name: dataclean-env
version: 2.0.0
description: >
  Real-world data cleaning and ETL environment for AI agent RL training.
  4 tasks including the novel Data Drift task — live streaming rows injected
  mid-episode every 5 steps, simulating a Kafka pipeline under continuous drift.
author: Team DataClean
tags:
  - openenv
  - data-cleaning
  - etl
  - real-world
  - tabular
  - data-drift
  - streaming

tasks:
  - id: task1
    name: Null Fixer
    difficulty: easy
    max_steps: 10
    description: Fix null values and wrong dtypes in a 50-row customer dataset.

  - id: task2
    name: Schema Normalizer
    difficulty: medium
    max_steps: 20
    description: Deduplicate, normalize country names, fix date format, fill nulls.

  - id: task3
    name: ETL Pipeline
    difficulty: hard
    max_steps: 30
    description: Merge two tables, remove outliers, add derived column.

  - id: task4_data_drift
    name: Data Drift (Streaming)
    difficulty: expert
    max_steps: 40
    description: >
      NOVEL. Live streaming transactions — fresh dirty rows injected every 5 steps
      mid-episode. Agent must clean continuously under real-time data drift.

environment:
  type: http
  port: 7860
  endpoints:
    reset:    POST /reset
    step:     POST /step
    state:    GET  /state
    tasks:    GET  /tasks
    grader:   GET  /grader
    baseline: GET  /baseline
    health:   GET  /health