Upload folder using huggingface_hub

client.py

@@ -51,15 +51,12 @@ class NL2SQLEnv:
         return self._parse_result(resp.json())
 
     def _parse_result(self, payload: Dict[str, Any]) -> StepResult:
-        obs_data = payload.get("observation", payload)
+        obs_data = payload.get("observation", {})
         
-        # SAFETY CHECK: If reward or score is None/null, default to 0.0
-        raw_reward = obs_data.get("reward")
-        safe_reward = float(raw_reward) if raw_reward is not None else 0.0
+        # Read standard OpenEnv top-level keys safely
+        safe_reward = float(payload.get("reward", 0.0))
+        safe_done = bool(payload.get("done", False))
         
-        raw_score = obs_data.get("score")
-        safe_score = float(raw_score) if raw_score is not None else 0.0
-
         obs = NL2SQLObservation(
             question=obs_data.get("question", ""),
             schema_context=obs_data.get("schema_context", ""),
@@ -70,12 +67,12 @@ class NL2SQLEnv:
             result_columns=obs_data.get("result_columns", []),
             step=obs_data.get("step", 0),
             max_steps=obs_data.get("max_steps", 5),
-            done=obs_data.get("done", False),
+            done=safe_done,
             reward=safe_reward,
-            score=safe_score,
+            score=float(obs_data.get("score", 0.0)),
         )
         return StepResult(
             observation=obs,
             reward=safe_reward,
-            done=obs.done,
+            done=safe_done,
         )

server/environment.py

@@ -152,7 +152,7 @@ class NL2SQLEnvironment(Environment):
         self._last_obs = obs
         return obs
 
-    def step(self, action: NL2SQLAction) -> NL2SQLObservation:
+    def step(self, action: NL2SQLAction) -> tuple[NL2SQLObservation, float, bool, dict]:
         """Execute the agent's SQL and return graded observation."""
         if self._task is None or self._example is None:
             # Called before reset — auto-reset
@@ -210,8 +210,11 @@ class NL2SQLEnvironment(Environment):
             reward=reward,
             score=score,
         )
+
         self._last_obs = obs
-        return obs
+        # CRITICAL FIX: Return the standard OpenEnv 4-part tuple!
+        info = {"score": score, "error": error}
+        return obs, reward, done, info
 
     @property
     def state(self) -> NL2SQLState: