A stateful UI to generate Python 3D Raycasting payloads for LLM Code Interpreters. It operates as a Read-Eval-Print Loop (REPL) to test Large Language Models (LLMs) on spatial reasoning, deterministic physics, and long-horizon pathfinding.
Welcome to the GROND-REPL Controller
This tool allows you to explore a 3D environment rendered entirely in ASCII by tricking an AI's Python Sandbox into running a retro raycasting engine. Since the AI doesn't have an interactive screen, you play by "correspondence" (sending it a list of moves, and it sending you back a snapshot).
How to Play:
- Design (Optional): Use the MATRIX ARCHITECT to paint a custom map topology.
- Queue Your Moves: Use the TACTICAL INPUT buttons or MACRO AUTO-GEN to plan your path.
- Generate: Click > EXECUTE FORGE in the Payload Forge panel.
- Execute: Copy the generated Markdown payload and paste it directly into an LLM with a Code Interpreter.
- Behold: The LLM will execute the code and reply with your rendered ASCII frame and Telemetry HUD.
⚠️ THE GOLDEN RULE: Do NOT Clear the Queue!
LLM code environments frequently spin up and shut down, meaning they forget where you were standing between prompts. To fix this, our payload speed-runs your entire historical run from the spawn point every single time. ONLY click "✕ RESPAWN" if you want to wipe the queue and start back at the very beginning of the level!
Reading the HUD
The status bar and the PAYLOAD LOAD meter estimate how big your payload is. SANITY is the share of your chosen context budget still free after the estimated working set (about 2.5x the payload). As it drops, the pixel face gets more battered. All numbers are estimates, not exact token counts.