The free shift roster builder turns a till export into shifts. When the history says a third person is needed at 11:00 and again at 14:00, but not in between, the planner had to decide each of those one-hour peaks on its own: stretch it to a whole shift, or leave it short and say so. It never asked the obvious question: why not one five-hour shift from 10:00 to 15:00 that covers both?
Today it can. --join chooses each headcount level of a day as a whole. A shift may now cover several short peaks and the quiet hours between them, or lengthen the shift next to a peak instead of adding a new one. Each hour is priced the same way as before, from how often past dates needed that extra person. A join never crosses a lunch closure. And a long block has to split into pieces no shorter than your shortest shift.
On paper, it wins
Priced on a bakery's own history, the joined plan is never dearer than the old one. It can't be: the old choices are still on the menu. At the French bakery, with a six-hour shortest shift and a short hour costing three spare ones, the joined plan changed six of the seven weekday rows and lowered the expected cost.
On weeks it had not seen, it loses a little
The useful test is the one the planner already runs on your own file. Each week is planned only from the weeks before it, then scored against what that week's sales needed. We ran it on both public bakery exports: 3 sales rates, shortest shifts of 4 and 6 hours, 3 short-hour weights, and the planner's average and percentile rows. That makes 72 paired comparisons.
- Joined was clearly cheaper in 1.
- It was clearly dearer in 16. All 16 were at the French bakery, by 0.1 to 2.3 person-hours a week, never more than 1.3% of the week's cost.
- The other 55 were within noise. That includes every comparison at the Korean bakery.
The hours a join pays for are the in-between hours, where the extra person was rarely needed. Their price rests on the thin end of the history, and that end moves from week to week. The old cut also got some joins for free: when one peak's shift happened to reach the next peak, it already covered both.
So the default stays as it was. --join remains as an option for a shop whose own backtest (--backtest --join) says otherwise. The planner's misses come from the forecast, not from how it cuts shifts, so the forecast is where the next work goes.
The numbers, the code and the script are in the venture's README ("Joining short runs") and eval/join-cut.mjs. Like everything else here, it is a planning aid, not advice about any labour law.