Root cause: ai_provider.py assumed a single JSON blob in an
'encrypted_value' column, but IntegrationConfig stores each credential
as a separate row with config_key + config_value columns.
Fixes:
- Query groups rows by provider, decrypts each config_value separately
- Correct decrypt_value(secret_key, ciphertext) argument order
- Filter by user_id so providers are per-user
- Callers (ai_chat.py, chat()) now pass user_id through
AI coaching service (backend/app/services/ai_provider.py):
- Two code paths: OpenAI-compatible (6 providers) + Anthropic (Claude)
- Streaming SSE responses via httpx async
- System prompt built from AGENT_GUIDE + live user context
- Graceful error handling for all provider failures
- Gemini adapter for Google's generateContent API
Chat endpoint (backend/app/routers/ai_chat.py):
- POST /api/ai/chat — SSE streaming response
- GET /api/ai/status — check configured provider
- History capped at 20 messages, content at 4K chars
Provider registry expanded to 20 providers:
- 9 device integrations (strava, polar, garmin, whoop, oura, coros, fitbit, withings, suunto)
- 8 AI providers (openai, openrouter, huggingface, groq, ollama, claude, gemini, custom_ai)
- 3 other (usda, nutritionix, telegram)
- Categorized: device, ai_provider, nutrition, notifications
- COROS: MCP bridge approach (no API integration needed)
- Strava: AI/ML warning per their ToS
Frontend ChatCoach.jsx to follow in next commit.
