- Added detect_model_capabilities function that analyzes HF model tags and pipeline
- Auto-detects capabilities for Hugging Face models when adding via admin or --model
- Special handling for known models like Qwen VL, LLaVA, Whisper, etc.
- Updated UI with hints about auto-detection

- Added capabilities TEXT column to models table
- Updated admin interface to show and edit model capabilities
- Default model has 'video to text, image to text' capabilities
- Added Capabilities column to models table

- Added function to retrieve only available models for frontend/API use

- Added available BOOLEAN column to models table
- Default model is marked as available
- Models specified via --model are marked as available
- Updated admin interface to show and edit availability status
- Added Status column to models table showing Available/Hidden

- Added UNIQUE constraint on path column in models table
- Added cleanup query to remove duplicate models, keeping only one per path

- Modified init_db to check for existing default model before inserting
- Prevents multiple insertions of the same model

- Added ensure_model_exists function to database.py
- Modified vidai.py to call ensure_model_exists when --model is specified
- Automatically detects Hugging Face vs local models

- Added models table to database
- Created admin routes for model CRUD operations
- Added models.html template
- Added models link to admin dropdown
- Implemented Hugging Face model download functionality
- Added download_huggingface_model function in utils.py

Fix timezone issue by explicitly treating SQLite timestamps as UTC before converting to Unix timestamp

Fix last seen timestamp parsing to handle both ISO and SQLite timestamp formats for proper timezone display

Fix cluster status GPU detection to use available_backends instead of gpu_info for more reliable GPU client identification

Show connected cluster nodes in analyze page right panel, displaying hostname, IP, GPU status, weight, and last seen time

Add documentations dropdown menu in navbar, moving API documentation and admin API documentation links

Add admin dropdown menu in navbar for admin users, moving train, cluster tokens, cluster nodes, and configurations links

- Store connected_at as UTC ISO string with Z suffix
- Handle datetime objects from database as UTC timestamps
- Ensure uptime calculation uses correct UTC times regardless of server timezone

- Store connected_at as proper UTC timestamp in database using FROM_UNIXTIME/datetime
- Update web interface to handle datetime objects and timestamps correctly
- Ensure uptime starts from actual connection time, not offset by timezone

- Fix cluster master to properly detect GPU backends from client capabilities
- Extract available_backends from capabilities list instead of gpu_info
- Ensure clients with GPU workers are correctly identified as GPU-enabled

- Fix cluster client uptime calculation to start from 0 by using explicit UTC timestamps
- Fix cluster client workers not showing by populating cluster_master.clients dictionary
- Ensure connected_at uses current UTC time instead of database CURRENT_TIMESTAMP

- Get real worker process information from cluster master instead of placeholder data
- Display correct number of workers and their actual backends
- Improve accuracy of cluster node statistics

- Update connected_at timestamp on each successful connection
- Uptime now resets to 00:00:00 on reconnections as expected

- Extract real client IP address from websocket connection
- Preserve connected_at timestamp for accurate uptime calculation
- Send full GPU device info from client to master for proper VRAM reporting