Repository navigation
V1.4.0 - Activities & Attributes
·
25 commits
to master
since this release
[1.4.0] - 2026-03-13
Added
Activities System
- Icebreaker: Character asks a personality-driven question with 10 randomized styles, answer to start a timed chat
- Two Truths & A Lie: Character generates 3 statements (2 truths, 1 lie), guess the lie then chat about it
- This or That: 7 rounds of rapid-fire preference pairs, then timed chat discussing choices
- Would You Rather: 5 rounds of fun dilemma scenarios, then timed chat about decisions
- Activity Game Context Persistence: Game results (answers, choices, guesses) now stored on conversations and injected into every chat prompt, so the AI remembers what happened during the game
- Activity-Specific Chat Prompts: Each activity mode gets a unique chat context instead of generic "random chat"
- Activities Hub: Designed with cards for all activity types
- Activities Settings Page: Configure character selection options (include away/busy characters)
- Character First Message Chance: Configurable probability for who messages first in activities
Random Chat & Matching
- Random Chat Feature: Timed chat with random unmatched characters, option to match after timer
- Blind Date Mode: Hidden identity chat where only first initial is shown
- Proactive First Message: Characters can send the first message in activity chats
Character Attributes
- Configurable Attribute System: Replaced Big Five personality with fully customizable character attributes
- Attribute-Driven Behavior: Character attributes included in schedule generation and chat prompts
LLM Configuration
- Proactive LLM Config: Independent LLM settings for proactive message generation
- Random Model Pool: Configure a pool of random models for any LLM type
- Double Text Chance: Configurable probability for characters to send follow-up messages
Chat Improvements
- Image Gallery: View all images in a conversation with navigation
- Early Reactions: Characters can react to messages earlier in the conversation
- Typing Indicator Alignment: Improved positioning of typing indicators
Fixed
- Avatar rotation and gallery not updating when messages with images are deleted
- Asterisk words being incorrectly removed (e.g., "I really like that" losing the word)
- Stale model selector in LLM settings after switching providers
- Daily auto-match enable not working
- Smiley/emoji stripping in messages
- Image rotation not showing images from older paginated messages
- Message pagination being reset during conversation
- Autoscroll reliability improvements
- Double text when response already has multiple messages
- Several performance optimizations