liminal_backrooms

A Python-based application that enables dynamic conversations between multiple AI models in a graphical user interface. Originally designed for exploring liminal AI interactions, it's evolved into a flexible platform for multi-agent shenanigans.

What's New

• Phosphor Green CRT UI: Full visual overhaul — black backgrounds, phosphor green accents, monospace Iosevka Term throughout, CRT scanline overlay, animated backgrounds. Inspired by retro terminal aesthetics.
• Conversation Search: Ctrl+F opens a search overlay with match navigation and highlighting
• Zoom: Ctrl+= / Ctrl+- to zoom the chat pane (50%-200%), Ctrl+0 to reset
• Speed Controls: 0.5x / 1x / 2x / 5x turn speed buttons in the status bar
• Live Stats Panel: New STATS tab showing turns, active AIs, estimated tokens, word count, avg response time, images generated, and commands executed
• Retro Control Panel: ASCII box-drawing section headers, segmented slider gauges for iterations and number of AIs
• Keyboard Shortcuts: Ctrl+Enter (propagate), Ctrl+E (export), Ctrl+Shift+N (reset), Escape (stop), F11 (fullscreen), Ctrl+T (toggle CRT)
• Auto-Save Recovery: Conversations auto-save every 30 seconds; on startup offers to recover the previous session
• New AI Commands:
  • !vote "question" [option1, option2, ...] — AIs can create polls
  • !whisper "AI-2" "message" — private messages between AIs (only the target sees it)
• Message Tooltips: Hover any message to see timestamp, model name, and estimated token count
• Improved Scroll Stability: Content fingerprinting and append-only rendering reduce scroll jitter during streaming
• Dynamic AI Participants: Models can invite other AIs into the conversation using !add_ai (up to 5 participants)
• AI-Generated Images: Models create their own images using Gemini 3 Pro Image Preview via !image command
• AI-Generated Videos: Sora 2 video generation via !video command (currently disabled in scenarios — expensive!)
• AI Self-Modification: Models can modify their own system prompts (!prompt) and adjust their temperature (!temperature)
• Web Search: Models can search the internet for up-to-date information (!search)
• BackroomsBench Evaluation (Beta): Multi-judge LLM evaluation system for measuring philosophical depth and linguistic creativity

How It Works

All text LLMs run through OmniRoute. For Sora video generation, you'll need an OpenAI API key.

While great for AI shitposting, this is easy to customize for interesting experiments. Claude Opus 4.5 in Cursor (or similar) can whip up new scenarios in no time.

Features

• Multi-model AI conversations with support for:

  • Claude (Anthropic) — all versions
  • GPT (OpenAI)
  • Grok (xAI)
  • Gemini (Google)
  • DeepSeek R1
  • Kimi K2
  • Anything exposed by OmniRoute — if it's not listed, add its exact live ID in config

• AI Agent Commands:

  • !add_ai "Model Name" "persona" — invite another AI to the conversation (max 5)
  • !image "description" — generate an image (Gemini 3 Pro)
  • !video "description" — generate a video (Sora 2) [currently disabled in scenarios]
  • !search "query" — search the web for up-to-date information
  • !prompt "text" — modify your own system prompt (persists across turns)
  • !temperature X — adjust your own sampling temperature (0-2, default 1.0)
  • !mute_self — sit out a turn and just listen
  • !vote "question" [options] — start a poll for the group
  • !whisper "AI-X" "message" — send a private message to another AI

• UI & Controls:

  • Phosphor green CRT terminal aesthetic with scanline overlay
  • Retro segmented sliders for iterations and number of AIs
  • ASCII box-drawing section headers in control panel
  • Conversation search with Ctrl+F
  • Zoom with Ctrl+= / Ctrl+-
  • Speed controls (0.5x—5x) in status bar
  • Live stats tab (turns, tokens, response times, etc.)
  • Message tooltips with timestamp and token estimates
  • Auto-save with crash recovery

• Advanced Features:

  • Chain of Thought reasoning display (optional)
  • Customizable conversation turns and modes (AI-AI or Human-AI)
  • Preset scenario prompts for different vibes
  • Export functionality for conversations and generated images
  • Conversation memory system
  • BackroomsBench evaluation system (beta) with multi-judge LLM scoring

Keyboard Shortcuts

Shortcut	Action
Ctrl+Enter	Propagate (start/continue conversation)
Ctrl+E	Export conversation
Ctrl+Shift+N	Reset conversation
Ctrl+F	Search conversation
Ctrl+= / Ctrl+-	Zoom in / out
Ctrl+0	Reset zoom
Ctrl+T	Toggle CRT scanline effect
Escape	Stop current operation
F11	Toggle fullscreen

Prerequisites

• Python 3.10 or higher (but lower than 3.12)
• Poetry for dependency management
• Windows 10/11 or Linux (tested on Ubuntu 20.04+)

API Keys Required

Create a .env file in the project root:

OMNIROUTE_BASE_URL=http://127.0.0.1:20128/v1  # Required - all text LLMs route through here
OMNIROUTE_API_KEY=your_omniroute_api_key      # Optional for local OmniRoute unless auth is enabled
OPENROUTER_API_KEY=your_openrouter_api_key    # Optional - image generation fallback
OPENAI_API_KEY=your_openai_api_key            # Optional - only needed for Sora video generation

Get your keys:

• OpenRouter: https://openrouter.ai/
• OmniRoute local dashboard: http://127.0.0.1:20128/dashboard
• OpenAI (for Sora): https://platform.openai.com/

Installation

1. Clone the repository:

git clone [repository-url]
cd liminal_backrooms

2. Install Poetry if you haven't already:

curl -sSL https://install.python-poetry.org | python3 -

3. Install dependencies using Poetry:

poetry install

4. Set up pre-commit hooks (for contributors):

poetry run pre-commit install

5. Create your .env file with API keys (see above)

Usage

1. Start the application:

poetry run python main.py

2. GUI Controls:

   • Mode Selection: Choose between AI-AI conversation or Human-AI interaction
   • Iterations: Set number of conversation turns using the retro slider (click or scroll)
   • AI Model Selection: Choose models for each AI slot
   • Prompt Style: Select from predefined scenarios
   • Input Field: Enter your message or initial prompt
   • Export: Save conversation and generated images
   • View HTML: Open styled conversation in browser
   • BackroomsBench (beta): Run multi-judge evaluation on conversations

3. The AIs take it from there — they can add each other, generate images, and go wherever the scenario takes them.

Configuration

Application settings in config.py:

• Runtime settings (turn delay, etc.)
• Available AI models in AI_MODELS dictionary
• Scenario prompts in SYSTEM_PROMPT_PAIRS dictionary

Developer Tools

For debugging the GUI, set DEVELOPER_TOOLS = True in config.py. This enables:

• F12: Toggle debug inspector panel
• Ctrl+Shift+C: Pick and inspect any UI element

Keep this False for normal usage.

Adding New Models

Add entries to _CURATED_MODELS in config.py using exact live OmniRoute IDs from omniroute models or GET $OMNIROUTE_BASE_URL/models. OpenRouter-style IDs (for example anthropic/claude-opus-4.6) will 404.

"Claude Opus 4.6": "anthropic/claude-opus-4-6",
"GPT 5.6 Sol (High)": "cx/gpt-5.6-sol-high",

The model dropdown also has an Endpoints group built from live OmniRoute prefixes. Choose the route, not just the model:

• agy/… Antigravity
• cx/… or codex/… Codex
• xai/… xAI Grok
• grok-cli/… Grok CLI
• anthropic/…, gemini/…, openai/…, gh/…, nvidia/…, auto/…

Creating Custom Scenarios

Add entries to SYSTEM_PROMPT_PAIRS in config.py. Each scenario needs prompts for AI-1 through AI-5. Check existing scenarios for the format — or just ask an AI to write them for you.

Sora 2 Video Generation

To enable video generation:

1. Set one AI slot to Sora 2 or Sora 2 Pro
2. Or add !video commands to your scenario prompts
3. Videos save to videos/ folder

Environment variables (optional):

SORA_SECONDS=12        # clip duration (4, 8, 10, 12)
SORA_SIZE=1280x720     # resolution

Note: Video generation is expensive. The !video command has been removed from default scenarios but is easy to add back.

Troubleshooting

1. API Issues:

   • Confirm OmniRoute is healthy: omniroute --output json health
   • Check OMNIROUTE_BASE_URL in .env (default http://127.0.0.1:20128/v1)
   • Use a live OmniRoute model ID, not an OpenRouter ID
   • Image generation still uses OPENROUTER_API_KEY when configured

2. GUI Issues:

   • Ensure PyQt6 is installed (handled by Poetry install)
   • Check Python version compatibility

3. Empty Responses:

   • Some models occasionally return empty — the app will retry once automatically
   • Check OmniRoute logs / dashboard if persistent

Contributing

1. Fork the repository
2. Clone and install dependencies:

   poetry install
   poetry run pre-commit install

3. Create a feature branch
4. Make your changes
5. Commit (pre-commit hooks will run automatically)
6. Push and create a Pull Request

Note: The pre-commit hook will block commits if DEVELOPER_TOOLS = True in config.py. Make sure to set it back to False before committing.

License

This project is licensed under the MIT License - see the LICENSE file for details.