TurnOrchestrator: Coordinates multi-agent turn-based simulation
- Perspective switching with FOV layer updates
- Screenshot capture per agent per turn
- Pluggable LLM query callback
- SimulationStep/SimulationLog for full context capture
- JSON save/load with replay support
New demos:
- 2_integrated_demo.py: WorldGraph + action execution integration
- 3_multi_turn_demo.py: Complete multi-turn simulation with logging
Updated 1_multi_agent_demo.py with action parser/executor integration.
Tested with Qwen2.5-VL-32B: agents successfully navigate based on
WorldGraph descriptions and VLM visual input.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
ActionParser: Extracts structured actions from LLM text responses
- Regex patterns for GO, WAIT, LOOK, TAKE, DROP, PUSH, USE, etc.
- Direction normalization (N→NORTH, UP→NORTH)
- Handles "Action: GO EAST" and fallback patterns
- 12 unit tests covering edge cases
ActionExecutor: Executes parsed actions in the game world
- Movement with collision detection (walls, entities)
- Boundary checking
- ActionResult with path data for animation replay
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Implements Python-side room graph data structures for LLM agent environments:
- Room, Door, WorldObject dataclasses with full metadata
- WorldGraph class with spatial queries (room_at, get_exits)
- Deterministic text generation (describe_room, describe_exits)
- Available action enumeration based on room state
- Factory functions for test scenarios (two_room, button_door)
Example output:
"You are in the guard room. The air is musty. On the ground you see
a brass key. Exits: east (the armory)."
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- 0_basic_vllm_demo.py: Single agent with FOV, grounded text, VLLM query
- 1_multi_agent_demo.py: Three agents with perspective cycling
Features demonstrated:
- Headless step() + screenshot() for AI-driven gameplay
- ColorLayer.apply_perspective() for per-agent fog of war
- Grounded text generation based on entity visibility
- Sequential VLLM queries with vision model support
- Proper FOV reset between perspective switches
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
John McCardle