fix: typed Action model, OPENAI_API_KEY support, proper spec compliance

inference.py

@@ -28,7 +28,7 @@ from openai import OpenAI
 
 API_BASE_URL = os.getenv("API_BASE_URL", "https://integrate.api.nvidia.com/v1")
 MODEL_NAME = os.getenv("MODEL_NAME", "google/gemma-4-31b-it")
-HF_TOKEN = os.getenv("HF_TOKEN")
+HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("OPENAI_API_KEY")
 LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
 
 MAX_STEPS = 50

models.py

@@ -1,16 +1,28 @@
 """
 Data models for the EU AI Act Compliance Auditor Environment.
 
-MCP pattern: no custom Action class needed — the framework provides CallToolAction.
-We define Observation and State for typed environment responses.
+Typed Action, Observation, and State models for OpenEnv spec compliance.
 """
 
 from typing import Any, Dict, List, Optional
 
 from pydantic import BaseModel, Field
 
+try:
+    from openenv.core.env_server.types import Action, Observation
+except ImportError:
+    from pydantic import BaseModel as Action
+    from pydantic import BaseModel as Observation
 
-class ComplianceObservation(BaseModel):
+
+class ComplianceAction(Action):
+    """Action for the Compliance Auditor — an MCP tool call."""
+
+    tool_name: str = Field(default="", description="Name of the audit tool to call")
+    arguments: Dict[str, Any] = Field(default_factory=dict, description="Tool arguments as JSON")
+
+
+class ComplianceObservation(Observation):
     """Observation returned after each environment interaction."""
 
     done: bool = False

server/app.py

@@ -17,8 +17,8 @@ from typing import Any, Dict, Optional
 from fastapi import Body, HTTPException
 from pydantic import BaseModel
 from openenv.core.env_server import create_app
-from openenv.core.env_server.types import Action, Observation
 
+from models import ComplianceAction, ComplianceObservation
 from server.environment import ComplianceAuditorEnvironment, QUERY_BUDGET
 from scenarios.registry import SCENARIO_LIST, DIFFICULTY_TIERS
 
@@ -26,8 +26,8 @@ from scenarios.registry import SCENARIO_LIST, DIFFICULTY_TIERS
 
 app = create_app(
     ComplianceAuditorEnvironment,
-    Action,
-    Observation,
+    ComplianceAction,
+    ComplianceObservation,
     env_name="compliance_auditor_env",
     max_concurrent_envs=5,
 )