Fix validator discovery for Phase 2: point openenv.yaml back to tasks.X.grader:grade

openenv.yaml

@@ -20,7 +20,7 @@ tasks:
       are fake, estimate the true final price against the shopper budget,
       and decide if the shopper should proceed.
     difficulty: "easy"
-    grader: "guardian_openenv.task_graders:grade_value_hotel_budget_guard"
+    grader: "tasks.value_hotel_budget_guard.grader:grade"
 
   - id: "airline_seat_upsell_gauntlet"
     name: "Navigate an airline upsell funnel before purchase"
@@ -30,7 +30,7 @@ tasks:
       is fake, calculate the true checkout total, and recommend whether to
       continue.
     difficulty: "medium"
-    grader: "guardian_openenv.task_graders:grade_airline_seat_upsell_gauntlet"
+    grader: "tasks.airline_seat_upsell_gauntlet.grader:grade"
 
   - id: "marketplace_ghost_checkout"
     name: "Run a ghost checkout before the shopper commits"
@@ -40,4 +40,4 @@ tasks:
       verify timers, and determine the real final total before the shopper
       falls for sunk-cost pressure.
     difficulty: "hard"
-    grader: "guardian_openenv.task_graders:grade_marketplace_ghost_checkout"
+    grader: "tasks.marketplace_ghost_checkout.grader:grade"

server/app.py

@@ -81,7 +81,7 @@ def list_tasks() -> list[dict]:
             "description": task.objective,
             "difficulty": task.difficulty,
             "has_grader": True,
-            "grader": f"guardian_openenv.task_graders:grade_{task.task_id.replace('-', '_')}",
+            "grader": f"tasks.{task.task_id}.grader:grade",
         })
     return results
 

test_grader_direct.py

@@ -6,9 +6,9 @@ from guardian_openenv.task_graders import (
 )
 
 GRADERS = [
-    ("value-hotel-budget-guard", grade_value_hotel_budget_guard),
-    ("airline-seat-upsell-gauntlet", grade_airline_seat_upsell_gauntlet),
-    ("marketplace-ghost-checkout", grade_marketplace_ghost_checkout),
+    ("value_hotel_budget_guard", grade_value_hotel_budget_guard),
+    ("airline_seat_upsell_gauntlet", grade_airline_seat_upsell_gauntlet),
+    ("marketplace_ghost_checkout", grade_marketplace_ghost_checkout),
 ]
 
 print("=== Test 1: Called with no arguments ===")

test_live_space.py

@@ -0,0 +1,72 @@
+"""Test the live HF Space grader endpoint with step+grade flow."""
+import requests
+import json
+
+BASE = "https://prince9868-guardian-agent.hf.space"
+
+tasks = [
+    "value_hotel_budget_guard",
+    "airline_seat_upsell_gauntlet", 
+    "marketplace_ghost_checkout",
+]
+
+print("=== Test: reset -> submit -> grade flow ===")
+for task_id in tasks:
+    # Reset
+    r = requests.post(f"{BASE}/reset", json={"task_id": task_id}, timeout=30)
+    print(f"RESET {task_id}: {r.status_code}")
+    
+    # Immediately submit
+    r = requests.post(f"{BASE}/step", json={"action_type": "submit_decision"}, timeout=30)
+    step_data = r.json()
+    reward = step_data.get("reward", {})
+    score_val = reward.get("score")
+    value_val = reward.get("value")
+    done = step_data.get("done")
+    info = step_data.get("info", {})
+    final_score = info.get("final_score")
+    
+    print(f"  STEP: score={score_val}, value={value_val}, done={done}")
+    print(f"  INFO final_score={final_score}")
+    
+    # Check ranges 
+    if score_val is not None:
+        if score_val <= 0.0 or score_val >= 1.0:
+            print(f"  *** reward.score OUT OF RANGE: {score_val}")
+    if value_val is not None:
+        if value_val == 0.0:
+            print(f"  ** reward.value is exactly 0.0 (might flag validator)")
+    if final_score is not None:
+        if final_score <= 0.0 or final_score >= 1.0:
+            print(f"  *** final_score OUT OF RANGE: {final_score}")
+    
+    # Grade
+    r = requests.post(f"{BASE}/grader", json={"task_id": task_id}, timeout=30)
+    grade_data = r.json()
+    gscore = grade_data.get("score")
+    print(f"  GRADER: score={gscore}")
+    if gscore is not None and (gscore <= 0.0 or gscore >= 1.0):
+        print(f"  *** GRADER score OUT OF RANGE: {gscore}")
+    
+    breakdown = grade_data.get("grader_breakdown", {})
+    for k, v in breakdown.items():
+        if isinstance(v, (int, float)) and (v <= 0.0 or v >= 1.0):
+            print(f"  *** breakdown.{k}={v} OUT OF RANGE")
+    print()
+
+print("=== Test: empty POST to /reset then /grader ===")
+r = requests.post(f"{BASE}/reset", timeout=30)
+print(f"RESET (empty): {r.status_code}")
+r = requests.post(f"{BASE}/grader", timeout=30)
+grade_data = r.json()
+print(f"GRADER (empty): score={grade_data.get('score')}")
+
+print()
+print("=== Test: POST /grade (alias) ===")
+for task_id in tasks:
+    requests.post(f"{BASE}/reset", json={"task_id": task_id}, timeout=30)
+    r = requests.post(f"{BASE}/grade", json={"task_id": task_id}, timeout=30)
+    print(f"  /grade {task_id}: {r.status_code} score={r.json().get('score')}")
+
+print()
+print("ALL LIVE TESTS COMPLETE")

tests/test_environment.py

@@ -12,10 +12,10 @@ from guardian_openenv.tasks import TASKS, TASKS_BY_ID
 
 def test_reset_returns_expected_task():
     env = GuardianReviewEnvironment()
-    observation = env.reset("value-hotel-budget-guard")
+    observation = env.reset("value_hotel_budget_guard")
 
-    task = TASKS_BY_ID["value-hotel-budget-guard"]
-    assert observation.task_id == "value-hotel-budget-guard"
+    task = TASKS_BY_ID["value_hotel_budget_guard"]
+    assert observation.task_id == "value_hotel_budget_guard"
     # remaining_steps equals the task's max_steps right after reset
     assert observation.remaining_steps == task.max_steps
     assert observation.current_decision.flagged_patterns == []
@@ -24,9 +24,9 @@ def test_reset_returns_expected_task():
 
 def test_state_reflects_task_after_reset():
     env = GuardianReviewEnvironment()
-    env.reset("airline-seat-upsell-gauntlet")
+    env.reset("airline_seat_upsell_gauntlet")
     state = env.state()
-    assert state.task_id == "airline-seat-upsell-gauntlet"
+    assert state.task_id == "airline_seat_upsell_gauntlet"
     assert state.difficulty == "medium"
     assert state.step_count == 0
     assert state.done is False
@@ -34,8 +34,8 @@ def test_state_reflects_task_after_reset():
 
 def test_step_increases_step_count():
     env = GuardianReviewEnvironment()
-    obs = env.reset("value-hotel-budget-guard")
-    task = TASKS_BY_ID["value-hotel-budget-guard"]
+    obs = env.reset("value_hotel_budget_guard")
+    task = TASKS_BY_ID["value_hotel_budget_guard"]
 
     result = env.step(GuardianAction(action_type=ActionType.INSPECT_SECTION, section_id=task.sections[0].section_id))
     assert env.state().step_count == 1
@@ -59,7 +59,7 @@ def test_reward_score_always_in_open_range():
 
 def test_invalid_section_penalised():
     env = GuardianReviewEnvironment()
-    env.reset("value-hotel-budget-guard")
+    env.reset("value_hotel_budget_guard")
     result = env.step(GuardianAction(action_type=ActionType.INSPECT_SECTION, section_id="NONEXISTENT"))
     assert result.reward.value < 0.0
 
@@ -68,7 +68,7 @@ def test_invalid_section_penalised():
 
 def test_dense_progress_and_final_grade():
     env = GuardianReviewEnvironment()
-    env.reset("value-hotel-budget-guard")
+    env.reset("value_hotel_budget_guard")
 
     env.step(GuardianAction(action_type=ActionType.INSPECT_SECTION, section_id="listing-banner"))
     env.step(GuardianAction(action_type=ActionType.INSPECT_SECTION, section_id="price-summary"))
@@ -99,7 +99,7 @@ def test_dense_progress_and_final_grade():
 
 def test_medium_task_completes():
     env = GuardianReviewEnvironment()
-    env.reset("airline-seat-upsell-gauntlet")
+    env.reset("airline_seat_upsell_gauntlet")
     result = env.step(GuardianAction(action_type=ActionType.SUBMIT_DECISION))
     assert result.done is True
     assert 0.0 < result.reward.score < 1.0
@@ -107,7 +107,7 @@ def test_medium_task_completes():
 
 def test_hard_task_completes():
     env = GuardianReviewEnvironment()
-    env.reset("marketplace-ghost-checkout")
+    env.reset("marketplace_ghost_checkout")
     result = env.step(GuardianAction(action_type=ActionType.SUBMIT_DECISION))
     assert result.done is True
     assert 0.0 < result.reward.score < 1.0
@@ -117,7 +117,7 @@ def test_hard_task_completes():
 
 def test_actions_after_done_return_reward_zero():
     env = GuardianReviewEnvironment()
-    env.reset("value-hotel-budget-guard")
+    env.reset("value_hotel_budget_guard")
     env.step(GuardianAction(action_type=ActionType.SUBMIT_DECISION))
     # Further steps should be no-ops with value=0
     noop = env.step(GuardianAction(action_type=ActionType.SUBMIT_DECISION))