env/reward.py

from .entities import Ticket
from .models import Action, RewardSignal


def compute_reward(
    previous: Ticket,
    current: Ticket,
    action: Action,
    distortion_penalty: float = 0.0,
    inner_product_penalty: float = 0.0,
) -> RewardSignal:
    base_value = 0.0
    reason = "neutral"

    if previous.status != "resolved" and current.status == "resolved":
        base_value = 1.0
        reason = "ticket_resolved"
    elif action.action_type == "escalate":
        base_value = 0.1
        reason = "escalated_for_support"
    elif current.attempts_used > previous.attempts_used:
        base_value = -0.05
        reason = "step_cost"

    penalty = max(distortion_penalty, 0.0) + max(inner_product_penalty, 0.0)
    if penalty > 0:
        reason = f"{reason}+quant_penalty"

    value = base_value - penalty
    value = max(-1.0, min(1.0, value))
    return RewardSignal(value=value, reason=reason)