Create a new environment

An environment is the task/world only — it owns the hidden target, the per-actor measurement, and the team reward. Communication (delay, partial observability) is handled by network.py, so you do not re-implement it.

The contract

from coadapt.env import Env, EnvSpec
from coadapt.spaces import Discrete   # or MultiDiscrete / Box
import chex, jax, jax.numpy as jnp

@chex.dataclass
class MyWorld:            # the hidden task state (a JAX pytree)
    target: jnp.ndarray
    t: jnp.ndarray

class MyEnv(Env):
    def __init__(self, N=8, horizon=60, nu=0.0):
        self.spec = EnvSpec(num_agents=N, meas_dim=3,
                            action_space=Discrete(5), horizon=horizon)
    def reset(self, key):
        # -> (measurement [N, meas_dim], world)
        ...
    def step(self, key, world, actions):
        # -> (measurement, world, reward_scalar, done, info)
        ...

Rules that keep it compatible

• reset/step are functional (take a key + state, return new state) so they scan/vmap.
• The measurement is what an actor senses of ITSELF/the world — not its neighbours. Neighbour information reaches an agent only through the network's delayed channel.
• reward is one shared team scalar in [0, 1] (cooperative).
• info should carry per_actor_error so the standard plots work.

Register it

# src/coadapt/envs/__init__.py
from .myenv import MyEnv
ENVS = { ..., "myenv": MyEnv }

It is now selectable as --env myenv everywhere.