Expose 3 graded tasks and strict in-range task scores for phase 2

AgentBox/src/env.py

@@ -24,11 +24,20 @@ class CodeGuardEnv:
         # Strict score interval for validator compatibility.
         return max(0.01, min(0.99, base_score))
 
+    def _get_all_task_scores(self, action: str) -> Dict[str, float]:
+        scores: Dict[str, float] = {}
+        for key, grader in GRADERS.items():
+            score = float(grader(action))
+            scores[key] = max(0.01, min(0.99, score))
+        return scores
+
     def reset(self) -> Dict[str, Any]:
         self.state = {
-            "score": 0.0,
+            "score": 0.01,
             "history": [],
             "task": TASKS[self.task_key],
+            "tasks": list(TASKS.values()),
+            "task_scores": {k: 0.01 for k in GRADERS.keys()},
         }
         self.current_step = 0
         self.done = False
@@ -59,7 +68,8 @@ class CodeGuardEnv:
 
         reward: float = compute_reward(self.state, action, base_score)
 
-        self.state["score"] = reward
+        self.state["score"] = max(0.01, min(0.99, base_score))
+        self.state["task_scores"] = self._get_all_task_scores(action)
         self.state["history"].append(
             {
                 "step": self.current_step,

AgentBox/src/tasks/easy.py

@@ -10,6 +10,8 @@ def get_task() -> Dict:
     return {
         "name": "lint_fix",
         "description": "Fix syntax errors in the given Python code.",
+        "grader": "grade",
+        "score_range": [MIN_TASK_SCORE, MAX_TASK_SCORE],
     }
 
 

AgentBox/src/tasks/hard.py

@@ -10,6 +10,8 @@ def get_task() -> Dict:
     return {
         "name": "refactor_types",
         "description": "Refactor code and add type hints.",
+        "grader": "grade",
+        "score_range": [MIN_TASK_SCORE, MAX_TASK_SCORE],
     }
 
 

AgentBox/src/tasks/medium.py

@@ -11,6 +11,8 @@ def get_task() -> Dict:
     return {
         "name": "vuln_patch",
         "description": "Remove unsafe function usage while preserving functionality.",
+        "grader": "grade",
+        "score_range": [MIN_TASK_SCORE, MAX_TASK_SCORE],
     }
 
 

src/env.py

@@ -0,0 +1 @@
+from AgentBox.src.env import *  # noqa: F401,F403

src/reward.py

@@ -0,0 +1 @@
+from AgentBox.src.reward import *  # noqa: F401,F403

src/tasks/__init__.py

@@ -0,0 +1 @@
+from AgentBox.src.tasks import *  # noqa: F401,F403

src/tasks/easy.py

@@ -0,0 +1 @@
+from AgentBox.src.tasks.easy import *  # noqa: F401,F403

src/tasks/hard.py

@@ -0,0 +1 @@
+from AgentBox.src.tasks.hard import *  # noqa: F401,F403

src/tasks/medium.py

@@ -0,0 +1 @@
+from AgentBox.src.tasks.medium import *  # noqa: F401,F403