Default provider to zerogpu_transformers on HF Spaces; drop bogus README env block

README.md

@@ -16,12 +16,6 @@ tags:
   - game
   - agent
   - minicpm
-env:
-  - PHANTOM_GRID_LLM_PROVIDER=zerogpu_transformers
-  - PHANTOM_GRID_ZEROGPU_MODEL_ID=openbmb/MiniCPM4.1-8B
-  - PHANTOM_GRID_ZEROGPU_DURATION=90
-  - PHANTOM_GRID_HOST=0.0.0.0
-  - PHANTOM_GRID_WITNESS_CHAT_TTS=0
 ---
 
 # 🕵️ Phantom Grid

config/settings.py

@@ -92,9 +92,14 @@ def load_settings() -> Settings:
         detected = shutil.which("llama-server") or shutil.which("llama-server.exe")
         server_bin = Path(detected) if detected else None
 
+    # HF Spaces auto-injects SPACE_ID. When present, default to the in-process
+    # ZeroGPU transformers backend so the Space never tries to invoke cmake or
+    # spawn llama-server. Explicit PHANTOM_GRID_LLM_PROVIDER still wins.
+    default_provider = "zerogpu_transformers" if os.getenv("SPACE_ID") else "llama_cpp_server"
+
     return Settings(
         games_dir=_env_path("PHANTOM_GRID_GAMES_DIR", PROJECT_ROOT / "data" / "games") or PROJECT_ROOT / "data" / "games",
-        llm_provider=os.getenv("PHANTOM_GRID_LLM_PROVIDER", "llama_cpp_server"),
+        llm_provider=os.getenv("PHANTOM_GRID_LLM_PROVIDER", default_provider),
         llamacpp_server_bin=server_bin,
         llamacpp_model_path=_env_path("PHANTOM_GRID_LLAMACPP_MODEL_PATH"),
         llamacpp_base_url=os.getenv("PHANTOM_GRID_LLAMACPP_BASE_URL", "http://127.0.0.1:8080/v1"),