Spaces:
Sleeping
Sleeping
Commit ·
b1b08c5
1
Parent(s): f18b8fb
Stabilize task grader discovery paths for Phase 2 validation
Browse files- openenv.yaml +3 -3
- server/app.py +6 -1
openenv.yaml
CHANGED
|
@@ -20,7 +20,7 @@ tasks:
|
|
| 20 |
are fake, estimate the true final price against the shopper budget,
|
| 21 |
and decide if the shopper should proceed.
|
| 22 |
difficulty: "easy"
|
| 23 |
-
grader: "
|
| 24 |
|
| 25 |
- id: "airline_seat_upsell_gauntlet"
|
| 26 |
name: "Navigate an airline upsell funnel before purchase"
|
|
@@ -30,7 +30,7 @@ tasks:
|
|
| 30 |
is fake, calculate the true checkout total, and recommend whether to
|
| 31 |
continue.
|
| 32 |
difficulty: "medium"
|
| 33 |
-
grader: "
|
| 34 |
|
| 35 |
- id: "marketplace_ghost_checkout"
|
| 36 |
name: "Run a ghost checkout before the shopper commits"
|
|
@@ -40,4 +40,4 @@ tasks:
|
|
| 40 |
verify timers, and determine the real final total before the shopper
|
| 41 |
falls for sunk-cost pressure.
|
| 42 |
difficulty: "hard"
|
| 43 |
-
grader: "
|
|
|
|
| 20 |
are fake, estimate the true final price against the shopper budget,
|
| 21 |
and decide if the shopper should proceed.
|
| 22 |
difficulty: "easy"
|
| 23 |
+
grader: "guardian_openenv.task_graders:grade_value_hotel_budget_guard"
|
| 24 |
|
| 25 |
- id: "airline_seat_upsell_gauntlet"
|
| 26 |
name: "Navigate an airline upsell funnel before purchase"
|
|
|
|
| 30 |
is fake, calculate the true checkout total, and recommend whether to
|
| 31 |
continue.
|
| 32 |
difficulty: "medium"
|
| 33 |
+
grader: "guardian_openenv.task_graders:grade_airline_seat_upsell_gauntlet"
|
| 34 |
|
| 35 |
- id: "marketplace_ghost_checkout"
|
| 36 |
name: "Run a ghost checkout before the shopper commits"
|
|
|
|
| 40 |
verify timers, and determine the real final total before the shopper
|
| 41 |
falls for sunk-cost pressure.
|
| 42 |
difficulty: "hard"
|
| 43 |
+
grader: "guardian_openenv.task_graders:grade_marketplace_ghost_checkout"
|
server/app.py
CHANGED
|
@@ -72,6 +72,11 @@ def metadata() -> dict:
|
|
| 72 |
@app.get("/tasks")
|
| 73 |
def list_tasks() -> list[dict]:
|
| 74 |
"""List all tasks with metadata — the validator discovers graders here."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
results = []
|
| 76 |
for task in TASKS:
|
| 77 |
results.append({
|
|
@@ -81,7 +86,7 @@ def list_tasks() -> list[dict]:
|
|
| 81 |
"description": task.objective,
|
| 82 |
"difficulty": task.difficulty,
|
| 83 |
"has_grader": True,
|
| 84 |
-
"grader":
|
| 85 |
})
|
| 86 |
return results
|
| 87 |
|
|
|
|
| 72 |
@app.get("/tasks")
|
| 73 |
def list_tasks() -> list[dict]:
|
| 74 |
"""List all tasks with metadata — the validator discovers graders here."""
|
| 75 |
+
grader_by_task_id = {
|
| 76 |
+
"value_hotel_budget_guard": "guardian_openenv.task_graders:grade_value_hotel_budget_guard",
|
| 77 |
+
"airline_seat_upsell_gauntlet": "guardian_openenv.task_graders:grade_airline_seat_upsell_gauntlet",
|
| 78 |
+
"marketplace_ghost_checkout": "guardian_openenv.task_graders:grade_marketplace_ghost_checkout",
|
| 79 |
+
}
|
| 80 |
results = []
|
| 81 |
for task in TASKS:
|
| 82 |
results.append({
|
|
|
|
| 86 |
"description": task.objective,
|
| 87 |
"difficulty": task.difficulty,
|
| 88 |
"has_grader": True,
|
| 89 |
+
"grader": grader_by_task_id.get(task.task_id, "guardian_openenv.task_graders:grade"),
|
| 90 |
})
|
| 91 |
return results
|
| 92 |
|