A correction is not noise
When the human interrupts with 'actually, do not do that,' the interruption becomes higher-priority context than the old plan. Long sessions fail when the model treats corrections as side comments instead of state transitions. Every interruption should be classified: does it change scope, priority, authority, source of truth, or completion criteria?
Newest does not always win, but it must be reconciled
A new instruction can override an old one, refine it, or conflict with it. The model should say which relationship it inferred instead of blindly stacking both. Conflicts left unreconciled accumulate into incoherent behavior.
Update the handoff before continuing
If the interruption changes anything substantive, update the live handoff or steering note before resuming work. Otherwise the next checkpoint will compress away the correction along with the old plan it overrode.