Add a plugin

The experiment registry accepts five plugin types: method, environment, metric, visualization, and obstacle generator.

Method example

from genedynamics.experiments.framework import MethodPlugin


class MyPlannerPlugin(MethodPlugin):
    @property
    def name(self):
        return "my_planner"

    def create_planner(self, env, energy, config):
        return MyPlanner(env=env, energy=energy, **config)

    def plan(self, planner, initial_state, rng):
        result = planner.plan(initial_state, rng)
        return {
            "states": result.states,
            "actions": result.actions,
            "initial_state": initial_state,
        }

Register the instance in register_all_plugins, add the method name to the V1 catalogue only when it is intended for release, and create one recipe that proves the complete path.

Required checks

1. Keep optional dependencies lazy and raise an install command that names the correct extra.
2. Return JSON-serializable result metadata and arrays with stable dimensions.
3. Accept the seed/random key supplied by the runner.
4. Add a unit test for the adapter boundary and an integration smoke test for the runner path.
5. Record device, compile/warm-up policy, and task dimensions in benchmarks.
6. Update configuration validation, compatibility, and user docs.

Environment plugins additionally define environment creation, energy creation, state dimension, position extraction, and optional task-owned reset/execution environment behavior. Metric plugins must keep scalar outputs serializable and may expose separate artifacts through pop_artifacts().