openenv.yaml

name: sieve
version: "1.0.0"
description: >
  A customer support reinforcement learning environment where an agent triages,
  classifies, drafts responses to, and prioritizes real-world support emails.
tags:
  - openenv
  - customer-support
  - email-triage
  - nlp

api:
  base_url: http://localhost:7860
  reset: POST /reset
  step: POST /step
  state: GET /state
  tasks: GET /tasks

tasks:
  - id: email_classification
    name: Email Classification
    difficulty: easy
    description: >
      Classify each incoming customer support email by category
      (billing/technical/general/spam/account/feature_request)
      and urgency (high/medium/low) using action_type='classify'.
    max_steps: 15
    score_range: [0.0, 1.0]

  - id: response_drafting
    name: Response Drafting
    difficulty: medium
    description: >
      Draft professional, empathetic responses to customer support emails.
      Each response must cover required keywords, exceed 50 characters,
      and maintain a professional tone. Use action_type='respond'.
    max_steps: 10
    score_range: [0.0, 1.0]

  - id: support_session
    name: Full Support Session
    difficulty: hard
    description: >
      Manage a queue of 15 mixed emails. Prioritize VIP customers first,
      then high-urgency emails. Choose the correct action (respond/escalate/archive)
      per email, provide category and urgency, and use email_id to select
      which email to process each step.
    max_steps: 40
    score_range: [0.0, 1.0]

observation_space:
  current_email:
    id: string
    subject: string
    body: string
    sender: string
    sender_tier: "standard | vip"
    received_minutes_ago: integer
  email_queue: "array of Email objects (populated in support_session only)"
  processed_count: integer
  step_count: integer
  task_id: string
  task_description: string
  available_actions: "array of strings"
  context:
    max_steps: integer
    remaining_steps: integer
    queue_size: integer

action_space:
  action_type: "classify | respond | escalate | archive | skip"
  category: "billing | technical | general | spam | account | feature_request"
  urgency: "high | medium | low"
  response_text: "string — required for respond"
  escalation_reason: "string — required for escalate"
  email_id: "string — used in support_session to select target email"

baseline:
  agent: gpt-4o-mini
  scores:
    email_classification: 0.930
    response_drafting: 0.920
    support_session: 0.882
    average: 0.911