I am not quite sure that if you close manually trades from a automatic system, it will be a bad idea. I mean: where is the proof of it? or is it just a believe or idea based on logic but not in solid data results? If the strategy goes down afterwards, can we ensure that it is so due to manually handling the trades? To proof so, one should have two accounts and act manually in one of them while the other keeps running without intervention.
Maybe if one does something with the system, it can help it to reorganize itself better for the current market conditions.
Another way to proof it is by inserting two random variables into the system while backtesting. The panic variable and the random manual intervention variable. The panic defines randomwly how much loss or win do we accept before closing the trade. The random manual intervention decides when the trades are closed, which is to say a random manual invervention. Playing with this and other variables we can see the effect of manual intervention. Maybe some conclusions can come from it. Maybe the DD and PF do not change much which can mean that manually intervention wont affect the long run meaning that if one really knows that the system is not adjusted to the market, the manual intervention can save losses but it wont affect the wins in the long run. Just an idea.

What nobody should do is to close trades due to panic because the trader does not like to see trades loosing. The system was already trained, so if maybe that that lossing trade will become a win one. The question maybe: Is this current market situation similar to those from the training period? If not, maybe the system cant adapt to the current market and to step into it can be helpful.

So it is possible that closing trades manually is good or bad depending on the market conditions. For that, two paralel systems should run. One with intervention and other without it but in different market conditions to see when it makes sense to close trades manualy, or if it does at all.

The system could have an alarm as well. For example: If the oil price or gold, or SPX500 or whatever or all of them are lower or higher now than the highest high or the lowest low of all the prices during the training time, the zorro window could prompt a message like: Not similar conditions haven been observed in the historical data used in the backtest. Manually intervention can now be considered.