fix(tts): make the ElevenLabs voice id field drive the voice that plays

[dyascj]

Fixes #153.

Root cause

Selecting ElevenLabs (settings dropdown or onboarding) sets speech.activeVoiceId to the first declared voice, Rachel 21m00Tcm4TlvDq8ikWAM. The "Custom Voice ID" box wrote to a different slot, providerConfigs.elevenlabs.voiceId, and the pipeline resolved activeVoiceId || ttsConfig.voiceId. Rachel is never empty, so every pasted id was shadowed. Reproduced against a mock ElevenLabs with the reporter's exact stored state: three requests to text-to-speech/21m00Tcm4TlvDq8ikWAM/stream while the custom id sat unused. That matches "ElevenLabs receives the requests, the voice does not change".

Change

• The ElevenLabs voice field edits activeVoiceId through handleTTSVoiceChange, the same path every other provider uses, with the declared voices as a native datalist. Onboarding twin gets the same treatment.
• One-time migration at startup (legacy-voice-migration.ts, decision in legacyVoiceToAdopt): if a legacy custom id exists and the active voice is empty or still the provider default, it moves across. The legacy slot is cleared either way so the migration can never fire twice or apply an ElevenLabs voice to another provider. Users who had a legacy id but are on a different TTS provider lose an id that was never in effect and will see the real slot when they switch back.
• Pipeline reads one slot again; the dead getter and setter are gone from the settings store.
• Second symptom from the screenshots: the TTS settings state now seeds its model list from the cache on init, so ElevenLabs shows the saved model instead of "Select model..." until someone hits refresh.

Verification

• pnpm test 440 pass (5 new for legacyVoiceToAdopt), pnpm check 0 errors 0 warnings.
• Live against the real ElevenLabs API on a Creator account:
  • Migration: seeded activeVoiceId Rachel plus a legacy custom id, reloaded, activeVoiceId became the custom id and the legacy slot was removed, API key kept.
  • Playback: chat turn produced three text-to-speech/JBFqnCBsd6RMkjVDRZzb/stream calls (George, not in our declared list), all 200 audio/mpeg.
  • Fresh user: picked ElevenLabs from the dropdown (voice resets to Rachel, field shows it), pasted the custom id, activeVoiceId updated, model dropdown showed "Eleven Turbo v2.5" after refresh and again after a full reload without refresh.
  • Invalid id: request goes out with the typed id, ElevenLabs returns 400, the chat bar shows "ElevenLabs error 400: An invalid ID has been received for voice". Before this fix a bad id never left the machine.
  • Onboarding: Services step field shows the current voice, typing writes it trimmed, datalist present.

Notes

• Touches the same voiceId: line in companion-chat.ts as Feature/multilingual tts #154; whichever merges second gets a one-line conflict.
• Not in scope, worth their own issues: the desktop "Download Save File" button in the reporter's fourth screenshot does nothing under WebView2 (db/export.ts uses an anchor download), and ElevenLabsTTS defaults to Bella when the voice is empty while the registry's first voice is Rachel.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
utsuwa	Ready	Preview	Aug 25, 2026 1:58pm