feat: Gymnasium style base environment

[cmunley1]

#1030

Gymnasium-style step()/reset() API implemented as a base resources server, with a simple compatible agent to drive the agent-environment interaction in this context, and an example blackjack resources server extending the base class.

resources_servers/base_gymnasium/ has GymnasiumServer base class. Meant to be subclassed, implement step() and optionally reset(), use with gymnasium_agent. Reset is used to init state or provide an additional initial message. Step handles verification, multistep logic, and stopping logic.

The base class handles /reset and /step endpoints, session IDs, and the HTTP layer.

class MyEnv(GymnasiumServer):
    async def step(self, action, metadata, session_id=None):
        return observation, reward, terminated, truncated, info

responses_api_agents/gymnasium_agent/ drives the loop: calls /reset, then alternates between model calls and /step until terminated or truncated. Handles token ids, usage, cookies

Examples

• blackjack_env - multi-step game, <action> tags, stateful

Removed multi-turn examples for follow up PR.

Docs

docs/resources-server/gymnasium-api.md covers usage and examples.

Impact on existing environments

Nothing existing changed. example_gymnasium + gymnasium_agent is a simpler way to write environments. It hides the HTTP endpoints, session management, and request/response types. Users just implement step(). Existing resources servers + agents path still works as before.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.