# Changelog

All notable changes to Konstruct are documented in this file.

## [1.0.0] - 2026-03-26

### Phase 10: Agent Capabilities
- Knowledge base ingestion pipeline — upload PDF, DOCX, PPTX, XLSX, CSV, TXT, Markdown; add URLs (Firecrawl scraping); add YouTube videos (transcript extraction)
- Async document processing via Celery — chunk, embed (all-MiniLM-L6-v2), store in pgvector
- KB management portal page with drag-and-drop upload, live status polling, delete, reindex
- Google Calendar OAuth per tenant — list events, check availability, create events
- Token auto-refresh with encrypted DB write-back
- Web search connected to Brave Search API (platform-wide key)
- Tool executor injects tenant_id/agent_id into all tool handlers
- System prompt includes tool result formatting instruction (no raw JSON)

### Phase 9: Testing & QA
- Playwright E2E test suite — 29 tests across 7 critical flows (login, tenants, agent deploy, chat, RBAC, i18n, mobile)
- Cross-browser testing — Chromium, Firefox, WebKit
- Visual regression snapshots at 3 viewports (desktop, tablet, mobile)
- axe-core accessibility scans on all pages
- Lighthouse CI score gating (>= 80 hard floor)
- Gitea Actions CI pipeline — backend lint + pytest → portal build + E2E + Lighthouse

### Phase 8: Mobile + PWA
- Responsive mobile layout with bottom tab bar (Dashboard, Employees, Chat, Usage, More)
- Full-screen WhatsApp-style mobile chat with back arrow + agent name header
- Visual Viewport API keyboard handling for iOS
- PWA manifest with K monogram icons
- Service worker (Serwist) with app shell + runtime caching
- Web Push notifications (VAPID) with push subscription management
- IndexedDB offline message queue with drain-on-reconnect
- Smart install prompt on second visit
- iOS safe-area support

### Phase 7: Multilanguage
- Full portal UI localization — English, Spanish, Portuguese
- next-intl v4 (cookie-based locale, no URL routing)
- Language switcher in sidebar (post-auth) and login page (pre-auth)
- Browser locale auto-detection on first visit
- Language preference saved to DB, synced to JWT
- Agent templates translated in all 3 languages (JSONB translations column)
- System prompt language instruction — agents auto-detect and respond in user's language
- Localized invitation emails

### Phase 6: Web Chat
- Real-time WebSocket chat in the portal
- Direct LLM streaming from WebSocket handler (bypasses Celery for speed)
- Token-by-token streaming via NDJSON → Redis pub-sub → WebSocket
- Conversation persistence (web_conversations + web_conversation_messages tables)
- Agent picker dialog for new conversations
- Markdown rendering (react-markdown + remark-gfm)
- Typing indicator during LLM generation
- All roles can chat (operators included)

### Phase 5: Employee Design
- Three-path AI employee creation: Templates / Guided Setup / Advanced
- 6 pre-built agent templates (Customer Support Rep, Sales Assistant, Marketing Manager, Office Manager, Project Coordinator, Finance & Accounting Manager)
- 5-step wizard (Role → Persona → Tools → Channels → Escalation)
- System prompt auto-generation from wizard inputs
- Templates stored as DB seed data with one-click deploy
- Agent Designer as "Advanced" mode

### Phase 4: RBAC
- Three-tier roles: platform_admin, customer_admin, customer_operator
- FastAPI RBAC guard dependencies (require_platform_admin, require_tenant_admin, require_tenant_member)
- Email invitation flow with HMAC tokens (48-hour expiry, resend capability)
- SMTP email sending via Python stdlib
- Portal navigation and API endpoints enforce role-based access
- Impersonation for platform admins with audit trail
- Global user management page

### Phase 3: Operator Experience
- Slack OAuth "Add to Slack" flow with HMAC state protection
- WhatsApp guided manual setup
- 3-step onboarding wizard (Connect → Configure → Test)
- Stripe subscription management (per-agent $49/month, 14-day trial)
- BYO API key management with Fernet encryption + MultiFernet key rotation
- Cost dashboard with Recharts (token usage, provider costs, message volume, budget alerts)
- Agent-level cost tracking and budget limits

### Phase 2: Agent Features
- Two-layer conversational memory (Redis sliding window + pgvector HNSW)
- Cross-conversation memory keyed per-user per-agent
- Tool framework with 4 built-in tools (web search, KB search, HTTP request, calendar)
- Schema-validated tool execution with confirmation flow for side-effecting actions
- Immutable audit logging (REVOKE UPDATE/DELETE at DB level)
- WhatsApp Business Cloud API adapter with Meta 2026 policy compliance
- Two-tier business-function scoping (keyword allowlist + role-based LLM)
- Human escalation with DM delivery, full transcript, and assistant mode
- Cross-channel bidirectional media support with multimodal LLM interpretation

### Phase 1: Foundation
- Monorepo with uv workspaces
- Docker Compose dev environment (PostgreSQL 16 + pgvector, Redis, Ollama)
- PostgreSQL Row Level Security with FORCE ROW LEVEL SECURITY
- Shared Pydantic models (KonstructMessage) and SQLAlchemy 2.0 async ORM
- LiteLLM Router with Ollama + Anthropic/OpenAI and fallback routing
- Celery orchestrator with sync-def pattern (asyncio.run)
- Slack adapter (Events API) with typing indicator
- Message Router with tenant resolution, rate limiting, idempotency
- Next.js 16 admin portal with Auth.js v5, tenant CRUD, Agent Designer
- Premium UI design system (indigo brand, dark sidebar, glass-morphism, DM Sans)