Custom Environments - ICRL

Python Interface

class MyEnv:
    def reset(self, goal: str) -> str:
        ...

    def step(self, action: str) -> tuple[str, bool, bool] | Awaitable[tuple[str, bool, bool]]:
        ...

Return value is (observation, done, success).

TypeScript Interface

interface Environment {
  reset(goal: string): string | Promise<string>;
  step(action: string): StepResult | Promise<StepResult>;
}

Where:

interface StepResult {
  observation: string;
  done: boolean;
  success: boolean;
}

Practical Rules

Store the goal in reset.
Keep observation text actionable.
Terminate episodes via done.
Set success only on true goal completion.

Reference Implementations

Python file environment: examples/file_api_env.py
Python coding environment: examples/harbor_coding_agent.py
TypeScript demos: icrl-ts/examples/*.ts