caption-conversation

name	caption-conversation
description	Show live, rolling captions of speech Jarvis hears — for accessibility, noisy rooms, or meetings — with optional translation. Use for "caption this", "turn on subtitles", "caption what they're saying", or "subtitle this meeting". Triggers: caption, captions, subtitles, transcribe live, what are they saying, accessibility, hard of hearing, meeting transcript.
license	MIT
compatibility	Requires JarvisVR agent-backend with ambient audio (perception.audio_scene). Protocol v1.1 §8.
metadata	{"agent":"communication-agent","category":"communication","version":"1.0","author":"jarvisvr"}
allowed-tools	show_live_caption

Caption Conversation

Render a live_caption panel that scrolls speech Jarvis hears in real time, tagged by speaker, optionally translated.

Steps

1. Ensure ambient audio is on: perception.request{stream:"ambient_audio", action:"start", reason:"user asked for live captions"}.
2. Open show_live_caption with an empty/seed lines array and the right speaker and max_lines.
3. Per audio window: append perception.audio_scene.ambient_transcript to lines (newest last), set speaker (user|other|jarvis|unknown), and holo.update. Trim to max_lines.
4. Translate (optional): if a target language is requested, translate each line, set translated:true and language (pair with live-translate).
5. Stop captions and the audio stream when the user is done.

Output

live_caption (show_live_caption, props per registry.json):

{ "widget_type": "live_caption",
  "transform": { "anchor": "head", "position": [0.0,-0.35,1.0], "billboard": true },
  "props": { "lines": [ "Hello there.", "How can I help you today?" ],
             "speaker": "other", "max_lines": 3, "language": "en", "translated": false },
  "interactions": ["grab","tap","resize"] }

Update as speech arrives (holo.update):

{ "object_id": "O_caption",
  "props": { "lines": [ "How can I help you today?", "I'd like the window seat." ], "speaker": "other" } }

Edge cases

• Multiple speakers → tag turns by speaker; "unknown" when diarization is uncertain.
• Background noise / no speech → keep the panel idle; don't caption music or ambient sounds (that's perception identify_sound).
• Long sessions → cap visible lines (max_lines ≤ 10); the full transcript can be saved as a note.
• Translated captions → set translated:true; consider a two-line layout (original + translation) via live-translate.
• Privacy → captioning others' speech is sensitive; it's user-initiated and perception.state reflects active capture. Stop promptly when asked.