ML Agent Bench crashes when an action omits Action Input #1804

Describe the bug

hr_ml_agent_bench.actions.get_action() returns an Action with args=None when a model emits an action name without an Action Input: block.

is_valid_action() only rejects string arguments, then unconditionally calls:

actual_args = action.args.keys()

so the normal malformed response:

Action: List Files

raises AttributeError: 'NoneType' object has no attribute 'keys' instead of being classified as an invalid action and re-prompted by the eval loop.

Expected behavior

Missing or otherwise non-dictionary action arguments should make is_valid_action() return False, not crash the evaluation.

Proposed fix

Require action.args to be a dictionary before comparing argument keys, and add local regression coverage for missing, malformed, and valid action inputs.