Putting ReSwift Actions on the Undo Stack Using Middleware


It took me a while to grok the use of “Middleware” in ReSwift. ReSwift’s own tests illustrate what you can do with it: you can filter or modify actions before they are passed to the Reducers, for example.

I use this in TableFlip to modify the undo stack now. Every action that’s undo-able is registered with the app’s NSUndoManager.

Take this enum of actions, for example:

struct AppState: ReSwift.StateType {
    var content: String = ""

enum DocumentAction: ReSwift.Action {
    case replaceContent(String, isInitialContent: Bool)

When a document is loaded, .replaceContent is dispatched with isInitialContent set to true. This flag is useful to not make the initial display of data undo-able. I’ll show you how I use that later. Keep in mind that when I initially created an action similar to this one, the isInitialContent flag wasn’t there. I discovered the need for it only later in the process.

“Doing” actions in the context of ReSwift means you invoke a Store’s dispatch method with a supported action. Like this:

let action = DocumentAction.replaceContent("new content", isInitialContent: false)

This is supposed to change the content of the document. Maybe a text field displays “new content” now; maybe it also changes the contents of a file or fires off a network request. That doesn’t matter right now. All that matters is that this action is triggered, and that I want to undo it.

Enter NSUndoManager. Registering an action with the NSUndoManager is as simple as calling registerUndoWithTarget(_:, selector:, object:). But you’ll have to do this for the redo counterpart as well. Traditionally, people put it into the same function:

class AnObject: NSObject {
    let undoManager = ... // maybe a property of the object
    func doSomething() {
        if undoManager.undoing {
            undoManager.registerUndoWithTarget(self, selector: #selector(doSomething), object: nil)
        undoManager.registerUndoWithTarget(self, selector: #selector(doSomething), object: nil)

This easily gets out of hand.

Let’s segue to registering undo actions in a bearable fashion before we continue to deal with the undo middleware.

Controlling What’s on the Undo Stack

NSUndoManager doesn’t accept the DocumentAction itself. Its API utilizes ye olde target–action to manipulate the undo stack. To transform my internal events to things that can be undone, I figured that blocks would be a good abstraction:

class UndoAction: NSObject {

    typealias Block = () -> Void

    let undoBlock: Block
    let undoName: String?
    let redoBlock: (Block)?

    init(undoBlock: Block, undoName: String? = nil, redoBlock: (Block)? = nil) {

        self.undoBlock = undoBlock
        self.undoName = undoName
        self.redoBlock = redoBlock

    var inverted: UndoAction? {

        guard let redoBlock = self.redoBlock else { return nil }

        // No need to pass a `undoName` as it will be associated with the current action
        // by the NSUndoManager.
        return UndoAction(undoBlock: redoBlock, redoBlock: undoBlock)

The inverted property switches the undo and redo block. That’s it.

To put an UndoAction on the stack, I don’t use the undoBlock directly because that doesn’t work. Instead, I register a callback which then invokes the undoBlock:

extension NSUndoManager {

    func registerUndo(action action: UndoAction) {

        registerUndoWithTarget(self, selector: #selector(performUndo(_:)), object: action)

    @objc private func performUndo(action: UndoAction) {


        if let inverted = action.inverted {
            // When you `registerUndoWithTarget` while undoing, the result is 
            // actually put on the "Redo" side. ¯\_(ツ)_/¯
            registerUndo(action: inverted)

That’ll work. But I want to set the caption of the undo-able action in the app’s Edit menu, too, without doing much on the client side, so I utilize the amazing “double dispatch” technique – which is fancy talk for take a dependency as method parameter and send it a message with your property:

extension UndoAction {
    func register(undoManager undoManager: NSUndoManager) {
        undoManager.registerUndo(action: self)

        // Set the action name in the menu, but don't overwrite
        // the inferred name for the redo counterpart.
        if !undoManager.undoing,
            let undoName = self.undoName {


With that stuff in place, registering an action is as simple as this:

let undoManager = ...
let action = UndoAction(undoBlock: { print("undone!") })
action.register(undoManager: undoManager)

Now back to ReSwift.

Inverting User-Triggered Events

For simplicity’s sake, I only put a DocumentAction.replaceContent in this example. This particular action can be used in three cases:

  1. To display initial contents (not undoable),
  2. to show different contents in the same document window (undoable), and
  3. to undo the change and show the old contents again (redoable).

The first two cases are easy. We just have to dispatch them to the Store so the event is passed to the Reducers which in turn modify the app state. Dispatching is a one-liner for each case:

// Show initial content
store.dispatch(DocumentAction.replaceContent("Initial content", isInitialContent: true))

// Replace content because of a server callback, user request, or whatever
store.dispatch(DocumentAction.replaceContent("Replacement", isInitialContent: false))

Now the undo middleware should register the inverted action to the undo stack. Creating an inversion is still missing. For 100% content replacements, this is easy: the inversion is a .replaceContent action with the content before changes have applied.

An exemplary story goes like this:

// State = ""
store.dispatch(DocumentAction.replaceContent("Initial content", isInitialContent: true))
// → State = "Initial content"

store.dispatch(DocumentAction.replaceContent("Replacement", isInitialContent: false))
// → State = "Replacement"

// → State = "Initial content"

Here, undoLastActionSomehow() essentially equals sending .replaceContent("Initial content", isInitialContent: false) from the app state’s perspective. But the undo stack won’t work if all we do is push new messages onto it. That’s what the undone flag is good for: so that the inverted action does not end up on the undo stack, but is ignored.

Not putting the inverted DocumentAction on the undo stack again is important. NSUndoManager puts the inversion on the redo stack for you in performUndo (see code above).

Figuring out how to compute the inverted case is pretty easy in our case. It’s the action’s own responsibility to know how to do that, given a context which provides the data it needs.

extension DocumentAction {
    func inverse(context state: AppState) -> DocumentAction? {

        switch self {
        case let .Replace(_, isInitial: isInitial):
            if isInitial { return nil }
            return DocumentAction.Replace(state.content, isInitial: isInitial)

Of course TableFlip has a lot of actions that need more context than this in order to not replace 100% of the document with every change.

Distinguishing between a regular action and an action that was triggered from the undo stack is important, so that when a user triggers “Undo Replace Content”, for example, the same action isn’t pushed to the undo stack again.

/// Wrapper around `DocumentAction` to flag an action as already
/// on the Undo-stack.
private struct Undone: Action {
    let action: DocumentAction

    init(_ action: DocumentAction) {
        self.action = action

Undone is not itself the reversal of the action. (We created that above already.) It’s just a simple marker. So there’s not much logic involved to make this come together:

private extension DocumentAction {

    var undone: Undone {
        return Undone(self)

    var isUndoable: Bool {

        switch self {
        case let .replaceContent(_, isInitialContent: isInitial): return !isInitial
        default: return true

I can now mark a copy of any DocumentAction as “on the undo stack” and determine if it should be undoable with the isUndoable property.

This is how everything comes together:

extension UndoAction {

    convenience init?(documentAction: DocumentAction, 
        context: AppState, 
        dispatch: ReSwift.DispatchFunction) {

        guard let inverseAction = documentAction.inverse(context: context)
            else { return nil }

        self.init(undoBlock: { dispatch(inverseAction.undone) },
                  redoBlock: { dispatch(documentAction.undone) })

With that in place, the (sole) DocumentAction can compute its inverse; the inversion is used to register undo actions with NSUndoManager. When the “Undo” menu item is clicked, the non-inverted original action will be put on the redo stack.

The Undo Middleware

ReSwift’s Middleware type is a bit … confusing at first:

public typealias Middleware = (ReSwift.DispatchFunction?, ReSwift.GetState) 
    -> ReSwift.DispatchFunction 
    -> ReSwift.DispatchFunction

public typealias DispatchFunction = (Action) -> Any
public typealias GetState = () -> StateType?

Returning a DispatchFunction from another DispatchFunction looked bogus to me at first: why return a closure that takes an Action and have it return another closure that takes an Action? – I didn’t realize this is a totally wrong interpretation until I read the unit tests I initially mentioned.

A Middleware ends up looking like this, with lots of explicit type annotations for your orientation:

let someMiddleware: Middleware = { (dispatch: ReSwift.DispatchFunction, getState: ReSwift.GetState) in
    return { (next: ReSwift.DispatchFunction) in
        return { (action: ReSwift.Action) in
            // Pass through without doing anything
            return next(action)

You see that the innermost block is the actual transformation of type DispatchFunction, taking an action and returning a transformed state, for example.

The block in the middle is not itself a DispatchFunction but it takes one and it returns one. The one it takes as parameter is the next middleware in the chain (or, in the cast of the last middleware, the invocation of the main Reducer), hence the name next. The one it returns is your transformation.

With this structure in mind, the undo middleware I use looks similar to this:

let undoManager: NSUndoManager = ...
let undoMiddleware: Middleware = { dispatch, getState in
    return { next in
        return { action in

            // Pass already undone actions through
            if let undoneAction = action as? Undone {
                return next(undoneAction.action)

            if let action = action as? DocumentAction 
                where action.isUndoable,
                let state = getState() as? AppState,
                let dispatch = dispatch,
                let undo = UndoAction(documentAction: action, context: state, dispatch: dispatch) {

                undo.register(undoManager: undoManager)

            return next(action)

Here, we only try to register actions we control and ignore the ReSwift.Init action, for example, by passing it through.

This was quite some work to write a ReSwift Middleware which registers events with the NSUndoManager. A lot of the work was targeted at creating UndoAction, though, which can be used independently of ReSwift in your Cocoa projects to encapsulate undoable actions. The rest was tricky only because reversing content changes is a non-trivial task. With the example I created, it was very easy to implement DocumentAction.inverse(context:).

The one thing that’s on my mind now is this: doesn’t the NSUndoManager own a piece of application state now? After all, it’s the only object that has knowledge about both the undo and redo stack.

There’s a proof-of-concept ReSwift Recorder library which records events so you can time-travel from within the app. Time-traveling achieves the same effect as undo/redo. Ditching the Cocoa framework’s well-established NSUndoManager for a proof-of-concept time travel library isn’t worth it for me, although I’d have preferred the conceptual purity of recording past events purely using ReSwift components.

Browse the blog archive