Schema Surgeon: An RL Environment for Database Schema Migration

Context: OpenEnv Hackathon, 2026 (team project). Code: github.com/MAUK9086/Schema_Surgeon_OpenEnv_Meta

Motivation

Real-world data cleaning involves long sequences of small, irreversible decisions. Benchmarks for language-model agents rarely test this. Schema Surgeon frames schema migration as a sequential decision problem in which the agent acts as a database reliability engineer.

Environment

• State: a collection of 50 documents with inconsistent schemas. The agent observes a sample of 10 documents plus global key-frequency statistics.
• Actions: rename_and_merge, cast_type, flatten_field, delete_key and terminate.
• Tasks (30 steps each, increasing difficulty):
  1. Key drift: merge uid / u_id into user_id.
  2. Type drift: e.g., cast string-typed ages to integers and prices to floats.
  3. Nested and ambiguous types: values such as "25", true or {"val": 25}, nested fields to flatten, and noise fields to remove.
• Reward: dense, equal to the change in the fraction of documents that validate against the target schema. Penalties apply to no-op actions (−0.1) and destructive actions (−0.5).
• Reproducibility: fixed datasets per task. The environment is served over a FastAPI WebSocket following the OpenEnv specification.
• Baseline: a gpt-4o-mini agent starts Task 1 at a score of 0.380 (19 of 50 documents already valid).