Time-on-task has been shown to predict student performance in Computer Science courses, making it a useful tool for teachers to identify which students need extra support. We attempt to add more granularity to our time-on-task metric by leveraging the exit code of the previous compile/run attempt to subdivide student time-on-task into time in an error state and time in an error-free state. We compare the predictive power of (1) keystroke-derived time-on-task, (2) error-free time-on-task, (3) error time-on-task, and (4) the ratio between error-free and error state time-on-task against assignment grades.