| """ |
| models.py - Pydantic data models for the OpenEnv Customer Support environment. |
| |
| Defines the action, observation, and state schemas used by the environment server, |
| client, and baseline agent. Fully compliant with the OpenEnv spec. |
| """ |
|
|
| from typing import List, Optional |
|
|
| from pydantic import BaseModel |
|
|
| try: |
| from openenv.core.env_server import Action, Observation, State |
| except ImportError: |
| |
| Action = BaseModel |
| Observation = BaseModel |
| State = BaseModel |
|
|
|
|
| class SupportAction(Action): |
| """An action taken by the support agent in response to a customer query. |
| |
| The agent sends a text message and optionally declares its intent |
| (e.g., classify the ticket, respond with a solution, ask for clarification, |
| escalate to a human, or close the ticket). |
| """ |
|
|
| message: str |
| """The agent's text reply to the customer.""" |
|
|
| intent: Optional[str] = None |
| """The agent's declared intent for this action. |
| |
| Must be one of: "classify", "respond", "clarify", "escalate", "close". |
| If None, the environment will infer intent from the message content. |
| """ |
|
|
|
|
| class SupportObservation(Observation): |
| """An observation returned by the environment after each step. |
| |
| Contains the full conversation history, the latest customer message |
| the agent must address, the current task difficulty, and scoring info. |
| """ |
|
|
| conversation: List[str] |
| """Full list of all messages exchanged so far (alternating customer/agent).""" |
|
|
| customer_query: str |
| """The latest customer message the agent must respond to.""" |
|
|
| task_name: str |
| """Difficulty tier for the current episode: "easy", "medium", or "hard".""" |
|
|
| info: Optional[str] = None |
| """Extra hints or context provided to the agent (e.g., knowledge-base snippets).""" |
|
|
| done: bool = False |
| """Whether the episode has ended (resolved, escalated, or max steps reached).""" |
|
|
| reward: Optional[float] = None |
| """Grader-assigned reward for the most recent step (0.0–1.0). None on the first obs.""" |
|
|
| cumulative_reward: float = 0.0 |
| """Running total reward accumulated across all turns in this episode.""" |
|
|
| turn_scores: List[float] = [] |
| """Per-turn reward breakdown (useful for hard multi-turn task analysis).""" |
|
|
|
|
| class SupportState(State): |
| """Internal state of the customer support environment for a single episode. |
| |
| Tracks the issue metadata, conversation progress, and grading context. |
| The environment uses this to drive step logic and evaluate the agent. |
| """ |
|
|
| issue_type: str = "" |
| """Category of the customer issue: "refund", "technical", "shipping", "billing", "account".""" |
|
|
| step_count: int = 0 |
| """Number of agent actions taken so far in this episode.""" |
|
|
| resolved: bool = False |
| """Whether the customer's issue has been successfully resolved.""" |
|
|
| episode_id: str = "" |
| """Unique identifier for the current episode.""" |
|
|
| task_name: str = "easy" |
| """Difficulty tier for the current episode: "easy", "medium", or "hard".""" |
|
|
| conversation_history: List[str] = [] |
| """Running log of all messages exchanged between the customer and the agent.""" |
|
|
| correct_answer: str = "" |
| """Ground-truth answer used by graders to evaluate the agent's reply.""" |
|
|
| max_steps: int = 10 |
| """Maximum number of agent steps allowed before the episode is forcibly ended.""" |
|
|
| cumulative_reward: float = 0.0 |
| """Running total of rewards accumulated across all turns.""" |
|
|
| turn_scores: List[float] = [] |
| """Per-turn reward values for detailed grader breakdown.""" |
|
|
| scenario_index: int = 0 |
| """Index of the chosen scenario within its category (for reproducibility).""" |
|
|
