akhilsu's picture
Upload 3 files
6d51f50 verified
Raw
History Blame Contribute Delete
3.11 kB
from models import ActionType, B2BSupportPayload, B2BSupportTriageAction
from server.support_triage_environment import B2BSupportTriageEnvironment
def test_reset_initializes_clean_state() -> None:
env = B2BSupportTriageEnvironment()
obs = env.reset(task_id="easy", seed=11)
assert obs.task_id == "easy"
assert obs.step_index == 0
assert 0.0 < obs.progress_score < 1.0
assert env.state.step_count == 0
assert env.state.applied_decisions == {}
def test_invalid_ticket_penalty() -> None:
env = B2BSupportTriageEnvironment()
env.reset(task_id="easy")
obs = env.step(
B2BSupportTriageAction(
action_type=ActionType.CLASSIFY,
ticket_id="WRONG",
payload=B2BSupportPayload(category="billing"),
)
)
assert obs.reward < 0
assert obs.last_action_error is not None
assert "ticket_id_mismatch" in obs.last_action_error
def test_hard_task_can_reach_full_score() -> None:
env = B2BSupportTriageEnvironment()
reset_obs = env.reset(task_id="hard")
ticket_id = reset_obs.visible_ticket.ticket_id
env.step(
B2BSupportTriageAction(
action_type=ActionType.CLASSIFY,
ticket_id=ticket_id,
payload=B2BSupportPayload(category="security"),
)
)
env.step(
B2BSupportTriageAction(
action_type=ActionType.SET_PRIORITY,
ticket_id=ticket_id,
payload=B2BSupportPayload(priority="urgent"),
)
)
env.step(
B2BSupportTriageAction(
action_type=ActionType.ROUTE,
ticket_id=ticket_id,
payload=B2BSupportPayload(
route_queue="security-incident-response",
sla_minutes=120,
escalate=True,
),
)
)
env.step(
B2BSupportTriageAction(
action_type=ActionType.DRAFT_REPLY,
ticket_id=ticket_id,
payload=B2BSupportPayload(
reply_text=(
"We have escalated this to our security team. "
"This issue is escalated. Please reset your API key now. "
"We will provide an update in 2 hours."
)
),
)
)
final_obs = env.step(
B2BSupportTriageAction(
action_type=ActionType.SUBMIT,
ticket_id=None,
payload=B2BSupportPayload(),
)
)
assert final_obs.done is True
assert 0.99 <= final_obs.progress_score < 1.0
def test_episode_ends_at_max_steps() -> None:
env = B2BSupportTriageEnvironment()
obs = env.reset(task_id="easy")
ticket_id = obs.visible_ticket.ticket_id
final_obs = obs
for _ in range(obs.max_steps):
final_obs = env.step(
B2BSupportTriageAction(
action_type=ActionType.CLASSIFY,
ticket_id=ticket_id,
payload=B2BSupportPayload(category="billing"),
)
)
assert final_obs.done is True
assert final_obs.last_action_error == "max_steps_reached"