Backend changes for full blueprint feature parity:
Database models:
- MessageTemplate: add body_ru (locale variant), event_type field
- AlbumTracker: add track_images, track_videos, notify_favorites_only,
include_people, include_asset_details, max_assets_to_show,
assets_order_by, assets_order fields
- New ScheduledJob model: supports periodic_summary, scheduled_assets,
and memory (On This Day) notification modes with full config
(times, interval, album_mode, limit, filters, sorting)
API:
- New /api/scheduled/* CRUD endpoints for scheduled jobs
- Tracker create/update accept all new filtering fields
- Tracker response includes new fields
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes 7 issues identified by code-reviewer agent:
1. (Critical) Webhook endpoint now validates X-Telegram-Bot-Api-
Secret-Token header against configured secret, and verifies
bot_token matches a stored NotificationTarget
2. (Critical) register/unregister webhook endpoints now require
JWT auth via Depends(get_current_user); register passes
secret_token to Telegram setWebhook
3. (Critical) Conversation dict now uses OrderedDict with LRU
eviction (max 100 chats); trim happens BEFORE API call
4. (Important) Tool-use responses no longer stored in conversation
history (avoids corrupted multi-turn state)
5. (Important) Singleton AsyncAnthropic client (module-level,
lazily initialized once) - no more connection pool leaks
6. (Important) Markdown retry now uses payload.pop("parse_mode")
instead of setting empty string
7. (Important) All user-controlled data wrapped in <data> tags
with _sanitize() helper (truncation + newline stripping)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Integrate Claude AI into the notification system for intelligent
conversational interactions and AI-powered captions.
New modules:
- ai/service.py: Claude API client with conversation history,
caption generation, and album activity summarization
- ai/telegram_webhook.py: Telegram webhook handler for incoming
bot messages, routes to AI service for responses
Features:
- Conversational bot: users chat with the bot about albums
- AI captions: intelligent notification messages based on album
context (people, locations, dates) - enabled per target via
"ai_captions" config flag
- Album summaries: "what's new?" triggers AI-generated overview
- /start command with welcome message
- Webhook register/unregister endpoints
Architecture:
- Per-chat conversation history (in-memory, capped at 20 messages)
- Graceful degradation: AI features completely disabled without
IMMICH_WATCHER_ANTHROPIC_API_KEY env var (zero impact)
- AI caption failure falls back to Jinja2 template rendering
- Health endpoint reports ai_enabled status
Config: IMMICH_WATCHER_ANTHROPIC_API_KEY, IMMICH_WATCHER_AI_MODEL,
IMMICH_WATCHER_AI_MAX_TOKENS
Server now has 45 API routes (was 42 after Phase 5).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Build a complete standalone web server for Immich album change
notifications, independent of Home Assistant. Uses the shared
core library from Phase 1.
Server features:
- FastAPI with async SQLite (SQLModel + aiosqlite)
- Multi-user auth with JWT (admin/user roles, setup wizard)
- CRUD APIs: Immich servers, album trackers, message templates,
notification targets (Telegram + webhook), user management
- APScheduler background polling per tracker
- Jinja2 template rendering with live preview
- Album browser proxied from Immich API
- Event logging and dashboard status endpoint
- Docker deployment (single container, SQLite in volume)
39 API routes, 14 integration tests passing.
Also adds Phase 6 (Claude AI Telegram bot enhancement) to the
primary plan as an optional future phase.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
