Skip to main content

Node Traversals with NodeCaret

NodeCaret offers a unified and efficient way for traversing the document tree, making it much easier to correctly implement traversals and avoid edge cases around empty nodes and collapsed selections.

These new low-level functions were all designed to work together as a fully featured relatively lightweight API to use in the core to allow us to gradually address some edge cases and then simplify and shrink the code. We expect higher-level utilities to be developed and shipped in @lexical/utils or another module at a later date. The current overhead should be less than 3kB in a production environment.

Concepts

The core concept with NodeCaret is that you can represent any specific point in the document by using an origin node, a direction that points towards an adjacent node (next or previous), and a type to specify whether the arrow points towards a sibling (breadth) or towards a child (child).

All of these types have a D type parameter that must be a CaretDirection, so you can not accidentally mix up next and previous carets. Many of them also have a T type parameter that encodes the type of the origin node.

tip

The methods of a caret are designed to operate on nodes attached to the origin in the designated direction, not the origin itself. For example, this code is a no-op because it will attach a node to the origin, and then remove the node that was just attached.

// The origin is unaffected (other than being marked dirty)
caret.insert($createTextNode('no-op')).remove();
warning

Carets are immutable, and designed for low-level usage. There is no attempt for carets to automatically update based on changes to the document (this is a common source of bugs when working with RangeSelection). Functions and methods that work with carets and are expected to change the structure of the document will always return a possibly new caret.

The origin of a caret is the exact version of the object that it was constructed with, all accessor methods on that origin will generally call origin.getLatest() so the operations will see the latest version.

NodeCaret

NodeCaret is any SiblingCaret or any ChildCaret

  • Typically constructed with $getChildCaretOrSelf($getSiblingCaret(origin, direction)) which returns a ChildCaret when the origin is an ElementNode

SiblingCaret

SiblingCaret is a caret that points towards a sibling of the origin

  • Constructed with $getSiblingCaret(origin: LexicalNode, direction: CaretDirection)
  • The next direction points towards the right
  • The previous direction points towards the left
→ direction: 'next'← direction: 'previous'
getParentAtCaret()origin.getParent()origin.getParent()
getNodeAtCaret()origin.getNextSibling()origin.getPreviousSibling()
insert(node)origin.insertAfter(node)origin.insertBefore(node)

ChildCaret

ChildCaret is a caret that points towards the first or last child of the origin

  • Constructed with $getChildCaret(origin: ElementNode, direction: CaretDirection)
  • The next direction points towards the first child
  • The previous direction points towards the last child
↘ direction: 'next'↙ direction: 'previous'
getParentAtCaret()originorigin
getNodeAtCaret()origin.getFirstChild()origin.getLastChild()
insert(node)origin.splice(0, 0, node)origin.append(node)

PointCaret

PointCaret is any TextPointCaret, SiblingCaret or ChildCaret. This type can be used to represent any point in the document that PointType can represent.

tip

Because TextPointCaret is a subclass of SiblingCaret, this type is really just used to document that the function will not ignore TextPointCaret

TextPointCaret

TextPointCaret is a specialized SiblingCaret with any TextNode origin and an offset property

  • Constructed with $getTextPointCaret(origin, direction, offset)
  • The offset property is an absolute index into the string
  • The next direction implies all text content after offset
  • The previous direction implies all text content before offset
warning

Since TextPointCaret is a specialization of SiblingCaret, the offset will be ignored by functions that are not also specialized to handle it.

TextPointCaretSlice

TextPointCaretSlice is a wrapper for TextPointCaret that provides a signed distance, it is just a data structure and has no methods.

  • Constructed with $getTextPointCaretSlice(caret, distance)
  • Math.min(caret.offset, caret.offset + distance) refers to the start offset of the slice
  • Math.max(caret.offset, caret.offset + distance) refers to the end offset of the slice
  • The direction of the caret is generally ignored when working with a TextPointCaretSlice, the slice is in absolute string coordinates

CaretRange

CaretRange contains a pair of PointCaret that are in the same direction. It is equivalent in purpose to a RangeSelection, and is what you would generally use for depth first traversals.

  • Constructed with $getCaretRange(anchor, focus) or $caretRangeFromSelection(selection)
  • The anchor is the start of the range, generally where the selection originated, and it is "anchored" in place because when a selection grows or shrinks only the focus will be moved
  • The focus is the end of the range, where the blinking cursor is, it's the current focus of the user
  • Anchor and focus must point in the same direction. The anchor points towards the first node in the range and the focus points towards the first node not in the range
  • The getTextSlices() method is essential to handle the literal edge cases where the anchor and/or focus are a TextPointCaret. These edges are not included in the default caret iteration of the CaretRange.
warning

If you are iterating a CaretRange you must consider the getTextSlices() separately, they are not included in the iteration. This is so you don't have to consider TextPointCaretSlice at every step. They are literal edge cases that can only be at the start and/or end and typically have special treatment (splitting instead of removing, for example).

Traversal Strategies

Adjacent Caret Traversals

The lowest level building block for traversals with NodeCaret is the adjacent caret traversal, which is supported directly by methods of NodeCaret.

getAdjacentCaret() - Gets a SiblingCaret for the node attached to origin in direction. If there is no attached node, it will return null

getParentCaret(rootMode) - Gets a SiblingCaret for the parent node of origin in the same direction. If there is no parent node, or the parent is a root according to rootMode, then it will return null. rootMode may be 'root' to only return null for RootNode or 'shadowRoot' to return null for RootNode or any ElementNode parent where isShadowRoot() returns true

getChildCaret() - Gets a ChildCaret for this origin, or null if the origin is not an ElementNode. Will return this if the caret is already a ChildCaret

For example, iterating all siblings:

// Note that NodeCaret<D> already implements Iterable<NodeCaret<D>> in this
// way, so this function is not very useful. You can just use startCaret as
// the iterable.
function *$iterSiblings<D extends CaretDirection>(
startCaret: NodeCaret<D>
): Iterable<SiblingCaret<D>> {
// Note that we start at the adjacent caret. The start caret
// points away from the origin node, so we do not want to
// trick ourselves into thinking that that origin is included.
for (
let caret = startCaret.getAdjacentCaret();
caret !== null;
caret = caret.getAdjacentCaret()
) {
yield caret;
}
}

Depth First Caret Traversals

The strategy to do a depth-first caret traversal is to use an adjacent caret traversal and immediately use a ChildCaret any time that an ElementNode origin is encountered. This strategy yields all possible carets, but each ElementNode in the traversal may be yielded once or twice (a ChildCaret on enter, and a SiblingCaret on leave). Allowing you to see whether an ElementNode is partially included in the range or not is one of the reasons that this abstraction exists.

function *$iterCaretsDepthFirst<D extends CaretDirection>(
startCaret: NodeCaret<D>
): Iterable<NodeCaret<D>> {
function step(prevCaret: NodeCaret<D>): null | NodeCaret<D> {
// Get the adjacent SiblingCaret
const nextCaret = prevCaret.getAdjacent();
return (
// If there is a sibling, try and get a ChildCaret from it
(nextCaret && nextCaret.getChildCaret()) ||
// Return the sibling if there is one
nextCaret ||
// Return a SiblingCaret of the parent, if there is one
prevCaret.getParentCaret('root')
);
}
// You may add an additional check here, usually some specific
// caret to terminate the iteration with (such as the parent caret
// of startCaret):
//
// `caret !== null || caret.is(endCaret)`
//
for (
let caret = step(startCaret);
caret !== null;
caret = step(caret)
) {
yield caret;
}
}

Normally this type of iteration would be done from a CaretRange, where you would specify a precise end caret (focus).

function $iterCaretsDepthFirst<D extends CaretDirection>(
startCaret: NodeCaret<D>,
endCaret?: NodeCaret<D>,
): Iterable<NodeCaret<D>> {
return $getCaretRange(
startCaret,
// Use the root as the default end caret, but you might choose
// to use startCaret.getParentCaret('root') for example
endCaret || $getBreadthNode($getRoot(), startCaret.direction)
);
}

To get all nodes that are entirely selected between two carets:

function *$iterNodesDepthFirst<D extends CaretDirection>(
startCaret: NodeCaret<D>,
endCaret?: NodeCaret<D>,
): Iterable<NodeCaret<D>> {
const seen = new Set<NodeKey>();
for (const caret of $iterCaretsDepthFirst(startCaret, endCaret)) {
const {origin} = caret;
if ($isChildCaret(caret)) {
seen.add(origin.getKey());
} else if (!$isElementNode(origin) || seen.has(origin.getKey())) {
// If the origin is an element and we have not seen it as a ChildCaret
// then it was not entirely in the CaretRange
yield origin;
}
}
}

Handling TextPointSlice

Future Direction

It's expected that higher-level abstractions will be built on top of this outside of the core, either in @lexical/utils or a separate companion package. This is just designed to be the lowest-level layer with a consistent and type-safe interface. That sort of abstraction will probably look a little bit like cheerio or jQuery, but for working with Lexical documents. It is not expected that more abstractions will be added to the core.

In order to reduce code size and eliminate bugs, more of the core will be refactored to use NodeCaret internally.

Once this happens, it's possible that the internal structure of PointType and/or RangeSelection may change to accommodate NodeCaret, as it is more resilient to document changes (only changes that directly affect the orgin node will "break" the point). A simple version of this would be to create a caret any time that the point changes, and use that caret as a fallback if the selection would otherwise be lost.

It may be the case that NodeCaret will become the lowest level API, working directly with private LexicalNode/ElementNode internals. When/if that happens, the methods on LexicalNode will remain for backwards compatibility, but overriding them will not be supported. It isn't particularly safe to override them as-is anyway, and these overrides are frequently the root cause of bugs (e.g. parents that remove themselves after an operation on a child, causing the point to be lost unless the caller was sophisticated enough to store the array of parents).

History

Before NodeCaret, Lexical's core API offered a relatively low-level DOM-like interface for working with nodes and traversing them. It has accumulated many functions over time for performing various kinds of traversals around the tree (finding ancestors, children, depth, siblings, etc.), but most of them are not implemented in a way that makes them easy to combine efficiently, and many of them have edge cases that are difficult to avoid and can't really be addressed without breaking compatibility.

Many of these functions also have a lot of edge cases, particularly around assuming the reference nodes are inclusive. Many are also left-to-right biased, don't offer an iterative version that can be aborted early or consumed on the fly, etc.

Refactoring many of these to use something like PointType would almost be sufficient for many of these use cases, but the representation of that type is inefficient and error-prone as any mutation to the tree requires that each point be manually recomputed. PointType is also directionless, forcing a specific left-to-right bias into most APIs. RangeSelection can be used in many cases because a direction can be inferred from any two different points, but that collapses with a single point. It's also impractical to use RangeSelection concurrently with mutations due to the problems with PointType.

NodeCaret was born out of frustration with these APIs and a desire to unify it all in a coherent way to simplify and reduce errors in the core.