v5: navigation graph, DFS backtracking, updated README

v4: extract valid actions free from play_action response, timeout=120

v3: exploration bias, untried directions, use_parallel=False, README report

v2: dead-end memory, valid-actions cache, score stagnation detection

Implement ReAct agent with 5 MCP tools

Updated agent with local option

simplified app

Update gradio sdk_version to 5.12.0

Remove gradio from requirements - already installed on HF Spaces

Initial template