Moving "it fixes if told" to weights: OPRD's mechanism for training long-term coding agents through scope-attached reminder distillation

In repository-scale issue resolution, not only code correctness but also "continuing to obey task constraints" and "reacting to tool responses" are effective. However, in the authors' diagnostic analysis, agents repeatedly failed in situations where they could not utilize information already present within the context for action.

Interestingly, simply inserting a reminder (advisory text) into the relevant part actually reduces this type of error. In other words, the model "fixes if told," but fails without being told.

However, inference-time reminders come with the costs of construction cost, context consumption, and dependence on scaffolds. OPRD utilizes this "fixes if told" behavior only during training to move it to the weights.