02
Environment type & API
LLM policy
tool call | final answer
messages + tool result
SQLite sandbox
Mental model
An OpenEnv environment served over HTTP from a Docker container. The client exposes the familiar reset(), step(action) and state() calls. One step is one assistant turn: a single tool call or the final answer.
from reward_atlas_sql_detective import SQLDetectiveClient, SQLAction
with SQLDetectiveClient.from_docker_image("rewardatlas/sql-detective:1.0") as env:
obs = env.reset()
obs = env.step(SQLAction(tool="list_tables", args={}))
print(obs.observation.tool_result)