Open-source release video

Video: final.mp4 (final-4) · 1920x1080 · 30 fps · 35.00 s · 18.0 MB · -14.1 LUFS / -1.1 dBTP · GitHub/chat copy: exports/final.github.mp4 (9.4 MB, under the 10 MB cap)

The request#

"Make a 35-second what's-new video for Black 26.1.0 — show what the 2026 stable style actually changes in people's code."

Mode: quick, publish-bound (researcher facts before the build, critic pass on the final). No questions asked. The opening line stated the assumptions:

Quick mode: 35 s, 16:9, 30 fps, no voice-over (on-screen text, a deep-house bed at 122 bpm, key clicks under the typing, one glitch on the bug-fix beat). The diff shown is Black 26.1.0 run on code that 25.12.0 had already formatted, which is what someone who upgrades sees in review. Every code line and terminal line is real output from both versions. Unofficial-summary label on screen and in the share copy; no logos.

Contract: upgrading to Black 26.1.0 (18 Jan 2026) turns on the 2026 stable style, and a few real patterns in code you already formatted change: redundant parentheses around assignment targets go, type comments are normalised, and a docstring after a comment is now treated as the module docstring.

The story, shot by shot#

The spine is the diff gutter: the red/green −/+ of the hook returns as the diff panels, the + rows of scene 5 and the $ prompt on the end card. A small "unofficial summary" pill sits top right on every frame.

Where every line on screen comes from#

• The diffs (evidence/samples/): pipeline.py and settings.py were written for this video (*.before.py), formatted with Black 25.12.0 (*.25.12.0.py, what users have today), then run through Black 26.1.0 (*.26.1.0.py). showtime code <25.12.0 file> --to <26.1.0 file> made the panels; after review round 1 the token files were rearranged, not re-typed (see Review round 1): the removed lines are kept as rows, and settings.py is split into two steps, 25.12.0 → +#4764 (diff-2a-docstring.json) → +#4645 = 26.1.0 (diff-2b-type-comments.json). Black 25.12.0 --preview on the 25.12.0 output gives byte-identical results to 26.1.0 for both files, which backs "All of it was in --preview during 2025" for these samples (the researcher checked 9 more).
• The terminal (evidence/terminal/session.log): a fresh CPython 3.12.9 venv, pip install black==26.1.0, black --check app.py, black app.py, captured 2026-09-27 in a neutral folder with a relative file argument (Black prints absolute paths for ., so black --check . was not used). On screen, the ~30 "Collecting / Using cached" lines are elided as "…" and the dependency list after black-26.1.0 as "…"; everything else is verbatim, emoji included (drawn with Noto Emoji so they look the same on every OS).
• Facts (researcher, primary sources fetched 2026-09-27: the GitHub release and PR API, CHANGES.md at the tag, PyPI JSON, the docs' stability policy): release published 2026-01-18T04:49Z; 9 stabilised features; 6 produce an upgrade diff on 25.12.0-formatted code (#4800, #4720, #4710 do not, so they are not shown); PR authors from repos/psf/black/pulls/<n>; pip install -U black installs 26.5.1 today, hence the "newer 26.x releases keep the 2026 style" note (stability policy). Not claimed: speed, user counts, "latest version", maintainer titles.
• Release-notes excerpt: showtime site component https://github.com/psf/black/releases/tag/26.1.0 ".markdown-body" --dark, cropped to the Highlights list (no GitHub chrome, no logos).

Commands, in order#

showtime is skills/showtime/bin/showtime; <job> is showtime-out/black-26-1-0-20260927-131402, <p> is <job>/project. The two Black versions live in scratch venvs outside the example.

showtime doctor --quick
showtime job init black-26-1-0 --mode quick --platform github --goal "Make a 35-second what's-new video ..." \
  --assumed "35 s, 16:9, 30 fps, no voice-over: on-screen text + deep-house bed (122 bpm) + keyclick/glitch SFX" \
  --assumed "the diff shown is 26.1.0 run on code 25.12.0 already formatted" --assumed "unofficial label; no logos" \
  --assumed "every diff and terminal line is real output from both versions"
showtime new dom <p> --title "Black 26.1.0" --duration 35      # the template's timeline rescaled to 35 s, then rewritten

# real material
venv-25.12.0/bin/black -q x.25.12.0.py; venv-26.1.0/bin/black -q x.26.1.0.py   # per sample, see evidence/samples
venv-25.12.0/bin/black --preview -q - < x.25.12.0.py | diff - x.26.1.0.py       # preview == 26.1.0
(fresh venv) pip install black==26.1.0; black --check app.py; black app.py      # -> evidence/terminal/session.log
showtime site component https://github.com/psf/black/releases/tag/26.1.0 ".markdown-body" --dark -o capture/notes-gh.png
showtime assets emoji 💥 -o <p>/img/    # also 💔 ✨ 🍰 (Noto, Apache-2.0, license sidecars kept)
showtime code samples/pipeline.25.12.0.py --to samples/pipeline.26.1.0.py -o <p>/code/diff-parens.json --theme houston
showtime code samples/settings.25.12.0.py --to samples/settings.26.1.0.py -o <p>/code/diff-comments.json --theme houston
#   (round 1 turned these into diff-1-unified.json, diff-2a-docstring.json, diff-2b-type-comments.json)

# sound: the mix is generated by a script so every key click sits on the frame its character is typed
showtime audio sfx keyclick --variants 4 -o <p>/audio/key.wav
showtime audio sfx glitch --key A --intensity 0.7 --seed 5 -o <p>/audio/glitch.wav
showtime audio sfx thock --intensity 0.5 --seed 2 -o <p>/audio/thock.wav
python3 <p>/tools/keyclicks.py          # writes audio/mix.json: deep-house bed (sections on the scene cuts), 57 keys + 3 returns, glitch, 4 thocks
showtime audio mix <p>/audio/mix.json -o <p>/work/mix.wav   # read the report; 3 rounds (see below)

# first look and fixes
showtime check <p>        # round 1: 24 contrast errors, 2 system-font warnings, holds -> fixed (see below)
showtime snap <p> --at 10.05,10.15,10.25,10.35   # CSS glitch vs WebGL signal-glitch: picked signal-glitch
showtime snap <p> --every 1                       # contact sheets, 4 rounds
showtime check <p>        # PASS, 0 warnings

# final, verify, deliver
showtime render <p> --job <job>                   # final.mp4 22.6 MB (crf 16) -> "render": {"crf": 18} in showtime.json
showtime qa <job>                                 # FAIL only on the github 10 MB cap for the master
showtime deliver exports <job> --targets github,chat,x,linkedin
showtime render <p> --job <job>                   # final-2 (laptop layout fix), final-3 (audio fix)
showtime qa <job> --platform youtube              # final-3: PASS
showtime deliver exports <job> --targets github,chat,x,linkedin
showtime qa <job>/exports/final-3.github.mp4 --platform github   # PASS, 9.4 MB
showtime review-pack <job>                        # self-review answered into review/round-1/FINDINGS.md

# review round 1 (critic): fixes, then final-4
showtime snap <job>/final-3.mp4 --at 20.93,21.0,22.5            # confirmed: output printed before Return
showtime snap <p> --at 5.5,6.3,8.0,11.0,12.5,13.2,13.7,15.5,21.2,27.4,29.5,32.6 --sheet
showtime check <p>                                # PASS, 0 warnings (after easing the change-2 dim, see below)
showtime render <p> --job <job>                   # final-4.mp4, poster = 32.6 s (not baked)
showtime qa <job>/final-4.mp4 --platform youtube  # PASS
showtime deliver exports <job> --targets github
showtime qa <job>/exports/final-4.github.mp4 --platform github   # PASS, 9.4 MB
showtime snap <job>/final-4.mp4 --at 0,6.9,11.0,12.6,13.3,13.8,21.3,27.4,29.5,32.6 --compare <job>/final-3.mp4

Iterations that mattered#

1. Contrast (check round 1). The first code theme (vitesse-dark) draws punctuation and comments at #666666 (3.1:1), and the focus dim (0.45) plus 35 % line numbers failed 4.5:1. Switched to houston (every token colour ≥ 8:1 on the panel), dim 0.7, line numbers at 62 %, lighter −/+ gutter colours.
2. System fonts. Geist Mono has no → glyph (U+2192) and a bare <code> fell back to Courier: the arrow is an inline SVG and code inherits the page font.
3. Readability holds. The removed lines needed 2 s before they collapse: diffs moved to 2.5 s / 1.8 s into their scenes; one right-column block was trimmed to fit its 7.2 s.
4. Glitch pick. Both were snapped mid-window; signal-glitch tears rows but keeps the code's structure, so it reads as "a bug", not as noise. It is the only glitch in the video.
5. Laptop layout (final → final-2). Pushed to 1.04x, the laptop ran under the label pill and touched the right edge; resized, and the terminal line height tightened so the last line clears the screen.
6. Audio (final-2 → final-3). The "break" under the terminal dropped the bed ~10 LU, so the music seemed to stop; section gain +4 dB, lighter key ducking, louder keys.
7. Review round 1 (final-3 → final-4). See the next section.

Review round 1#

A critic pass on final-3 said "ship after fixes": no blockers, two should-fixes and five polish notes. Each one was checked first against final-3's frames at its cited time (the review pack's frames, fresh snaps and a before/after sheet), and each held up. What changed in final-4, each checked with a full-res snap of the new render at the cited time and a before/after sheet against final-3:

• Should-fix, scene 5 (27.4 s / 29.5 s): the highlight boxes crossed out two of the three features. The boxes' percentages resolved against the padded card, not the image, so every box sat 12-15 px low and the top borders ran through the #4777 and #1879 lines. The image and boxes now share an unpadded wrapper and the boxes are set from the 1440x1195 excerpt (#4489 13.6 %/12.3 %, #4777 47.6 %/4.4 %, #1879 52.3 %/8.7 %). Each box now encloses its whole entry.

• Should-fix, scene 3 (10.4-14.8 s): the type-comment edit sat under the docstring headline. Lines 6-7 used to show red minus signs from the first frame of the scene. The scene now plays the changes in order: 25.12.0's file, with lines 6-7 unmarked; #4764 adds line 3 (11.7 s), which is highlighted; then "Change 3" arrives (13.2 s), change 2 steps back to 66 % opacity, and at 13.5 s lines 6-7 turn red and their fixed versions open green. Change 3 is on screen for 4.2 s. A full step back (42 %) made check flag the change-2 text as readable for only 3.1 s, so the step back is gentler.

• Polish, scene 2: the diff never showed − and + together. Both diffs now hold as unified diffs, with the removed lines in red above the added lines (the code panel is a little taller and uses a slightly smaller font, size 3.6, so 12 rows fit).

• Polish, 11.0 s: the gutter jumped 1, 2, 4, 5. The gutter now counts the file on screen: the old file's numbers before a change and the new file's after it. Removed lines get a minus.

• Polish, 21.2-21.5 s: short near-silence. It comes from the composed break itself: the bed alone dips to -40...-46 dBFS for about 0.6 s at a bar end, and ducking can't cause that. A soft thock (-8 dB) now fills it, landing with "Oh no! … 1 file would be reformatted." at 21.2 s. Checking this turned up a timing bug: output printed about 0.2 s before the Return key, while the caret was still on the command (snapped at 21.0 s on final-3). Every output line now appears just after its Return.

  I can't listen in this session, so this was judged only from metering.

• Polish, poster. Frame 0 stays the hook, since baking another frame there would flash. poster.jpg is now the end card at 32.6 s, where "black 26.1.0" is about 140 px tall and still reads at feed size. Upload it as the thumbnail wherever a platform takes one.

• Polish, end card. The note "newer 26.x releases keep the 2026 style" went from 2.6 to 3.1 cqh (about 33 px at 1080p).

Features shown#

new dom (retimed to 35 s) · showtime code --to → code-block (diff, highlight, dim, size; two panels swapped by a CSS step for a two-step diff) · ST.onSeek (a small page script that keeps removed lines on screen and renumbers the gutter) · device-frame (laptop, live HTML screen) · typewriter (uniform cadence; key clicks computed from it) · notifications (hook) · transitions wipe (linear, 3 cuts), signal-glitch (WebGL, once; CSS glitch snapped and rejected), stagger · site component (release-notes body, cropped) · assets emoji (Noto) · audio compose deep-house 122 bpm via the mix (sections on the cuts, section_gain, ducking under 60 key tracks) · audio sfx keyclick (4 variants), glitch, thock · audio mix report · check, snap (project and video), render --job with "render": {"crf": 18}, qa --platform, deliver exports --targets github,chat,x,linkedin, review-pack · researcher facts (every item → PR, date, author). Not used: retime -d (the brief's cut-down is covered by example 17), export html (not asked).

QA summary#

• final.mp4 (final-4): PASS with --platform youtube (0 fail, 0 warn): h264 High yuv420p, 1920x1080 30 fps, faststart, BT.709, 35.00 s, -14.1 LUFS, -1.1 dBTP, LRA 4.5, no silence, no black or frozen stretches, frame 0 flows into frame 1 (no poster flash), no attribution required. Against the job's github platform the master FAILs only file_size (18.0 MB vs the 10 MB cap): justified, since the master is the upload copy; the GitHub copy is the export below.
• exports/final.github.mp4 (final-4): PASS --platform github, 9.4 MB, -14.3 LUFS, -0.8 dBTP.
• Other exports made for final-3 (not shipped here, same picture): chat 9.4 MB, x 10.8 MB, linkedin 10.8 MB.
• check: PASS, 0 errors, 0 warnings (6 info notes, all about image, transition or mid-stagger layers).
• Critic: no sub-agent tool in this session, so the maker answered the critic brief itself (review/round-1-FINDINGS.md, verdict "ship", two should-fixes found on final-2 and fixed in final-3). A critic pass on final-3 followed. It found two should-fixes and five polish notes, all fixed in final-4 (see Review round 1). A human look is still welcome.

Sources and licenses#

Credits ship in credits.txt; the share copy with the label is share.txt.

The project files on GitHub