, so `A<font color=red> </font>B` coloured B. In a gap holding such a pair the tags now keep their source order (opening before the whitespace, closing after; whole at the next line's start when the whitespace became the break). Other tags are anchored as before. - carry_srt_markup took the cue's {\anN} from the first chunk that had one, so chunks before it were left unplaced. The alignment is now read from all chunks first. - look.py --compare built the display conversion from the first input's tags and ran it on both pictures: an SDR clip beside a PQ one was read as PQ, and an HDR clip given second was not tone-mapped. Each input now gets display_conversion() of its own probe; BT.709 SDR gets none. Swept: look.py is the only tool that converts two inputs for display (color.py has one input; grid.py's no-tone-map is decided); reinsert/carry_srt_markup have one caller (caption.py layout_cues). Tests: test_markup_around_whitespace_only_keeps_its_place, test_a_split_cue_takes_its_alignment_from_any_piece, test_look_compare_converts_each_input_by_its_own_tags (all fail on the parent commit). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit 95b6ee8997b69dd09a874c14dbdaf01dee86bfcf) * fix(timecode): count the start label at the delivered rate, from a late picture's first frame; sync seeks early Three CodeRabbit findings on #469: - cut.first_frame_from (re-encode) and fit --method trim clamp the start to the picture's first frame before the grid lookup: a source whose video starts after the file's origin was stamped a frame before its picture (00:59:59:21 for a cut from 0.2 s of a picture starting at 0.5 s). sync's padded head still counts back. - output_timecode_args takes the delivered rate: the source label is read at the source's rate and written at the output's. fit passes --fps (or the rate cfr_args forces) and relabels whenever the rate changes, without copying the source's tmcd track (run_keeping_subtitles(relabel_timecode=True)); cut's re-encode passes the rate cfr_args forces on a VFR source; sync does not change the rate. render.py reaches frame.fps through fit, so it is covered. - sync --trim-second seeks SEEK_MARGIN early and drops the margin on the output side (on the drift-corrected clock), as cut.py's re-encode does. SEEK_MARGIN moves to _common.decision so both import it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs (cherry picked from commit c54f59fbd527e31ebd5418fd14e4732bc97418dd) * fix(batch): a watch pass replaces only the files it delivered, checked per path where each is written The replace permission was a Boolean taken when the item started and handed to every step as --overwrite, so a hand-edited sidecar (.srt) or a file swapped in meanwhile was replaced. Now each step gets a snapshot of the files this item delivered (size, mtime, inode) and its own output check passes only those, still unchanged; deliver() compares the final path again under the output lock. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(timecode): round the start label once at the delivered rate; fit trim seeks early The source label and the offset to the first frame were rounded apart, so a rate change could land one frame off (4.8 frames written as 4). A delivered rate below 1 fps divided by a zero-length day; it now writes no label. fit --method trim seeked straight to its start before -i, which can lose the frames before a keyframe on a B-frame HEVC source; it now seeks SEEK_MARGIN early and drops the margin on the output side, as cut.py does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * test(sync): read the first frame count FFmpeg 8's ffprobe prints for the stream On macOS and Windows (FFmpeg 8) csv output repeats the stream section, so the count read back as '92\n\n92'. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs * fix(batch): record each delivered file as it was when written, not when the item ends caption.py writes its SRT before the video finishes encoding. batch stated every reported path after the whole item completed, so an SRT corrected while the video encoded was recorded as batch's own and the next watch pass replaced it. Each writer now reports the identity its file has right after the write (runner.report_committed, from run()'s place step, place_output, write_srt and remember_written), and deliver() records the final path under the output lock. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013tPrsGXLuaGU7t643QNUZs --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Claude <noreply@anthropic.com> | 6 days ago |
| Files | Last commit | Last update |
|---|---|---|
| 6 days ago | ||
| 6 days ago | ||
| 6 days ago | ||
| 6 days ago |