Hey all,
I did a lot of focus tests the last weeks and wrote a big sampler application as well.
I will provide the source code as Gist in the PR description at one point.
First of all, things look pretty good! There are some problematic cases, and at least one bug from what I can see.
Note: All the problematic cases below also affect all other Controls that commit their value on focus loss, like DatePicker, Spinner, ...
What works:
- By default, the focus loss commit works well with all Controls out of the box
- Mnemonics work. They will first request focus before they trigger the action
Problematic cases:
- A Tab selection change is fired before the focus is requested on the TabPane. While the focus lost commit still works, it is in my opinion too late. Think about disabling a Tab when something in a Table is invalid (which we know after the commit)
- List/Tree/Table/View: Selection and focus is changed before the actual cell container will receive the focus. Even more weird: For both Tables, the selection change is fired BEFORE the focus change. List- and TreeView have the correct order. Again, in case we want to disable this Control when another Table is invalid, which we know right at the commit, we will still first trigger a selection and focus change
Buggy cases:
- As John also mentioned, non focus traversable Controls are completely broken. A non focus traversable Button or CheckBox will not request any focus. Therefore, Cells, DatePicker, ... can not commit their value. Other Controls set to non focus traversable like TextField or the TextField inside the DatePicker will request focus, the DatePicker button will not. This seems like a bug to me. Could also be something we want to ignore, as we can say: In this case, all focus loss commits are broken, so it is up to you. Feedback welcome
Other cases:
- MenuBar, Menu, MenuItem will not trigger focus at all. This might be expected. I don't know what to think about that. Feedback welcome
-- Marius
Gesendet: Freitag, 24. Oktober 2025 um 06:58
Betreff: Re: Re: Re: Allowing a cell to commit the value on focus loss
Those are good points. I will have a look. I do agree that the focus traversable behavior is questionable.
I would expect that those problems also exist for Controls like the DatePicker, since it does also commit its value on focus loss.
We might not be able to 'fix' mnemonics focus loss, in this case the responsibility is indeed an application responsibility.
-- Marius
Gesendet: Mittwoch, 15. Oktober 2025 um 18:52
Betreff: Re: [External] : Re: Allowing a cell to commit the value on focus loss
There is also the focus traversable flag that interacts with this, but perhaps there is a bug. When a button has focusTraversable set to false, clicking it will not give it focus. One may say that a property named "focus traversable" would only affect focus *traversal* with the keyboard (as I'd hardly call clicking with the mouse "traversal").
That still leaves mnemonic short-cuts and default actions for buttons though. Pretty sure those also don't focus the button and aren't intended to, yet do execute the action.
--John
On 15/10/2025 17:20, Andy Goryachev wrote:
This looks like a bug to me, really. What is the main purpose of the focus subsystem?
I know we like to reinvent the wheel, but focus in Swing works as expected, and one does get focus lost event on mouse press, and the target button gets the focus. Why should FX be different?
-andy
Hi Marius,
Focus lost is currently sort of a proxy of starting an interaction with a new control, but not all controls gain focus when interacted with. Buttons are one of those (either with mouse press or keyboard short cut), but there is I think also the scroll wheel that can interact with a control without focusing it (and perhaps even popup menu's).
I can only think of one half-baked solution to this:
- Have a new Event type that is always targetted at the current focus owner ("InterestLostEvent" ? :))
- This event is automatically fired by Scene just before an event is fired that is not targetted at the current focus owner, AND the last event fired did have the focus owner as target
What would happen in practice then would be something like:
- User edits field, keypress events go to current focus owner
- User does something else (moves mouse, scrolls, presses a hotkey, or presses a button):
- An InterestLostEvent is fired at the current focus owner BEFORE the new event is fired
- The delayed new event is now fired
- No further InterestLostEvents are fired until the focus owner has received a normal event again
- User goes back to editing after playing with the mouse; events targetted at the focus owner renew the interest in that control, and so next time an InterestLostEvent is fired again when needed
It feels a bit awkward, especially because simple things like mouse moves may trigger it already (but a mouse move may trigger something that requires the model to be up to date...); perhaps it would need to be selective in some way so one can choose to only be interested in the InterestLostEvent on focus loss and mouse clicks.
I can immediately see some problems as well. Some controls I think allow editing without focus gain/loss at all (I think some controls can be edited by just scrolling the mouse wheel over them). When should those controls "commit" their values...?
--John
On 15/10/2025 16:39, Marius Hanl wrote:
Hi John,
you are right that there might be corner cases. I hope that we could, what Andy suggests, find all cases and have a deeper look at them.
We can also check whether the focus delegation API from Michael is something that could help us here (but might be completely unrelated).
The other options as you also mentioned, also have their problems. Even debouncing a commit on every keystroke can be unreliable if the user is too fast.
I really hope we can make the focus loss reliable, as we then do not need much of an API changes inside the Cell Framework.
-- Marius
Gesendet: Montag, 13. Oktober 2025 um 19:32
Betreff: Re: Allowing a cell to commit the value on focus loss
I wonder if we should find out exactly why onFocusLost does not work in these cases, as expected. Then, if I understand the proposal correctly, we won't need any API changes.
-andy
Hi Marius,
This may be unrelated, but it may be problematic to rely on committing values using focus lost:
I've built a lot of code that relies on focus lost to "commit" values to some underlying model. However, I noticed that a focus lost handler for committing values is insufficient when an action is triggered that doesn't trigger a loss of focus. For example, if I have a field "email address" and a Button "Send Email", and I have a focus lost handler to commit the email address textfield to an underlying model, then pressing the Button will not trigger that handler and the underlying model may not have been updated with the latest edits.
Solutions to trigger the correct action are all a bit tricky or annoying:
- Query all fields for their current contents as focus lost is not entirely reliable for this purpose
- Have fields update models immediately (which would be on every key press...) -- this is not very efficient, and can get in the way of validation / model restrictions
- Have controls listen to a "COMMIT" event (this is fired at the current focus owner by the Button). This event may be veto'd if committing the value resulted in a validation error, in which case the button press is cancelled
I don't like any of these, but using the last option at the moment because I like constant updates and having to requery UI components even less...
--John
I noticed however that if you edit some field (it doesn't have to be in a table view, just a regular field), and have a focus lost handler that commits the value, that this focus lost handler is insufficient...
All,
I created an initial poc 1* to support developers to commit the cell value when the focus is lost 2* (including 3*).
More specifically, this gives the maximum flexibility to choose what should happen when the focus is lost or the editing index changed (which may happen when clicking into another cell while editing).
All information mentioned here are also in the description of the PR.
API
- Instead of calling `cancelEdit`, every cell now calls `stopEdit` when the focus is lost or the editing index changed. The default behavior is cancelling the edit, but developers can now override the behavior and allow a `commitEdit` instead
- There are multiple 'events' that can lead to a editing change. Every change will now call `stopEdit`.
It is therefore the responsibility of the developer to decide, when it makes sense to actually commit the value instead of cancelling it. This decision was made as the behavior is manipulating the editing index, but you as a developer can as well. We do not really know what intention led to e.g. a change of the editing index.
- Every `MOUSE_PRESSED` shifts the focus to the cell container, which is undesired in case of editing the cell. So this event is now consumed.
- All `TextField` cells now commit their value (instead of cancel) on focus loss
- `TextField` Escape handling was badly implemented (it was never really called, as the cell container handled Escape before)
Considerations
- I tried to make the API minimal, and without breaking changes (other than the `TextField` cells committing their values, but we may split this up)
- The Cell Container focus behavior is, well, weird right now. That is why consuming the event is needed to better support this PR. One thing we may can consider is using the `focusWithin` property instead for all 4 Cell Containers and not calling `requestFocus` for nearly every `MOUSE_PRESSED` event. If we decide so, this is needs to be done before merging this PR.
- Clicking the `ScrollBar` now commits/cancels the edit. I checked other applications and this is very common. But something I need to note here. This probably can be fixed in the same way mentioned above (`focusWithin`)
- It might be hard for a developer to exactly know the cause why `stopEdit` is called. This does not seem like a problem, as e.g. for a `TextField`, you normally register listeners for e.g. pressing the Escape key on it, so you keep full control.
Another Approach
- Another Approach I tested could be to request the focus to a cell when clicked/edited, to ensure that the focus listener is ALWAYS called before another cell will reach the editing state. Again, we probably need to change the focus handling to e.g. use the `focusWithin` property. With this approach, we can only call `stopEdit` when the focus changed (since it is now called always), but not when the editing index changed.
-- Marius
