Space Demo Video (space-demo-kit)

The kit renders a Gradio-style "app window" video: a hero intro where the input opens near full-frame with the prompt typing in large text (readable at mobile feed size), a morph into the app layout with the Space's real header (emoji, author/name, likes, hardware), a Generate press, a progress sweep, and the real output filling the output panel edge-to-edge. Text auto-fits its panels (short answers render huge, long ones stay readable). Branding is extracted from the Space page automatically. Output: 1440x1080 (4:3), 30 fps, ~12–17 s, H.264, silent unless the task itself involves audio.

Kit location on this machine: /root/code/space-demo-kit Render command (the ONLY way to produce the video):

cd /root/code/space-demo-kit && python3 render.py manifests/<name>.json
# -> out/<name>/<name>.mp4  +  out/<name>/<name>-contact.jpg

Golden rules

1. Never design video frames by hand. No PIL cards, no ffmpeg drawtext/overlay design, no custom HTML. The kit owns 100% of the design; you only supply facts and assets via a manifest JSON.
2. Real outputs. Generate the output by calling the live Space (gradio_client) with a real input. Only mock when live inference is impossible (quota, broken Space), and then label it honestly in the manifest meta field (e.g. "mock demo assets") and tell the user.
3. Real branding. Pull emoji, colorFrom, colorTo, likes, and hardware from https://huggingface.co/api/spaces/<id> (cardData + runtime.hardware.current). If the user gives a MODEL id, resolve its demo Space first (model page / spaces field of https://huggingface.co/api/models/<id>).
4. Prompts under 160 chars. The renderer truncates displayed prompts around 150 chars; pick or write inputs that fit.
5. Verify before delivering: ffprobe the MP4 (duration, 1440x1080) and look at the contact sheet (no overflowing text, output visible). Then send the MP4 to the user as media.
6. One look only. Never put a "style" key in a manifest — the app-window template is the only production design and the renderer ignores styles. Anything under experiments/ is a rejected prototype: never use it as an example or reference.

Asset quality bar (hard requirements)

• Inputs must be real content: a real photograph, illustration, or clip at least 512 px on the short side. NEVER an emoji, icon, favicon, logo, or clipart — tiny images upscale into blurry blobs and ruin the video. Verify every asset with file + dimensions before using it.
• Best input source: the Space's own example assets (the examples in its Gradio config, or files in the Space repo via https://huggingface.co/api/spaces/<id>/tree/main). They are known to work with the model.
• Before/after coherence (edit, a2a): the output MUST actually be derived from the input. Never pair two unrelated assets — an incoherent before/after is worse than no video. If live inference fails: retry via gradio_client (not hand-built HTTP calls), try another example input, and if it still fails, report the failure to the user and STOP. Do not fabricate the pair.

Template selection

Text-to-image Spaces use t2v: turn the generated image into a short static clip first (ffmpeg -loop 1 -i out.webp -t 6 -r 30 -pix_fmt yuv420p static.mp4) and pass it as video (see manifests/z-image-turbo-demo.json). There is no t2i template.

Manifest schema

Common fields (all templates):

{
  "name": "my-demo",                    // output folder + file name
  "template": "edit",
  "space": {
    "id": "owner/space-name",
    "emoji": "🧬",                      // from cardData
    "colorFrom": "blue",                // from cardData
    "colorTo": "gray",                  // from cardData
    "likes": 27,                        // from the API
    "hardware": "zero-a10g"             // runtime.hardware.current
  },
  "meta": "live output · zero-a10g",    // small honest footer note; use "mock demo assets" for mocks
  "button": "Generate",                 // optional; e.g. "Enhance", "Animate"
  "input_label": "source image",        // optional label overrides
  "prompt_label": "edit instruction",
  "output_label": "edited image",
  "audio_gain": 2.0,                    // optional; default 2.0 (with limiter)
  "hero": false                         // optional; set false to skip the full-frame intro
}

Plus the template-specific fields from the table. All asset paths are absolute paths on this machine. Store working assets under /root/code/<project-dir>/ as usual.

Workflow

1. curl -s https://huggingface.co/api/spaces/<id> → metadata for space{} fields.
2. Pick the template from the table. If the Space's task fits none of them, say so and pick the closest; do NOT invent a new design.
3. Capture assets: choose a good input (the Space's own example assets are ideal), run the live Space with gradio_client, download the real output. Verify the output file with ffprobe/file before rendering.
4. Write manifests/<name>.json in the kit.
5. Render (command above). First run downloads fonts/emoji (needs network) and, for video inputs, extracts frames. A render takes roughly 3–5 minutes — do not assume it hung.
6. Verify MP4 + contact sheet, then deliver the MP4 to the user as a media message.

Codex dispatch (OpenClaw — mandatory)

Demo-video requests are heavy (live inference + a ~2 min render). Do NOT run inference or rendering inline in the main session — it blocks you for many minutes. ALWAYS dispatch a codex session per the codex-monitor skill (setsid, </dev/null, CODEX_HOME pinned) and let the CODEX_FINISHED wake drive verification + delivery. Include in the codex prompt: "Read /root/code/space-demo-kit/skills/space-demo-video/SKILL.md and follow it exactly. Use the kit's render.py for ALL visuals; do not design any frame yourself."

Gotchas

• Long Space ids auto-shrink in the footer; long repo names get ellipsized in the header — fine.
• The renderer needs Playwright + Chromium (already installed system-wide here).
• Audio templates (avatar, a2a) mux the real audio automatically, delayed to the moment playback starts on screen, at 2x gain with a limiter.
• gallery wants exactly 4 outputs for the 2x2 grid.
• vqa answers auto-fit the output panel: keep them 30–60 words so the type stays large.
• Do not edit templates/app-window.html or render.py for a one-off video. If a Space genuinely needs a new mode, report that instead of hacking around the kit.