The development of large language models offers new possibilities for enhancing adaptive scaffolding of student learning in game-based learning environments. In this work, we present a novel framework for automatic plan generation that utilizes text-based representations of students' actions within a game-based learning environment, Crystal Island, to inform adaptive scaffolding of student goal setting and planning, which are critical elements of self-regulated learning. Plan generation is the task of automatically generating a set of low-level actions that contribute toward accomplishing a target goal given a sequence of student gameplay and their prior completed goals. We investigate the use of two pre-trained large language models, T5 and GPT-3.5, in the plan generation framework. The models utilize 144 middle school students gameplay data, encompassing a total of 11,610 event sequences, as input. The plans generated by the model are subsequently evaluated against plans crafted by students during gameplay utilizing an in-game planning support tool in Crystal Island. We compare automatically generated plans to students' manually generated in terms of the number of matching low-level actions, the number of actions that match when mapped to higher-level categories of actions, and the distribution of categories of actions within plans. Results indicate that automatically generated plans from both models largely align in terms of the high-level categories of actions that are included, but the generated plans feature fewer low-level actions than students' plans. Plans generated by T5 align more closely with student plans, whereas GPT-3.5, though not following student planning patterns, produces valid plans as well. These findings suggest that LLMs show significant promise for automatically generating plans that can be used to devise run-time adaptive scaffolding for student planning in game-based learning environments.