openenv.yaml

name: email-sorting-env
version: "1.0.0"
description: "A real-world email sorting environment where an AI agent classifies emails as spam, important, or promotion. Includes easy, medium, and hard difficulty levels with meaningful reward signals."

author: "Mayur Patil"
task_type: classification

observation_space:
  type: dict
  fields:
    email:
      type: dict
      fields:
        subject:
          type: string
          description: "The subject line of the email"
        body:
          type: string
          description: "The body content of the email"
        sender:
          type: string
          description: "The sender email address"
    step:
      type: integer
      description: "Current step number in the episode"
    max_steps:
      type: integer
      description: "Maximum steps allowed per episode"
    total_reward:
      type: float
      description: "Cumulative reward so far"
    done:
      type: boolean
      description: "Whether the episode is finished"
    valid_actions:
      type: list
      description: "List of valid actions the agent can take"

action_space:
  type: string
  values:
    - spam
    - important
    - promotion
  description: "Agent must classify each email into one of these three categories"

reward:
  range: [-0.5, 1.0]
  description: "Positive reward for correct classification, negative for wrong. Higher reward for harder difficulty."
  breakdown:
    easy_correct: 0.5
    medium_correct: 0.75
    hard_correct: 1.0
    easy_wrong: -0.5
    medium_wrong: -0.3
    hard_wrong: -0.1
    invalid_action: -0.2

tasks:
  - id: easy_sorting
    name: "Easy Email Sorting"
    description: "Sort obviously spam emails like prize winning scams"
    difficulty: easy
    max_steps: 10
    grader:
      type: llm
      prompt_template: "Score the agent's email classification performance from 0.0 to 1.0 based on accuracy. The agent must correctly classify obvious spam emails (prize scams, fake giveaways) vs important emails. A score of 1.0 means perfect classification, 0.0 means all wrong."

  - id: medium_sorting
    name: "Medium Email Sorting"
    description: "Distinguish between newsletters, promotions, and important emails"
    difficulty: medium
    max_steps: 10
    grader:
      type: llm
      prompt_template: "Score the agent's email classification performance from 0.0 to 1.0. The agent must distinguish between spam, promotion, and important emails across 3 categories. A score of 1.0 means perfect classification, 0.0 means all wrong."

  - id: hard_sorting
    name: "Hard Email Sorting"
    description: "Detect phishing emails disguised as banks and trusted services"
    difficulty: hard
    max_steps: 10
    grader:
      type: llm
      prompt_template: "Score the agent's email classification performance from 0.0 to 1.0. The agent must detect subtle phishing emails disguised as legitimate bank and service communications. A score of 1.0 means perfect classification, 0.0 means all wrong."

episode:
  max_steps: 10
  reset_on_done: true

api:
  endpoints:
    - name: reset
      method: POST
      path: /reset
      description: "Reset environment and return initial state"
    - name: step
      method: POST
      path: /step
      description: "Take action and return next state, reward, done, info"
    - name: state
      method: GET
      path: /state
      description: "Return current state without taking action"