- Voice clone integration: When user selects voice clone in Write phase,
backend uses their uploaded voice sample + scene script text to generate
audio via qwen3/minimax/cosyvoice voice clone APIs
- Multi-tenant workspace storage: All podcast assets (audio, video, images,
charts) now use workspace-specific directories per user
- Chart preview improvements: Card-based B-Roll charts UI with thumbnails,
takeaway text, and action buttons; public endpoint for image serving
- Voice clone caching: In-memory LRU cache for voice samples (avoids
re-downloading per scene); frontend caches voice clone metadata
- Thread pool for voice clone: Audio generation uses ThreadPoolExecutor to
avoid blocking the FastAPI event loop
- Auto-detect voice clone IDs (vc_*, MY_VOICE_CLONE) to route correctly
- DB fallback for voice sample URL: Fetches from ContentAsset if not passed
- Fixed API URL resolution for chart previews
- Fixed GlassyCard DOM warnings for motion props
- Fixed ScriptGenerationProgressView syntax error
- Fixed usePodcastWorkflow scriptData reference
- Add pandas to requirements-podcast.txt for usage tracking
- Fix LLM prompt to return plain strings instead of objects for enhanced_ideas
- Add object-to-string normalization for LLM responses that return objects
- Skip bible generation in podcast-only mode (onboarding disabled)
- Skip alerts polling in AlertsBadge when in podcast-only demo mode
- Remove hardcoded preferred_provider=huggingface in podcast handlers
- Set preferred_provider=None to respect GPT_PROVIDER env var
- Change default model from Qwen to gpt-oss-120b:cerebras (the model user had access to)
- WaveSpeed will now use gpt-oss-120b model instead of Qwen
