Hi Panos,
This is still the same known issue from your earlier thread, where deanrie confirmed it and suggested the .cursor/rules workaround. The team is tracking it.
Were you able to set up the rule file @deanrie suggested? If so, and the model is still ignoring it, that’s a useful data point. Starting a fresh chat when you switch back to plan mode (rather than continuing the same conversation after implementing) also tends to help, since the large prior context makes it more likely that the model drifts into agent behavior.