Customize the composition

Turn the content you wrote into a validated, render-ready bundle. This step catches mistakes (wrong color format, a value that doesn't match a declared variable) locally, so the render API never sees a bad payload.

You don't edit the composition's HTML. HyperFrames injects your variable values at render time — the composition reads them with window.__hyperframes.getVariables(). So "customizing" means producing a clean variables.json and staging the composition directory.

Steps

1. Make sure your generated content is at /.agents/workspace/output/proposed.json (from generate-script).

2. Validate it against the chosen composition and write the clean variables:

   python3 /.agents/workspace/scripts/customize.py \
     /.agents/workspace/compositions/<chosen-id> \
     /.agents/workspace/output/proposed.json \
     /.agents/workspace/output/variables.json
   

   • On success it writes /.agents/workspace/output/variables.json (your values plus declared defaults for anything you left out) and exits 0.
   • On failure it exits non-zero and prints exactly what's wrong (unknown variable, wrong type, a color that isn't a hex value, a number out of range). Read the message, fix proposed.json, and run it again. Do not skip ahead to rendering with an invalid file.

3. Copy the chosen composition into the output area so the render step has a self-contained directory:

   rm -rf /.agents/workspace/output/composition
   cp -r /.agents/workspace/compositions/<chosen-id> /.agents/workspace/output/composition
   

4. Bake in the voiceover. The renderer reads the <audio> src from static HTML, so substitute the synthesized URL into the staged composition (or strip the audio element if there's no narration):

   python3 /.agents/workspace/scripts/apply_voiceover.py \
     /.agents/workspace/output/composition \
     "$(cat /.agents/workspace/output/voiceover_url.txt 2>/dev/null)"
   

   Free-composition case: the shipped starters already have the voiceover <audio> slot. But if you free-authored the composition's HTML (custom video, not a starter), you MUST include the same slot or the video renders silent — apply_voiceover.py only substitutes a slot that exists. Put this inside the composition root, near the other timed elements, with the literal placeholder src (apply_voiceover fills it; it's stripped if there's no narration). Use a real, literal data-duration in seconds — your composition's own length — e.g. for a 10-second video:

   <audio id="voiceover" data-start="0" data-duration="10"
          data-track-index="0" data-volume="1" src="__VOICEOVER_URL__"></audio>
   

   • The ONLY placeholder is src="__VOICEOVER_URL__" (apply_voiceover substitutes it). Every other attribute must be a concrete value — do not leave a literal like <comp-duration> in the HTML; the renderer parses data-duration as a number, and a non-number is silently dropped → silent video. (You may also omit data-duration entirely; it then defaults to the audio's own length.)
   • Do NOT use data-has-audio or data-composition-variable — the renderer reads a static src; the placeholder + apply_voiceover is the working path.

Multi-turn edits

On a follow-up ("make the headline punchier", "switch to a dark background"), you don't need to start over:

• Edit only the changed keys in /.agents/workspace/output/proposed.json.
• Re-run customize.py to re-validate and rewrite variables.json.
• Keep the same staged composition unless the user wants a different starter.

Output of this step

• /.agents/workspace/output/variables.json — validated values.
• /.agents/workspace/output/composition/ — the staged composition directory.

Then move to render-and-return.