Regarding using generalized difference-in-differences, we normally use the joint null test for 2 or 3 years before the event year. For example, in this answer, @Thomas Bilach mentioned we may perform the joint null test for 3 years before the event year.
However, how do we handle this when we have a date variable and the data is at the date-level? How many dates before the event we should go to test for parallel trends?
The equation I used in that response generalizes to any number of leads. I don't explicitly state that we should use a joint null test "3 years before the event year" specifically, though I would argue that three or more leads before some event (i.e., treatment) is a sound recommendation. And to be clear, my endorsement of three or more leads before treatment is not specific to a panel in yearly time units. I would likely suggest three or more leads if the panel was in yearly, quarterly, monthly, weekly, daily, or even hourly time units. Obviously, you should choose a frequency that's most appropriate to your research endeavor, but my advice wouldn't necessarily vary by the panel's temporal identifier. The total number of periods before treatment also matters. Suppose you only observe entities for three days before the administration of treatment. This leaves you with, at most, two event leads before the first day of treatment, assuming you omit one period as a reference.
Note that there isn't an ideal number of leads and/or lags to include in a difference-in-differences analysis, and oftentimes the structure is arbitrarily chosen. In general, the farther back in time you go, the more likely it is that your treatment and control groups were on a stable, parallel trajectory before the shock. Even if you suspect some anticipatory response to treatment, you should still demonstrate a common evolution of the trends before that anticipation period.
The choice of "how many" leads (lags) to include in your analysis should be tailored to your specific study. A minimum of three leads seems reasonable in my estimation, though you don't have to limit yourself. Remember, the goal is to assess the "joint significance" of the adoption leads. In practice, you might want to consider a slightly smaller threshold for significance once the number of pretreatment leads gets arbitrarily large.
And lastly, a large $p$-value resulting from a joint null test doesn't "prove" parallel trends. In fact, you should interpret insignificance in this context as one piece of evidence supporting—not confirming—parallel trends. We can never really "confirm" anything, as this assumption is not directly testable.
