Skip to content

CHANGELOG.md

Latest commit

 

History

History
203 lines (155 loc) · 16.7 KB

CHANGELOG.md

File metadata and controls

203 lines (155 loc) · 16.7 KB
Raw

0.10.3 — 2026-08-31

  • The caption file's header no longer opens the transcript. A platform .vtt starts with WEBVTT / Kind: captions / Language: en, and only the first of those was being dropped — so since 0.10.2 made captions the normal path for URL runs, every such transcript.txt began with two lines nobody said. Everything ahead of the first timecode is file header by spec, so that is what is skipped now; NOTE, STYLE and any header key added later are covered too, and cue text is untouched.

0.10.2 — 2026-08-28

  • URL downloads now use the platform's own captions (#25, by @Daily-AC). The "existing subtitles" fast path only ever held for local files — a URL run never asked yt-dlp for captions and always fell through to Whisper, even when the platform had a transcript sitting right there. crv now reads the video's metadata once and asks for a single exact language code (preferring --lang, then the video's own language, then English; manual tracks before auto-generated) — not --sub-langs all, which on YouTube means 157 auto-translated tracks, and not a regex, which still matches them. Fail-open: any metadata failure just means Whisper runs as before.
  • A partial download no longer hands ffmpeg a subtitle file (#25). The source.* fallback glob now screens out .vtt/.srt-style sidecars, which could previously be picked up as "the video" and die with no video stream.

0.10.1 — 2026-08-27

  • --to keeps the frame timestamps again (#19, #21). The window's -t was an output-side option, so ffmpeg decoded past the window end and the showinfo log covered more frames than were written — the count mismatch silently threw away every timestamp (no frames.json, no frame timestamps: line). -t now sits on the input side with -ss.
  • A window now reaches shipped captions too (#20, #23). Sidecar .srt/.vtt files and embedded subtitle streams are clipped to --from/--to, so a windowed run gets a windowed transcript whichever path produced it.
  • --text-anchors lands on the right frames under --from (#20, #24). Anchor frame indices are rebased onto the window start — an input-side seek restarts ffmpeg's frame counter at zero, so source-clock cue times pointed start x fps frames too late.
  • A broken Pillow install fails loudly instead of silently skipping dedup (#22). Pillow is a declared dependency; when it cannot import, the run now raises with a reinstall hint instead of producing #19's exact signature (no frames.json, no timestamps, exit 0) from a different cause.

All four were reported — and three arrived as ready pull requests with regression tests — by @Jassu225. Thank you.

0.10.0 — 2026-08-22

  • --from / --to analysis window (#16): analyse a slice of a long video without re-encoding it first. The frame budget and the transcript follow the window; reported timestamps stay on the source clock.
  • --frame-width (#17): control the extracted frame width (default 640).

0.9.3 — 2026-08-18

  • A failed ffmpeg no longer reads as an empty video (#15). extract_frames() and extract_frames_adaptive() never checked ffmpeg's exit code, so a run whose ffmpeg invocation failed outright looked identical to a video with nothing to extract — 0 frames, exit 0 on older versions, and a generic "no frames" error on newer ones. Frame extraction now raises with ffmpeg's actual stderr (ffmpeg frame extraction failed (exit 8). ffmpeg said: …) whenever ffmpeg exits non-zero and produced nothing, the same policy the whisper branch has always had. Thanks @HPoyisis for the precise report, the shim reproduction, and the fix suggestion.

0.9.2 — 2026-08-18

  • YouTube downloads work again out of the box. YouTube now requires solving a JavaScript "n challenge" before serving video data; yt-dlp older than 2026.07 can't solve it, so URL inputs failed with HTTP Error 403: Forbidden or "n challenge solving failed". The dependency is now yt-dlp[default,deno]>=2026.7.4, which pulls in yt-dlp's official EJS challenge solver and a bundled Deno runtime — no system Node/Deno setup needed. Existing installs: pip install -U claude-real-video fixes it.

0.9.1 — 2026-08-17

  • ffmpeg 9.0 compatibility (#14). ffmpeg 9.0 removed the long-deprecated -vsync option, so every extraction failed with "No frames could be extracted". crv now probes the installed ffmpeg once and uses -fps_mode vfr when available (ffmpeg 5.1+), falling back to -vsync vfr on older builds. Thanks @volpatocode for the precise report.

0.9.0 — 2026-08-11

  • Memory across videos. crv now keeps a local index of everything it has watched (transcript lines + on-screen text, timestamps included), so you can ask across your whole library instead of re-processing:

    crv-ask "pricing strategy"     # → which video, which second, the exact line
crv-ask 定價                    # CJK works: trigram FTS + substring fallback
crv-ask --list                 # everything you've watched, newest first
crv-ask --prune 200            # keep the newest 200, reclaim the space

    Re-running the same source with the same options now says "already watched" and points at the existing analysis in 0.04s instead of re-processing (measured against 5.5s); different options or --overwrite re-analyse as before. Everything is local — one SQLite file at ~/.crv/memory.db, user-only permissions, no embeddings, no network. First indexing prints a one-line notice; opt out entirely with CRV_NO_MEMORY=1.

  • MCP server: 2 tools → 5. New: search_memory (ask across every watched video), list_watched (check before re-watching), get_transcript (words only, no frames — saves context when you don't need images). watch_video now indexes automatically.

  • Hardened against review findings before release: FTS5 queries are injection-safe and verified literally after tokenizer fallback, empty queries and unbounded limits are rejected, the schema refuses to downgrade an index written by a newer build, migrations are atomic (no executescript inside the write lock), foreign keys are on, the DB file is chmod 600, and concurrent writers take BEGIN IMMEDIATE (6 simultaneous writers tested clean).

0.8.2 — 2026-08-08

  • Pro pointer at the end of a run now shows current pricing ($29 one-time, launch code through Aug 31) instead of the expired founder price. Opt out with CRV_NO_HINT=1 as before.

0.8.1 — 2026-08-03

  • Registry metadata only: mcp-name marker in the README and a server.json for the official MCP Registry (registry.modelcontextprotocol.io). No code changes.

0.8.0 — 2026-08-02

  • MCP server. pip install 'claude-real-video[mcp]' then crv-mcp — the crv pipeline over the Model Context Protocol, so Claude Desktop, Claude Code, Cursor and any MCP client can ask for a video to be watched. Two tools: watch_video (transcript + first batch of keyframes as inline images) and get_frames (page through the rest). Analyses cached per source under ~/.cache/crv-mcp; frames are resized to 768px before returning so responses stay context-friendly. Verified end-to-end on Claude Code with a real model run.

0.7.19 — 2026-07-31

  • Fixed the broken images on the PyPI page. The README pointed at docs/… with relative paths. GitHub resolves those; PyPI does not, so the demo poster, the animated demo and the contact-sheet example had all been rendering as broken-image icons on the package page — since at least 0.7.17, not just this release. They now use absolute raw.githubusercontent.com URLs, which render in both places.

0.7.18 — 2026-07-31

  • --yt-dlp-arg: pass raw options straight through to yt-dlp, repeatable. Asked for by @IamBennyOuO (issue #12) for pre-production workflows where the working answer to YouTube's JS challenges changes week to week.
crv "<url>" --yt-dlp-arg=-S --yt-dlp-arg=res:1080
crv "<url>" --yt-dlp-arg=--remote-components --yt-dlp-arg=ejs:github

Works on both paths: appended to the command line when the yt-dlp executable is used, parsed with yt-dlp's own parser when the Python API is used, so the accepted syntax is identical either way.

  • Docs corrected. turbo was described in three different ways across the CLI help and the README — "large-v3 accuracy" in one place, "close to large-v2 accuracy" in another. Per the model card, turbo is a pruned and finetuned large-v3: much faster than large, with a minor quality trade-off. The "~8x the speed" figure was removed everywhere; it had no source — no turbo benchmark exists in this repo and OpenAI publishes no multiplier.

0.7.17 — 2026-07-29

Isolated installs (pipx install / uv tool install) could not fetch URLs at all — fixed by @IamBennyOuO (PR #11):

  • yt-dlp and whisper have always been dependencies, so both are installed. But those installers expose only crv's own entry points on PATH, so shutil.which("yt-dlp") found nothing and crv stopped with "yt-dlp not found". URL downloads were dead for anyone who installed that way.
  • crv now falls back to yt-dlp's Python API when the executable is absent, keeping the same cookie-retry order as the command-line path.
  • whisper is probed with importlib.util.find_spec rather than PATH — and deliberately not imported, since that pulls in torch on every silent video.
  • --cookies-from-browser specs are parsed with yt-dlp's own parser, so the accepted syntax stays identical to the CLI.

0.7.16 — 2026-07-21

Batch-hardening, ported from a 2,181-video field report against crv Pro:

  • dedup: new "action channel" — a handful of 32x32 cells changing hard (>45/255) marks a frame as new regardless of percentage. Small-in-frame fast action is no longer deduplicated away (synthetic repro: 1/10 action frames survived -> 10/10).
  • all ffprobe/ffmpeg output decoded with errors="replace" — Latin-1-ish metadata no longer crashes a run.
  • --max-frames default now scales with duration: clamp(150, seconds*1.5, 600); explicit value still wins.
  • --min-frame-interval added as an alias of --fps-floor; both help texts now say plainly it is seconds per frame.
  • distribution: Claude Code plugin marketplace support (/plugin marketplace add HUANGCHIHHUNGLeo/claude-real-video).

0.7.8 (2026-07-15)

  • The end-of-run Pro pointer now shows this run's real numbers (deduped visual-change count and changes/min) instead of a generic line — only when the video actually has them; static or very short videos keep the quiet one-liner. Opt out unchanged: CRV_NO_HINT=1.

0.7.7 (2026-07-15)

  • faster-whisper support — new [fast] extra. pip install 'claude-real-video[fast]' and crv automatically transcribes in-process with faster-whisper (CTranslate2): same model names, same transcript.txt + transcript.json output, several times faster and lighter on RAM than the whisper CLI. No new flags — if the package is importable it's used, and any failure falls back to the whisper CLI automatically.
  • The "install whisper" hint shown when no transcriber is present now mentions both options.

0.7.6 (2026-07-13)

  • Per-frame source timestamps — frames.json (issue #7). Every kept frame_XXX.jpg now knows exactly which second of the original video it came from, surviving extraction, deduplication, --max-frames thinning and renaming. A machine-readable frames.json is written next to the frames: {"file", "timestamp_sec", "timestamp", "selection_reason"} — so you can cite visual evidence with a timestamp, align frames with transcript.json segments, or feed the mapping into video-RAG / lecture-note pipelines.
  • Timestamps come from ffmpeg's showinfo log on the same select pass (no extra decode pass, VFR-accurate); the adaptive extractor gets them the same way. If the log and the extracted files ever disagree in count, crv writes no timestamps rather than wrong ones.
  • viewer.html shows each keyframe's MM:SS.mmm on its cell, and the lightbox gains "play video from here" — click a keyframe, jump the player to that exact moment.
  • MANIFEST.txt points the reading LLM at frames.json so reports can cite [frame_012 @ 00:03:41.2] instead of "somewhere in the middle".

0.7.4 (2026-07-11)

  • Dedup was blind to local change — fixed. Benchmarking (benchmark/benchmark.md) caught the old comparator — a 16x16 downscaled signature with a 25/255 per-channel tolerance — measuring thin pen strokes, caption/text-card swaps and small UI updates as exactly 0.0% difference, so it dropped them at any --dedup-threshold, including 0. Measured damage: a handwriting clip collapsed 52 frames to 3 (blank page + finished page, no progression), a text-card reel lost 2 of its 5 caption cards, and a screen recording lost its most information-dense final state.

  • The fix adds a second, settled-local detector next to the (unchanged) global one: on a finer 192px signature it looks for a region that differs strongly from every recent kept frame — with 1px shift tolerance so film grain/jitter don't trigger, and a stricter second tolerance pass so soft-contrast drift (smoke, clouds) doesn't — and that is no longer changing toward the next frame: a settled new state, not motion mid-flight. A cooldown stops continuously-"settling" motion (a waving flag pausing every second) from taking a frame each time. Full derivation and before/after numbers: benchmark/benchmark.md.

  • Default behavior changes: content the old dedup was blind to now produces more frames (that's the point). On the benchmark set: handwriting clip 3 → 9 frames (progression visible), text-card reel 3 → 8 (all 5 cards present), screen demo 6 → 15 (final summary state included), JFK 1962 footage 53 → 77 (+45%, real settled pose changes on grainy film — the cost of the fix, kept in check by the cooldown), NASA launch 20 → 28 (ignition/liftoff sequence now sampled more densely), slow orbital pan 101 → 109, speaker+slides 87 → 90. Runtime on the 8.8-min launch video: +3s (10 → 13s end to end).

  • Package metadata could report a stale version (e.g. 0.4.0) on editable installs made before the version bump — the dist-info was generated at install time and never refreshed. __init__.py's fallback is now kept in sync with pyproject.toml; if crv --version-style checks show an old number on an editable install, re-run pip install -e . to regenerate the metadata.

  • report.html now labels frames kept by the settled-local detector with their settled-change %.

  • benchmark: run_benchmark.sh now downloads the JFK test clip via yt-dlp instead of shipping a 13MB mp4 in the repo; added a "v0.7.4 dedup fix — before/after" section to benchmark.md.

  • Note: the final frame is exempt from the "scene has settled" motion check (the closing state of a video is always considered), but still has to pass both contrast gates.

0.7.3 (2026-07-10)

  • --whisper-model now accepts turbo (close to large-v2 accuracy at ~8x the speed; needs openai-whisper>=20240930). Default stays base for fast first runs; sharper transcripts are one flag away.
  • Transcription failures now print whisper's actual error instead of a silent "(none — transcription failed)".

0.7.2 (2026-07-10)

  • Safer output directories: running into a folder that already holds a previous analysis is now refused, so two videos can never mix frames or audio. Pass the new --overwrite flag to replace it (only crv's own artifacts are removed). Recommended: one folder per video.
  • Fail loudly on bad sources: zero extracted frames now raises a clear error (incomplete download / not a playable video / check ffmpeg) instead of quietly producing an empty result; partial-download leftovers (.part/.ytdl/.tmp) are no longer picked up as the video.
  • Honest silent-video diagnosis: a video with no audio track now says so, instead of telling you to install whisper.
  • Cleaner output: the temporary 16kHz audio.wav used for transcription is removed after Whisper finishes (--keep-audio's audio.m4a is untouched).
  • Windows fix: viewer.html is read/written as UTF-8 explicitly — CJK content no longer crashes on cp1252.
  • __version__ now reports the installed package version.
  • Docs: README install commands show the [whisper] extra (extras never auto-install), skill-install instructions clone the repo first, Options table lists all flags, and --text-anchors wording matches reality (sidecar/embedded subtitles only).

0.7.1 (2026-07-10)

  • Timestamped transcript: every analysis now also writes transcript.json — the same transcript as per-line segments with start/end times (from Whisper segments, or the video's own subtitle cues when available). Pipe it into your own tools, or give your LLM timings instead of a wall of text.
  • README: build-in-public link; crv-web footer credit.