- Add MercadoLibre OAuth, listings, orders, webhooks and category search
- New marketplace_external_bp.py, meli_service.py, marketplace_external_service.py
- New marketplace_external.html/js with ML management UI
- Inventory: bulk publish to ML with category autocomplete, listing type and shipping selectors
- Inventory: new .btn--meli styles, select/label CSS fixes
- WhatsApp bridge: rate limiting, 440/515/408 error handling, stale watchdog
- DB migration v3.4_meli_integration.sql for marketplace_listings, orders, sync_queue
- Add Celery tasks for ML sync and webhook processing
- Sidebar: MercadoLibre navigation link
- Add QWEN (qwen3.6) as primary AI backend with short system prompt
- Hermes remains as fallback with 45s timeout
- Increase QWEN timeout to 35s, max_tokens to 4000
- Add conversation history loading from whatsapp_messages (last 4 msgs)
- Persist detected vehicle in whatsapp_sessions table
- Add 'limpiar chat' / 'nuevo chat' / 'reset' commands to clear history
- Fix CSS conflict: rename whatsapp chat-panel classes to wa-chat-panel
- Fix JS ID conflicts with chat.js widget (waChatPanel, waChatMessages, etc.)
- Improve no-stock response: conversational with alternatives
- Split search_query by | for multi-part lookups
- Add DEMO_PROMPTS.md and DEMO_PROMPTS_V2.md
Cambios implementados:
1. Connection pooling (tenant_db.py):
- psycopg2.pool.ThreadedConnectionPool para master y tenants
- Wrapper _PooledConnection que devuelve al pool en .close()
- Cero cambios en blueprints (backward compatible)
2. Tabla inventory_stock_summary + triggers (v3.2):
- O(1) stock lookup en vez de SUM() sobre historial completo
- Trigger AFTER INSERT en inventory_operations recalcula stock
- Poblada inicialmente en ambos tenants
- Refactor en 6 archivos de servicios para usar la nueva tabla
3. Fix N+1 en process_sale (pos_engine.py):
- Precarga retail_price en bulk query FOR UPDATE
- Elimina SELECT individual por item en loop
4. Índices críticos:
- idx_parts_name_part + pattern_ops (master)
- idx_inv_ops_inventory_branch_created (tenants)
- idx_wi_part_stock_positive (master, ya existía desde Fase 1)
Tests: 73/73 pasando (compat + fase3 + fase5 + fase6)
Migración: v3.2_db_performance.sql
- Kiosk mode: fullscreen, wake lock, auto-login, context menu block, PWA/Capacitor detection
- AI vision: camera photos analyzed by Gemma 3 27B vision model via OpenRouter
- AI part classification: auto-suggest name/brand/category when entering part number
- Public catalog chatbot: /api/chat endpoint with rate limiting, chat widget on catalog page
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Chat now fetches tenant inventory summary (brands, counts, low-stock)
and injects it into the AI system prompt so responses prioritize local
stock. VIN decoder uses free NHTSA vPIC API to decode 17-char VINs and
auto-fills the vehicle selector dropdowns when a catalog match is found.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Symptom diagnosis: AI now detects vehicle symptoms and suggests probable parts
2. Smart quotations: "cotizame frenos completos" returns multiple search_queries (pipe-separated), backend searches each term and deduplicates
3. Photo identification: camera button in chat widget uploads image and sends placeholder prompt (ready for vision model)
4. Multilanguage: AI detects user language and responds accordingly, search_query always in English
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
3 reintentos con espera 5s/10s/15s. Si agota reintentos,
muestra mensaje amigable en vez de error.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Validacion _validate_model() verifica que el modelo termine en ':free'.
Si alguien cambia MODEL a uno de pago, el chatbot lanza error en vez
de generar costos. Alternativas documentadas en comentario.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- System prompt: SIEMPRE devuelve search_query en ingles
- Diccionario de traducciones (balatas→Brake Pad, etc.)
- Busca directo sin preguntar mas info
- Fallback: extrae keywords si AI no da search_query
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
