Done status based on action in OpenAI Atari

Question

I am attempting to create a neural network to play the emulated atari game "BreakoutDeterministic". The action space for the game is [0,1,2,3].

When inputting

frame, reward, is_done, _ = env.step(env.action_space.sample())

The game will play out and and the "is_done" variable will eventually be set to True (when the game ends)

When inputting

frame, reward, is_done, _ = env.step(3)

The game will end but done will not be set to True. Instead I have to manually end the process.

Any explanation to why this happens and how I can fix it?

Full code:

import gym

env = gym.make('BreakoutDeterministic-v4')
frame = env.reset()
env.render()

is_done = False
while not is_done:
  frame, reward, is_done, _ = env.step(env.action_space.sample())
  #frame, reward, is_done, _ = env.step(3)
  env.render()