Overlay display duration is now computed server-side from estimated
reading time (text length / speaking rate), not from the audio
element's 'ended' event — edge-tts streams MP3 without a duration
header, and OBS's embedded browser fires 'ended' unreliably for that,
which is why the earlier event-based fix didn't hold up.
Per-streamer broadcasts are now serialized through a queue, so a
message waiting on an AI reply can no longer be overtaken by a later
message from a different viewer that doesn't need one.
Also fix the Google Sign-In button intermittently not rendering: the
GSI script tag is async/defer, and both index.html and send.html/js
called google.accounts.id.initialize() without waiting for it to
actually finish loading — a race that failed silently until the
script was cached from a previous load.
Dashboard now shows unique visitors (via sender link), total messages,
and unique senders per streamer, tracked via new viewer_visits/messages
tables.
The previous overlay TTS-overlap fix had a race: audio not having
arrived yet was treated the same as "no audio at all", letting the
queue advance before slow-to-synthesize audio (e.g. a long question)
ever started playing. Now "not arrived yet" is explicitly distinct
from "finished playing", with a grace-period fallback for genuine TTS
failures.
Also drop the hardcoded AI name from ai-prompt.txt since it's now
injected dynamically from the streamer's configured ai_name.
New profanity.js does whole-word matching against a Russian/English
root list (avoids false positives like "победа", "тебя", "сукно").
Applied once per message before it's used anywhere downstream
(display, TTS, AI reply input) — screen shows "..." in place of a
censored word, while the TTS text uses "бип" so it's heard as a beep
instead of read literally.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Previously the viewer's message displayed immediately and the AI
reply showed up whenever DeepSeek happened to respond, landing with
an awkward gap. Now, when the AI checkbox is checked, both texts are
fetched/prepared before anything is broadcast, so the original
message and the AI reply queue back-to-back on the overlay. Falls
back to just showing the original message if the AI call fails.
Behavior without the checkbox is unchanged (still immediate).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Viewers can check a box on the send page to get a short AI-generated
reply to their message, voiced with a distinct TTS voice and queued
on the overlay after the original message. The system prompt lives
in ai-prompt.txt (editable without touching code) instead of being
hardcoded. TTS for both viewer messages and AI replies now announces
who it's from ("Сообщение от X." / "Ответ от ИИ.") before the text.
No-ops entirely if DEEPSEEK_API_KEY isn't set.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
