_generate_planning_step should support image inputs

Currently, _generate_planning_step only processes textual messages. In multimodal workflows, it’s often necessary to include images (e.g., screenshots or visual context) as part of the reasoning input.