add grader endpoint, fix reward bounds and openenv.yaml

openenv.yaml

@@ -1,3 +1,4 @@
+spec_version: 1
 name: sql-correction-env
 version: "1.0.0"
 description: >
@@ -48,7 +49,7 @@ action_space:
       description: The agent's corrected SQL query
 
 reward:
-  range: [0.0, 1.0]
+  range: [0.001, 0.999]
   description: >
     1.0 = exact match, 0.7 = right tokens minor structure diff,
     0.4 = most keywords correct, 0.2 = basic structure present, 0.0 = invalid SQL.
@@ -76,5 +77,7 @@ tasks:
 endpoints:
   reset: POST /reset
   step:  POST /step
-  state: GET /state
+  state: POST /state
+  tasks: GET /tasks
+  grader: POST /grader
   health: GET /health

sql_env/grader.py

@@ -53,7 +53,7 @@ def grade(action: SQLAction, task: SQLTask) -> SQLReward:
     # ── Level 1: Exact match ─────────────────────────────────
     if agent == correct:
         return SQLReward(
-            value=_clamp(0.99),
+            value=_clamp(0.999),
             reason="Exact match — perfect correction."
         )
 
@@ -93,7 +93,7 @@ def grade(action: SQLAction, task: SQLTask) -> SQLReward:
         )
 
     # ── Level 0: No recognizable SQL ─────────────────────────
-    return SQLReward(value=_clamp(0.01), reason="Response is not valid SQL.")
+    return SQLReward(value=_clamp(0.001), reason="Response is not valid SQL.")
 
 
 def generate_feedback(action: SQLAction, task: SQLTask, reward: SQLReward) -> str: