Function _run_simplerenv_eval

tests/examples/test_simplerenv.py:51–137

Shared helper: setup sim, start server, run rollout, assert results.

(
    env: dict,
    blocks: list,
    server_model_key: str,
    client_env_name_old: str,
    client_env_name_new: str,
    server_startup_env_var: str,
)

def _run_simplerenv_eval(
    env: dict,
    blocks: list,
    server_model_key: str,
    client_env_name_old: str,
    client_env_name_new: str,
    server_startup_env_var: str,
) -> None:
    """Shared helper: setup sim, start server, run rollout, assert results."""
    # Step 1: Setup sim (shared across both benchmarks)
    with timed("step 1: sim venv setup (setup_SimplerEnv.sh)"):
        run_bash_blocks(
            [find_block(blocks, "setup_SimplerEnv.sh", language="bash")],
            cwd=REPO_ROOT,
            env=env,
        )

    model_server_host = "127.0.0.1"
    model_server_port = 5559

    # Step 2: Server — inject test-specific flags
    server_code = find_block(blocks, server_model_key, language="bash").code
    server_code += f" --device cuda:0 --host {model_server_host} --port {model_server_port}"

    # Step 3: Rollout — substitute test-safe values
    rollout_code = replace_once(
        replace_once(
            replace_once(
                replace_once(
                    replace_once(
                        find_block(blocks, client_env_name_old, language="bash").code,
                        "--n-episodes 10",
                        "--n-episodes 1",
                    ),
                    "--policy-client-port 5555",
                    f"--policy-client-port {model_server_port}",
                ),
                "--max-episode-steps 300",
                "--max-episode-steps 2",
            ),
            "--n-envs 5",
            "--n-envs 1",
        ),
        client_env_name_old,
        client_env_name_new,
    )

    assert_port_available(model_server_host, model_server_port)
    model_server_proc, server_log = start_server_process(server_code, cwd=REPO_ROOT, env=env)
    with timed("step 2: server startup"):
        wait_for_server_ready(
            proc=model_server_proc,
            host=model_server_host,
            port=model_server_port,
            timeout_s=float(os.getenv(server_startup_env_var, str(DEFAULT_SERVER_STARTUP_SECONDS))),
            server_log=server_log,
        )