English | 中文说明

Turn a system description (one JSON file) into a terminal-style, always-running architecture diagram, or a soft light-theme infographic that moves: fixed layout, packets flowing along the wires, a scrolling log, counters, bars that flip state, side triggers that light up in turn. Output is an H.264 mp4 (X / Xiaohongshu ready) or a live web page. It is also a Claude Code style skill (SKILL.md).

Credits - please read

• The idea, the look and the motion grammar come from an architecture-diagram clip by @thedelost (a Codex agent-tree panel, generated with GPT, screen-recorded as a web page): https://x.com/thedelost/status/2105398038026195279. It spread via a quote-post by @slashui: https://x.com/slashui/status/2105850132365443528.
• examples/codex-agents/ is a recreation of that original picture (layout, content and the three-tempo motion, redrawn from frame captures of the clip). The design belongs to @thedelost. The method (motion grammar, config-driven template) is re-implemented here and is not their code.
• examples/agent-architecture/ re-animates the static Xiaohongshu infographic 《AI Agent 的完整架构》 by 小红书 @林纾 (posted Sep 4). Layout, wording and colours follow the original; the design belongs to the original author. The footer of the video says so.
• examples/airbnb/ visualises figures Airbnb leaders stated in the Latent.Space interview https://www.latent.space/p/airbnb. The numbers are Airbnb's own account; the animation counters are illustrative.

terminal-dark, 4:5 - recreation of @thedelost's Codex agent tree	light-pastel, 3:4 - 林纾's AI Agent architecture

Airbnb example (terminal-dark, Chinese content, figures from the interview):

Videos: examples/codex-agents/codex-agents.mp4, examples/agent-architecture/agent-architecture.mp4, examples/airbnb/airbnb.mp4 (about 28-30 s each, 30 fps, silent AAC track so chat apps do not treat them as GIFs).

Requirements: Python 3.8+ (standard library only), Chrome or Chromium, ffmpeg. No pip packages.

python3 scripts/render.py --config examples/codex-agents/config.json --out out.mp4
python3 scripts/check_frames.py --config examples/codex-agents/config.json --out-dir frames --repeat

• render.py options: --width --height --duration --fps (default from the config), --chrome --ffmpeg (default: found on PATH), --html-out page.html (keep the self-contained live page), --keep-frames DIR, --audio none, --crf.
• check_frames.py samples ~120 time points, measures text overflow / overlap from the DOM, writes a few PNGs and (with --repeat) re-renders each exported frame after seeking away to prove it is identical. Exit status 1 on any problem.
• To use as a skill, put this folder where your agent loads skills (for Claude Code: ~/.claude/skills/live-panel/).

1. Copy an example config. 2. Edit text, numbers, colours, positions. 3. Render, run check_frames.py, look at the PNGs. Nothing in assets/template.html has to change.

"canvas": { "preset": "3:4", "duration": 28, "fps": 30, "preroll": 0 },   // or width/height explicitly
"theme":  { "preset": "light-pastel", "colors": { "pk": "#f0575f" } }      // preset + your overrides

Positions in a config are absolute canvas pixels, so a config is authored for one canvas; render.py --width/--height only rescales (letterboxes) it. For another ratio, copy the config and re-place the boxes.

A theme controls: font and size, line height, the colour table (colors, any names you like), box mode (segmented or solid), corner radius, border width, wire width, packet size and glow, background. Add your own preset by editing THEME_PRESETS at the top of the boot code in the template, or override per config.

Full tables are in references/config-schema.md.

• elements: text, box (fill, border colour, radius, dash, container, per-line items/bar/runs), line, path (polyline with rounded corners, arrow head, dash, when/then colour change, embedded packets), glyph, rule, flow (packets along a polyline, optional when), tarrow (trigger arrows), log.
• machines (everything that changes, all pure functions of time): counter, cycle, gauge, any_low, lane, triggers.
• when / then on boxes, paths, flows, lines and runs: highlight things from machine variables ({"var":"seq.i","in":[0,1,2]}; a list of conditions means AND).
• Motion rules: references/motion-grammar.md.

The page exposes window.seek(t). Every visual is a pure function of t; the few "random" picks use a fixed-seed integer hash; no wall clock, no Math.random, no CSS animation. The renderer calls seek(i/fps) and screenshots. In a normal browser (no ?manual) the same function is driven by requestAnimationFrame.

No font files are bundled. The stacks fall back through installed fonts: terminal themes use JetBrains Mono (OFL) -> IBM Plex Mono (OFL) -> DejaVu Sans Mono -> Noto Sans Mono CJK SC (OFL) -> any monospace; light-pastel uses Noto Sans CJK SC (OFL) / PingFang SC / Microsoft YaHei -> Helvetica/Arial, and Noto Serif CJK SC for the big title. Install JetBrains Mono and Noto CJK for the exact look in the examples; emoji icons need a colour emoji font (e.g. Noto Color Emoji). Different fonts change glyph widths slightly, so re-run check_frames.py after changing fonts.

• All three examples rendered with scripts/render.py; check_frames.py reports 0 problems on ~120 sampled time points each. (It does catch real faults: the first Airbnb layout had a log time column overlapping the next column and was flagged.)
• Determinism: the Codex example rendered twice gives byte-identical decoded frames for all 900 frames (ffmpeg -f framemd5 hashes equal); same for the agent-architecture example; check_frames.py --repeat re-screenshots after seeking away and compares bytes.
• Fidelity of the Codex recreation against frame captures of the original (side by side at the same moment): box positions and sizes within a few pixels; same four-colour code, same three-tempo behaviour (packets on every wire, trigger rotation with typed advice and calls/tokens accumulation, gauges flipping sharp/split with the legend swapping, spinners and done states, 4-row log with newest row bright, forks counter ~16/s, prompt cursor). Differences: the original's font is a slab-like monospace and ours is JetBrains Mono; the original's trigger arrows are thin solid, ours dashed; bar texture, packet glow and trails are close but not identical; the original's gauge values, log wording and timing were read by eye from 2 fps frames, so the sequences differ; the third advice text (before done) was not legible in the frames and is a placeholder written for this repo; log time runs at an assumed 3x.

MIT for the code (see LICENSE). The visual design that examples/codex-agents/ recreates belongs to @thedelost, and the infographic re-animated in examples/agent-architecture/ belongs to 小红书 @林纾; those examples are credit-bearing recreations and are not covered by the MIT grant. Airbnb figures are Airbnb's own statements from a public interview.

把一份系统描述（一个 JSON）变成终端风格、一直在运行的动态架构图，或者一张会动的浅色粉彩信息图：版面不动，连线上有光点流动，日志滚动，计数跳动，进度条过阈值翻状态，侧栏触发点依次点亮。产出 H.264 mp4（适合发 X 和小红书）或可直接打开的网页。

出处：动法和样式学自 @thedelost 的 Codex agent 分工动态图（https://x.com/thedelost/status/2105398038026195279），经 @slashui 引用转发（https://x.com/slashui/status/2105850132365443528）。examples/codex-agents/ 是对原图的复刻，设计归原作者；examples/agent-architecture/ 把小红书 @林纾 的《AI Agent 的完整架构》静态图做成动态版，原图设计归林纾，画面底部有出处；examples/airbnb/ 的数字出自 Latent.Space 访谈，为 Airbnb 自述，画面里跳动的计数为示意。

用法：python3 scripts/render.py --config <配置> --out out.mp4；自查：python3 scripts/check_frames.py --config <配置> --out-dir frames --repeat。主题用 theme.preset（terminal-dark / light-pastel），画幅用 canvas.preset（4:5 1200x1500、3:4 1080x1440、1:1 1080x1080）。坐标是画布像素，换画幅需要另存一份配置重排。