Status of this Document

This document is a preliminary draft of a specification for HTML editing APIs, mainly defining selections, execCommand() and related functionality. It replaces a couple of old sections of the HTML specification, and the selection part of the old DOM Range specification. Issues that don't necessarily need discussion should preferably be filed in the issue tracker. Feedback that needs discussion should preferably be sent to the public-webapps list, with the string "[editing]" somewhere in the subject, CC'd to ayg@aryeh.name to make sure I see it. I will handle all feedback I'm aware of, either addressing it immediately or filing a bug if it needs longer-term consideration and is not already filed.

Copyright © 2014 the Contributors to the HTML Editing APIs Specification, published by the W3C Editing APIs Community Group under the W3C Community Contributor License Agreement (CLA). A human-readable summary is available.

This specification, along with the accompanying JavaScript implementation and tests, may also be used under the terms of the CC0 1.0 Universal License. Thus anyone may reuse, modify, and redistribute them without restriction.

This specification was published by the W3C Editing APIs Community Group. It is not a W3C Standard nor is it on the W3C Standards Track. Please note that under the W3C Community Contributor License Agreement (CLA) there is a limited opt-out and other conditions apply. Learn more about W3C Community and Business Groups.

Table of contents

  1. Status of this Document
  2. Table of contents
  3. Introduction
  4. Tests
    1. General remarks
    2. Command development tests
    3. Command conformance tests
  5. Issues
  6. Selections
  7. Commands
    1. Properties of commands
    2. Supported commands
    3. Enabled commands
  8. Methods to query and execute commands
  9. Common definitions
  10. Common algorithms
    1. Assorted common algorithms
    2. Wrapping a list of nodes
    3. Allowed children
  11. Inline formatting commands
    1. Inline formatting command definitions
    2. Assorted inline formatting command algorithms
    3. Clearing an element's value
    4. Pushing down values
    5. Forcing the value of a node
    6. Setting the selection's value
    7. The backColor command
    8. The bold command
    9. The createLink command
    10. The fontName command
    11. The fontSize command
    12. The foreColor command
    13. The hiliteColor command
    14. The italic command
    15. The removeFormat command
    16. The strikethrough command
    17. The subscript command
    18. The superscript command
    19. The underline command
    20. The unlink command
  12. Block formatting commands
    1. Block formatting command definitions
    2. Assorted block formatting command algorithms
    3. Block-extending a range
    4. Recording and restoring overrides
    5. Deleting the selection
    6. Splitting a node list's parent
    7. Canonical space sequences
    8. Indenting and outdenting
    9. Toggling lists
    10. Justifying the selection
    11. Automatic linking
    12. The delete command
    13. The formatBlock command
    14. The forwardDelete command
    15. The indent command
    16. The insertHorizontalRule command
    17. The insertHTML command
    18. The insertImage command
    19. The insertLineBreak command
    20. The insertOrderedList command
    21. The insertParagraph command
    22. The insertText command
    23. The insertUnorderedList command
    24. The justifyCenter command
    25. The justifyFull command
    26. The justifyLeft command
    27. The justifyRight command
    28. The outdent command
  13. Miscellaneous commands
    1. The copy command
    2. The cut command
    3. The defaultParagraphSeparator command
    4. The paste command
    5. The redo command
    6. The selectAll command
    7. The styleWithCSS command
    8. The undo command
    9. The useCSS command
  14. Additional requirements
  15. Acknowledgements

Introduction

Bug 16206.

A piece of software that claims to implement this specification must follow every normative requirement in it. Every requirement in this specification is normative unless stated otherwise.

These are comments. All comments other than this one are non-normative.

This is a note. All notes other than this one are non-normative.

This is an open issue. All issues other than this one are non-normative.

The remainder of this section is not normative, and nor is any preceding section.

This specification defines commands to edit HTML documents programmatically. The APIs specified here were originally introduced in Microsoft's Internet Explorer, but have subsequently been copied by other browsers in a haphazard and imprecise fashion. Although the behavior specified here does not exactly match any browser at the time of writing, it can serve as a target to converge to in the future.

Where the reasoning behind the specification is of interest, such as when major preexisting rendering engines are known not to match it, the reasoning is available by clicking the "comments" button on the right (requires JavaScript). If you have questions about why the specification says something, check the comments first. They're sometimes longer than the specification text itself, and commonly record what the major browsers do and other essential information.

The principles I've used for writing this specification so far are:

Tests

This section is not normative.

General remarks

There are two groups of tests currently associated with this document: selection tests, and command tests. These tests are a non-normative part of this specification. The command tests are available in two formats, development tests and conformance tests. This has a lot to do with the history of how the specification was developed: selection was originally part of the defunct DOM Range spec, while the command parts were developed on their own. The selection things have been merged into the command spec, but neither the spec text nor the tests have been fully unified at the time of this writing.

Thus the selection tests live in their own directory, selecttest/. They use the standard testharness.js framework developed by James Graham. They're entirely self-contained, sharing nothing with the command tests (except invoking the testharness.js library). At the time of this writing, they're not nearly as comprehensive as the command tests, although they do test some functionality comprehensively.

On the other hand, the command-related part of this specification is developed in tandem with a more or less complete JavaScript implementation. The implementation is used for creating tests, of two different types: development tests and conformance tests. The two types of test share most of the same code, starting with the multi-thousand-line implementation itself. The actual tests run are largely the same in either case, but the way they're run is very different.

The development tests were the original ones, and were designed to assist in writing the specification from scratch. Given an input, they print out the spec's and browser's output to allow manual inspection. They can store the spec's output and will raise an alert if it changed since the last run, as a form of regression testing, but they don't print out counts of passed/failed tests. They are not designed to run exactly the same across browsers and are tolerant of minor variations. Development tests are likely not very useful for anyone other than the spec's editor.

Conformance tests were added later. They run the same tests as the development tests, but in an entirely non-interactive format, using the testharness.js framework like the selection tests. They always run the same set of tests, don't vary behavior between browsers (hopefully), and are unforgiving of any deviation. However, they're significantly more cumbersome to set up and use, so they're less useful for developing the specification itself.

There's a suite of predefined tests for each command (~30–300 at the time of this writing). These are the only tests run for the conformance tests, and in the regression tests they can be run by clicking the "Run tests" button. For the regression tests, you can also enter your own tests manually in the box provided. In any event, the test input is a snippet of HTML, which must have a selection marked in it. There are three ways to mark a selection's start or end:

  1. Square brackets mark a selection inside a text node, like <b>foo[bar]baz</b> for a selection whose start and end nodes are in the text node foobarbaz, with start offset 3 and end offset 6. Do not use square brackets where there's no text node: <b>[]</b> is bad, <b>{}</b> is correct.
  2. Curly braces mark a selection inside an element, like <b>{foobarbaz</b>} for a selection whose start node is the element <b>, whose start offset is 0, whose end node is the root of the editable region, and whose end offset is 1. Do not use curly braces in the middle of a text node: foo{bar}baz is bad, foo[bar]baz is correct.
  3. The data-start and data-end attributes mark a selection inside an element if a curly brace can't be put there in text/html. For instance, <table><tr>{<td>foo</td>}</tr></table> doesn't work because when the fragment is parsed, the curly braces end up outside the table. Instead, you have to do <table><tr data-start=0 data-end=1><td>foo</table>.

Every input must have exactly one start marker and one end marker, which will be removed from the DOM before the test is run. You can mix and match marker types, e.g., [foo}.

There is one special test type that behaves differently, "multitest". This allows running several tests in succession, which is needed at least for testing the effect of commands' state override or value override. The syntax is JSON, and looks like

[HTML input,
  [command name 1, command value 1],
  [command name 2, command value 2],
  . . .]

where all the variables are properly-quoted JSON strings. queryCommand*() are not run for multitests.

In all cases, prefixing the test string (or the first string in the test array, for arrays) with "!" has a special effect. The "!" will be stripped, and the test will be added to an array of bad tests. These tests will be omitted from the conformance tests. This is used to mark tests where the reference implementation is known to currently give bad results, either because of a bug in the reference implementation or a bug in the spec. Bad tests will still be run as development tests.

Commands that are expected to vary significantly based on the value of the CSS styling flag are run twice. The first time runs execCommand("styleWithCSS", false, "false") before every command, and the second runs execCommand("styleWithCSS", false, "true") before every command. All other commands other than multitests run execCommand("styleWithCSS", false, "false") before every command. The extra tests are not run as regression tests, in IE or Opera, because they don't implement the styleWithCSS command (but of course the conformance tests always all run).

The implementation is also used for an actual rich-text editor, but it's currently more of a toy than anything. Significant functionality is probably broken, especially outside of Gecko/WebKit. I might spend some more time getting it to work right, but it's certainly not going to be very useful on real-world sites, since there's no way any of this stuff will work in IE8.

Command development tests

The development tests are mostly useful for developing the spec itself, not for use by authors or browser implementers. They consist of a suite of fully automated tests plus a few separate suites of manual tests. The tests are:

The automated tests run the JavaScript implementation of the specification on a particular input, then run the browser's implementation on the same input for comparison. The results of running automated tests are placed in a table, with rows marked as passing or failing based on whether the browser output is "close enough" to the spec output. Since the tests are designed for debugging the spec rather than actually testing conformance, minor variations are allowed to avoid having browsers fail many tests for uninteresting reasons. Passes and fails are based only on execCommand() output, not queryCommand*(): the latter is sanity-checked, and colored green or red if it's known to be right or wrong on general principle, but spec and browser output are not compared.

The tests will optionally store the specification's result for each test in localStorage, and will raise an alert for any new test (no stored output), any test whose spec output is different from the last run, and any test whose spec output is otherwise clearly bad (e.g., producing a non-serializable DOM). This is mostly useful for debugging and regression-testing the spec itself, and is probably not interesting to anyone other than me.

When a test runs, first the code sets up a contenteditable div with the given contents and sets the selection as requested. Then it runs queryCommandIndeterm(), queryCommandState(), and queryCommandValue(), and their values are noted. Then it runs execCommand(). Finally, it runs queryCommandIndeterm(), queryCommandState(), and queryCommandValue() again. Then it adds the output to the table.

The manual tests are much like the automated tests, with some key differences. They only test one command each, with one input value (if applicable). When a test is run, everything proceeds as in the automated case, but instead of running execCommand() for the browser tests, the user is asked to hit the appropriate key (backspace, delete, enter, etc.). Thus when running the tests for the first time, the user has to hit a key repeatedly, perhaps a few hundred times. The browser's result is then cached in localStorage so no manual intervention is required on subsequent runs except for newly-added tests, but the cached entries can be cleared if necessary.

The development tests have been tested and largely work in the latest versions (at the time of this writing) of IE, Firefox, Chrome, and Opera. Since the implementation of the spec is in JavaScript, it's vulnerable to bugs in browsers' JavaScript implementations. I work around or warn about some of these, but not all. The most correct results will probably be in Firefox or Chrome: both IE and Opera have serious known bugs that corrupt spec output for many tests. The tests are still useful for reviewing the browser output, but spec output in those browsers should be sanity-checked and compared against another browser's spec output in case of doubt.

Command conformance tests

The conformance tests operate more like one would expect from tests. Once generated, they consist of a single page, conformancetest/runtest.html, which runs all the tests and prints out a table of passes and fails. Like many other recent web standards suites, they use the testharness.js framework, and browser implementers should be able to make them part of their automated regression test frameworks.

For manual inspection, a version of the conformance tests is also available that runs only some of the tests at once: conformancetest/splitruntest.html. This runs only one group of tests at a time, which takes only a few seconds instead of a minute or more. This makes debugging particular tests much faster. In all other respects, it should behave identically to runtest.html.

Unlike the development tests, the conformance tests don't run the JavaScript implementation as part of the test. Instead, the tests' expected values are generated by a separate page, conformancetest/gentest.html. That page will output the expected values for all tests in a format to be copied to conformancetest/data.js, where it will be used by runtest.html.

This separation has a few benefits. First of all, it means the conformance tests take a long time to generate, but run much faster than the development tests: on the order of one minute instead of five or more. (Browser implementations of execCommand() are currently much faster than the spec's JS reference implementation, it seems.) Second of all, it means that all browsers are running against the same expected results. Third of all, the fact that all expected results are saved in a file allows much more systematic regression testing of the spec and the reference implementation. Changes in expected results will be recorded in the spec's version history and can be matched up to changes in the reference implementation or the spec, instead of being stored transiently in the spec editor's localStorage and then lost.

There are a couple of disadvantages as well. For one thing, new tests can't easily be added live, so the conformance tests aren't useful for experimenting with how implementations work. For another thing, the tests' expected results can only be generated all at once (not per-command), which takes minutes. A third issue is that they provide no useful feedback at all about user actions such as hitting Enter: they tell you what the insertParagraph command does in the browser, but not whether it matches up to any user action. Commands like the insertText command just fail all tests in most browsers. Thus they don't replace the manual development tests at all. (Manual conformance tests will eventually be added.)

As might be expected, browsers don't generate exactly the same expected results from the tests. Clearly incorrect expected results (like a DOM that doesn't round-trip through text/html) are automatically rejected, with a printed warning, but some discrepancies remain. At the time of this writing, Firefox 8.0a2 and Chrome 15 dev generate identical results, except that Firefox omits one test (due to incorrect serialization of <xmp>) and Chrome gets one test wrong (due to a style resolution bug). Thus I generate the tests right now in Firefox, since all its generated tests are at least correct. IE9 and Opera 11.50 don't yet generate or run the tests usefully at all in this initial version, but this will be remedied soon.

Event firing is currently tested in a totally separate file, event.html. In the long term, this will probably be merged into runtest.html so that it's tested more thoroughly.

Issues

This section is not normative.

This specification is mostly feature-complete. It's more or less fully implemented in JavaScript, and has been tested on a fairly significant amount of artificial input. It has not been tested on real-world sites that use execCommand(), and has not been thoroughly reviewed by anyone other than me. It should be considered mostly stable and awaiting implementater review and feedback.

Significant known issues that I need feedback on, or otherwise am not planning to fix just yet:

A variety of other issues are also noted in the text, formatted like this. Feedback would be appreciated on all of them.

TODO:

Also TODO: Things that are only implemented by a couple of browsers and may or may not be useful to spec:

Things I haven't looked at that multiple browsers implement:

Also need to look at contenteditable=plaintext-only.

Things that would be useful to address for the future but aren't important to fix right now are in comments prefixed with "TODO".

Selections

IE9 and Firefox 6.0a2 allow arbitrary ranges in the selection, which follows what this spec originally said. However, this leads to unpleasant corner cases that authors, implementers, and spec writers all have to deal with, and they don't make any real sense. Chrome 14 dev and Opera 11.11 aggressively normalize selections, like not letting them lie inside empty elements and things like that, but this is also viewed as a bad idea, because it takes flexibility away from authors.

So I changed the spec to a made-up compromise that allows some simplification but doesn't constrain authors much. See discussion. Basically it would throw exceptions in some places to try to stop the selection from containing a range that had a boundary point other than an Element or Text node, or a boundary point that didn't descend from a Document.

But this meant getRangeAt() had to start returning a copy, not a reference. Also, it would be prone to things failing weirdly in corner cases. Perhaps most significantly, all sorts of problems might arise when DOM mutations transpire, like if a boundary point's node is removed from its parent and the mutation rules would place the new boundary point inside a non-Text/Element node. And finally, the previously-specified behavior had the advantage of matching two major implementations, while the new behavior matched no one. So I changed it back.

See bug 15470. IE9, Firefox 12.0a1, Chrome 17 dev, and Opera Next 12.00 alpha all make the range initially null.

For the stuff about defaultView, see the comments on document.getSelection().

Every document with a non-null defaultView has a unique Selection object associated with it. Selection objects are known as selections. Each selection is associated with a single range, which may be null and is initially null. This one selection must be shared by all the content of the document (though not by nested documents), including any editing hosts in the document. Editing hosts that are not inside a document cannot have a selection.

This is a requirement of the HTML spec. IE9 and Opera Next 12.00 alpha seem to follow it, while Firefox 12.0a1 and Chrome 17 dev seem not to. See test, Mozilla bug, WebKit bug.

A document's selection is a singleton object associated with that document, so it gets replaced with a new object when Document.open() is called.

See bug 15470. IE9 and Opera Next 12.00 alpha allow the user to reset the range to null after the fact by clicking somewhere; Firefox 12.0a1 and Chrome 17 dev do not. I follow Gecko/WebKit, because it lessens the chance of getRangeAt(0) throwing.

The user agent should allow the user to change the active document's selection. If the user makes any modification to a selection, the user agent must create a new range with suitable start and end and associate the selection with this new range (not modify the existing range). The user agent must not allow the user to set a selection's range to null if it was not already null.

This matches Firefox 12.0a1, as far as I can tell. Chrome 17 dev and Opera Next 12.00 alpha return copies from getRangeAt(), so the requirement isn't testable for them. IE9 threw weird exceptions in my testing, so I match the only browser I could test.

Once a selection is associated with a given range, it must continue to be associated with that same range until this specification requires otherwise.

For instance, if the DOM changes in a way that changes the range's boundary points, or a script modifies the boundary points of the range, the same range object must continue to be associated with the selection. However, if the user changes the selection or a script calls addRange(), the selection must be associated with a new range object, as required elsewhere in this specification.

This paragraph is vague. It needs to be replaced by detailed conformance requirements saying exactly what to do for particular keystrokes, like we have for backspace/delete/etc.

If the selection's range is not null and is collapsed, then the caret position must be at that range's boundary point. When the selection is not empty, this specification does not define the caret position; user agents should follow platform conventions in deciding whether the caret is at the start of the selection, the end of the selection, or somewhere else.

This short-changes Mac users. See bug 13909.

Each selection has a direction, either forwards or backwards. If the user creates a selection by indicating first one boundary point of the range and then the other (such as by clicking on one point and dragging to another), and the first indicated boundary point is after the second, then the corresponding selection must initially be backwards. Otherwise, it must be forwards (including if the user didn't create the selection, created it by selecting an entire part of the page using a keyboard shortcut, etc.).

Wouldn't it make more sense if addRange()/removeRange() reset direction?

Selections also have an anchor and a focus. If the selection's range is null, its anchor and focus are both null. If the selection's range is not null and its direction is forwards, its anchor is the range's start, and its focus is the end. Otherwise, its focus is the start and its anchor is the end.

interface Selection {
  readonly attribute Node? anchorNode;
  readonly attribute unsigned long anchorOffset;
  readonly attribute Node? focusNode;
  readonly attribute unsigned long focusOffset;

  readonly attribute boolean isCollapsed;
  void               collapse(Node node, unsigned long offset);
  void               collapseToStart();
  void               collapseToEnd();

  void               extend(Node node, unsigned long offset);

  void               selectAllChildren(Node node);
  void               deleteFromDocument();

  readonly attribute unsigned long rangeCount;
  Range              getRangeAt(unsigned long index);
  void               addRange(Range range);
  void               removeRange(Range range);
  void               removeAllRanges();

  stringifier;
};

See also nsISelection.idl from Gecko. This spec doesn't have everything from there yet, in particular selectionLanguageChange() and containsNode() are missing. They are missing because I couldn't work out how to define them in terms of Ranges.

Originally, the Selection interface was a Netscape feature. The original implementation was carried on into Gecko (Firefox), and the feature was later implemented independently by other browser engines. The Netscape implementation always allowed multiple ranges in a single selection, for instance so the user could select a column of a table. However, multi-range selections proved to be an unpleasant corner case that web developers didn't know about and even Gecko developers rarely handled correctly. Other browser engines never implemented the feature, and clamped selections to a single range in various incompatible fashions.

This specification follows non-Gecko engines in restricting selections to at most one range, but the API was still originally designed for selections with arbitrary numbers of ranges. This explains oddities like the coexistence of removeRange() and removeAllRanges(), and a getRangeAt() method that takes an integer argument that must always be zero.

All of the members of the Selection interface are defined in terms of operations on the range object (if any) represented by the object. These operations can raise exceptions, as defined for the Range interface; this can therefore result in the members of the Selection interface raising exceptions as well, in addition to any explicitly called out below.

What happens if you try to put a selection in some node that's not part of the selection's document? Assuming it works, how is it presented to the user?

What does getSelection().getRangeAt(0).detach() do?

selection . anchorNode

Returns the element that contains the start of the selection.

Returns null if there's no selection.

selection . anchorOffset

Returns the offset of the start of the selection relative to the element that contains the start of the selection.

Returns 0 if there's no selection.

selection . focusNode

Returns the element that contains the end of the selection.

Returns null if there's no selection.

selection . focusOffset

Returns the offset of the end of the selection relative to the element that contains the end of the selection.

Returns 0 if there's no selection.

The anchorNode attribute must return the context object's anchor's node, or null if the anchor is null.

The anchorOffset attribute must return the context object's anchor's offset, or 0 if the anchor is null.

The focusNode attribute must return the context object's focus's node, or null if the focus is null.

The focusOffset attribute must return the context object's focus's offset, or 0 if the focus is null.

collapsed = selection . isCollapsed()

Returns true if there's no selection or if the selection is empty. Otherwise, returns false.

selection . collapse(node, offset)

Replaces the selection with a collapsed one at the given position.

Throws an IndexSizeError exception if offset is negative or longer than node's length.

selection . collapseToStart()

Replaces the selection with an empty one at the position of the start of the current selection.

Throws an InvalidStateError exception if there is no selection.

selection . collapseToEnd()

Replaces the selection with an empty one at the position of the end of the current selection.

Throws an InvalidStateError exception if there is no selection.

selection . extend(node, offset)

Changes the focus while leaving the anchor in place.

Throws an InvalidStateError if there's no selection, an InvalidNodeTypeError if node is a doctype, and an IndexSizeError exception if offset is negative or longer than node's length.

The isCollapsed attribute must return true if the anchor and focus are the same (including if both are null). Otherwise it must return false.

The collapse(node, offset) method must create a new range, set both its start and end to (node, offset), and set the context object's range to the newly-created range.

For collapseToStart/End, IE9 mutates the existing range, while Firefox 9.0a2 and Chrome 15 dev replace it with a new one. The spec follows the majority and replaces it with a new one, leaving the old Range object unchanged.

The collapseToStart() method must throw an InvalidStateError exception if the context object's range is null. Otherwise, it must create a new range object, set both its start and end to the context object's range's start, and then set the context object's range to the newly-created range.

The collapseToEnd() method must throw an InvalidStateError exception if the context object's range is null. Otherwise, it must create a new range object, set both its start and end to the context object's range's end, and then set the context object's range to the newly-created range.

Reverse-engineered circa January 2011. IE doesn't support it, so I'm relying on Firefox (implemented extend() sometime before 2000) and WebKit (implemented extend() in 2007). I'm mostly ignoring Opera, because gsnedders tells me its implementation isn't compatible.

The extend(node, offset) method must run these steps:

  1. Gecko raises a nonstandard exception, WebKit initializes to a zero-length selection at the given point, Opera silently ignores it. Gecko's behavior of throwing wins, but with a standardized exception type.

    If the context object's range is null, throw an InvalidStateError exception and abort these steps.

  2. Let anchor and focus be the context object's anchor and focus, and let new focus be the boundary point (node, offset).
  3. Firefox 12.0a1 seems to mutate the existing range. IE9 doesn't support extend(), and it's impossible to tell whether Chrome 17 dev or Opera Next 12.00 alpha mutate or replace, because getRangeAt() returns a copy anyway. Nevertheless, I go against Gecko here, to be consistent with collapse().

    Let new range be a new range.

  4. Gecko sets the direction backwards here. Why backwards? I don't know. I'm ignoring direction for collapsed selections for now.

    If node's root is not the same as the context object's range's root, set new range's start and end to (node, offset).

  5. Otherwise, if anchor is before or equal to new focus, set new range's start to anchor, then set its end to new focus.
  6. Otherwise, set new range's start to new focus, then set its end to anchor.
  7. Set the context object's range to new range.
  8. If new focus is before anchor, set the context object's direction to backwards. Otherwise, set it to forwards.

Reverse-engineered from WebKit circa January-February 2011. They were the ones who introduced it, and so far only Gecko has copied them in Firefox 4, which isn't even a final release yet, so presumably WebKit is the one to track. Still, it probably isn't so widely used yet, so I follow Gecko when it makes more sense.


selection . selectAllChildren(node)

Replaces the selection with one that contains all the contents of the given element.

selection . deleteFromDocument()

Deletes the selection.

Based mostly on Firefox 9.0a2. It has a bug that I didn't reproduce, namely that if you pass a Document as the argument, the end offset becomes 1 instead of the number of children it has. It also throws a RangeException instead of DOMException, because its implementation predated their merging.

IE9 behaves similarly but with glitches. It throws "Unspecified error." if the node is detached or display:none, and apparently in some random other cases too. It throws "Invalid argument." for detached comments (only!). Finally, if you pass it a comment, it seems to select the whole comment, unlike with text nodes.

Chrome 16 dev behaves as you'd expect given its Selection implementation. It refuses to select anything that's not visible, so it's almost always wrong. Opera 11.50 just does nothing in all my tests, as usual.

The selectAllChildren(node) method must run the following steps:

  1. The new range replaces any existing one, doesn't mutate it. This matches IE9 and Firefox 12.0a1. (Chrome 17 dev and Opera Next 12.00 alpha can't be tested, because getRangeAt() returns a copy anyway.)

    Let range be a new range.

  2. Set range's start to (node, 0).
  3. Set range's end to (node, number of node's children).
  4. Set the context object's range to range.
  5. Set the context object's direction to forwards.

This is the one method that actually mutates the range instead of replacing it. This matches IE9 and Firefox 12.0a1. (Chrome 17 dev and Opera Next 12.00 alpha can't be tested, because getRangeAt() returns a copy anyway.)

The deleteFromDocument() method must do nothing if the context object's range is null, and otherwise must invoke the deleteContents() method on the context object's range.

Should we replace the range rather than mutating it here? This is currently the only Selection method that actually mutates an existing range in place.

selection . rangeCount

Returns the number of ranges in the selection (either 0 or 1).

selection . getRangeAt(index)

Returns the selection's range, if index is 0.

Throws an IndexSizeError exception if index is not 0, or if there is no range in the selection.

selection . addRange(range)

Adds the given range to the selection.

selection . removeRange(range)

Unselects everyting, if range is in the selection. (Use removeAllRanges() instead.)

selection . removeAllRanges()

Unselects everything.

The rangeCount attribute must return 0 if the context object's range is null, otherwise 1.

IE9 and Firefox 4.0 return the same object every time, as the spec says. Chrome 12 dev and Opera 11.10 return a different object every time.

The getRangeAt(index) method must throw an IndexSizeError exception if index is not 0, or if the context object's range is null. Otherwise, it must return a reference to (not a copy of) the context object's range. (Thus subsequent calls must return the same object if nothing has removed the context object's range in the meantime.)

IE9 and Firefox 4.0 store a reference, as described here. Chrome 12 dev and Opera 11.10 appear to store a copy, so changes don't affect the selection.

Chrome 15 dev seems to ignore addRange() if there's already a range. IE9 replaces the existing range. Firefox 9.0a2, of course, just gives you a multi-range selection. IE is likely to behave closest to Firefox, and is also more useful than silent failure, so the spec goes with that.

The addRange(range) method must set the context object's range to a reference to (not a copy of) range. Since range is added by reference, subsequent calls to getRangeAt(0) must return the same object, and any changes that a script makes to range after it is added must be reflected in the selection, until something else removes or replaces the context object's range.

The removeRange(range) method must set the context object's range to null if its range is currently range, otherwise do nothing.

Do we check for object equality here or just equality of boundary points?

The removeAllRanges() method must set the context object's range to null and its direction to forwards.

The stringifier must . . .

Complicated. The Selection stringifier is magical like innerText. See W3C bug 10583.

This specification extends several interfaces to provide entry points to the interfaces defined in this specification.

window . getSelection()
document . getSelection()

Returns the Selection object for the window, which stringifies to the text of the current selection.

partial interface Document {
  Selection getSelection();
};

Originally Gecko returned the stringification here (bug 636512 fixed that).

Now, what happens if you create a Document object with no defaultView (say via document.implementation.createHTMLDocument("")) and call getSelection() on it? IE9 seems to return a different Selection object. Firefox 12.0a1 and Opera Next 12.00 alpha return the same object as for the current window. Chrome 17 dev returns null. See discussion. There's no meaningful selection associated with such a document, so we follow WebKit and require returning null.

This leaves open the question of which documents should actually return null. We somewhat cheat by deferring the question to the definition of defaultView, even though at the time of this writing that's underdefined and not very interoperable.

The getSelection() method on the Document interface must return null if the context object's defaultView is null, and the context object's selection otherwise.

partial interface Window {
  Selection getSelection();
};

The getSelection() method on the Window interface must return the same thing as calling getSelection() on the document returned by the context object's document property.

Commands

Properties of commands

This specification defines a number of commands, identified by ASCII case-insensitive strings. Each command can have several pieces of data associated with it:

Supported commands

If you try doing anything with an unrecognized command (except queryCommandSupported), IE10 Developer Preview throws an "Invalid argument" exception. Firefox 15.0a1 throws NS_ERROR_NOT_IMPLEMENTED on querying indeterm/state/value, and returns false from execCommand/queryCommandEnabled. Chrome 19 dev returns false from everything. Opera Next 12.00 alpha throws NOT_SUPPORTED_ERR for execCommand and returns false for enabled/state/value. Originally I went with IE, although of course with a standard exception type. But after discussion (WebKit bug, Mozilla bug), I changed to match WebKit (except that I return "" for value instead of false). The issue is that there are a whole bunch of IE commands that no one else supports or wants to support, and throwing on execCommand() would make lots of pages break. WebKit was unwilling to take the compat risk, so we took the safer option.

Some commands will be supported in a given user agent, and some will not. All commands defined in this specification must be supported, except optionally the copy command, the cut command, and/or the paste command. Additional vendor-specific commands can also be supported, but implementers must prefix any vendor-specific command names with a vendor-specific string (e.g., "ms", "moz", "webkit", "opera").

I.e., no trying to look good on lazy conformance tests by just sticking in a stub implementation that does nothing.

A command that does absolutely nothing in a particular user agent, such that execCommand() never has any effect and queryCommandEnabled() and queryCommandIndeterm() and queryCommandState() and queryCommandValue() each return the same value all the time, must not be supported.

In a particular user agent, every command must be consistently either supported or not. Specifically, a user agent must not permit one page to see the same command sometimes supported and sometimes not over the course of the same browsing session, unless the user agent has been upgraded or reconfigured in the middle of a session. However, user agents may treat the same command as supported for some pages and not others, e.g., if the command is only supported for certain origins for security reasons.

Authors can tell whether a command is supported using queryCommandSupported().

Enabled commands

At any given time, a supported command can be either enabled or not. Authors can tell whether a command is currently enabled using queryCommandEnabled(). Commands that are not enabled do nothing, as described in the definitions of the various methods that invoke commands.

Testing with bold:

IE10PP2 seems to return true if the active range's start node is editable, false otherwise.

Firefox 6.0a2 seems to always return true if there's anything editable on the page, and throw otherwise. (This is bug 676401.)

Chrome 14 dev seems to behave the same as IE10PP2.

Opera 11.11 seems to always return true if there's anything editable on the page, and false otherwise.

Firefox and Opera behave more or less uselessly. IE doesn't make much sense, in that whether a command is enabled seems meaningless: it will execute it on all nodes in the selection, editable or not. Chrome's definition makes sense in that it will only run the command if it's enabled, but it doesn't make much sense to only have the command run if the start is editable.

It's not clear to me what the point of this method is. There's no way we're going to always return true if the command will do something and false if it won't. I originally just stuck with a really conservative definition that happens to be convenient: if there's nothing selected, obviously nothing will work, and we want to bail out early in that case anyway because all the algorithms will talk about the active range. If there are use-cases for it to be more precise, I could make it so.

Bug 16094 illustrated that we don't really want to be able to modify multiple editing hosts at once, nor do we want to do anything if the start and end aren't both editable, so I co-opted this definition to fit my ends.

Among commands defined in this specification, those listed in Miscellaneous commands are always enabled, except for the cut command and the paste command. The other commands defined here are enabled if the active range is not null, its start node is either editable or an editing host, its end node is either editable or an editing host, and there is some editing host that is an inclusive ancestor of both its start node and its end node.

Methods to query and execute commands

We fire events as requested in bug 13118. This is a new feature does not currently match any browser. If you are implementing this, please make sure to file any feedback as bugs. The spec is not finalized yet and can still be easily changed.

[Constructor(DOMString type, optional EditingBeforeInputEventInit eventInitDict)]
interface EditingBeforeInputEvent : Event {
  readonly attribute DOMString command;
  readonly attribute DOMString value;
};

dictionary EditingBeforeInputEventInit : EventInit {
  DOMString command;
  DOMString value;
};

[Constructor(DOMString type, optional EditingInputEventInit eventInitDict)]
interface EditingInputEvent : Event {
  readonly attribute DOMString command;
  readonly attribute DOMString value;
};

dictionary EditingInputEventInit : EventInit {
  DOMString command;
  DOMString value;
};

We have two different interfaces because we might want to add additional members to the input event but not the beforeinput event, such as a list of nodes that were affected.

When an EditingBeforeInputEvent object is created, the command and value attributes must both be initialized to the empty string, unless otherwise specified.

When an EditingInputEvent object is created, the command and value attributes must both be initialized to the empty string, unless otherwise specified.

TODO: Define behavior for show UI.

When the execCommand(command, show UI, value) method on the HTMLDocument interface is invoked, the user agent must run the following steps:

  1. If only one argument was provided, let show UI be false.
  2. If only one or two arguments were provided, let value be the empty string.
  3. For supported: see comment before Supported commands.

    For enabled: I didn't research this closely, but at a first glance, this is possibly how Chrome 14 dev and Opera 11.11 behave. Maybe also Firefox 6.0a2, except it throws if the command isn't enabled, I think. IE9 returns true in at least some cases even if the command is disabled. TODO: Is this right? Maybe we should be returning false in other cases too?

    If command is not supported or not enabled, return false.

  4. If command is not in the Miscellaneous commands section:

    We don't fire events for copy/cut/paste/undo/redo/selectAll because they should all have their own events. We don't fire events for styleWithCSS/useCSS because it's not obvious where to fire them, or why anyone would want them. We don't fire events for unsupported commands, because then if they became supported and were classified with the miscellaneous events, we'd have to stop firing events for consistency's sake.

    1. Let affected editing host be the editing host that is an inclusive ancestor of the active range's start node and end node, and is not the ancestor of any editing host that is an inclusive ancestor of the active range's start node and end node.

      Such an editing host must exist, because otherwise the command would not be enabled.

    2. Dispatch an event at affected editing host that uses the EditingBeforeInputEvent interface. The event's type attribute must be initialized to "beforeinput"; its isTrusted, bubbles, and cancelable attributes must be initialized to true; its command attribute must be initialized to command; and its value attribute must be initialized to value.
    3. If the value returned by the previous step is false, return false.
    4. If command is not enabled, return false.

      We have to check again whether the command is enabled, because the beforeinput handler might have done something annoying like getSelection().removeAllRanges().

    5. Let affected editing host be the editing host that is an inclusive ancestor of the active range's start node and end node, and is not the ancestor of any editing host that is an inclusive ancestor of the active range's start node and end node.

      This new affected editing host is what we'll fire the input event at in a couple of lines. We want to compute it beforehand just to be safe: bugs in the command action might remove the selection or something bad like that, and we don't want to have to handle it later. We recompute it after the beforeinput event is handled so that if the handler moves the selection to some other editing host, the input event will be fired at the editing host that was actually affected.

  5. Take the action for command, passing value to the instructions as an argument.
  6. If the previous step returned false, return false.
  7. If command is not in the Miscellaneous commands section, then dispatch an event at affected editing host that uses the EditingInputEvent interface. The event's type attribute must be initialized to "input"; its isTrusted and bubbles attributes must be initialized to true; its command attribute must be initialized to command; and its value attribute must be initialized to value.
  8. Return true.

When the queryCommandEnabled(command) method on the HTMLDocument interface is invoked, the user agent must run the following steps:

  1. See comment before Supported commands.

  2. Return true if command is both supported and enabled, false otherwise.

When the queryCommandIndeterm(command) method on the HTMLDocument interface is invoked, the user agent must run the following steps:

  1. For supported: see comment before Supported commands.

    What happens if you call queryCommand(Indeterm|State|Value)() on a command where it makes no sense?

    IE9 consistently returns false for all three. However, any command that has a state defined also has a value defined, which is equal to the state: it returns boolean true or false.

    Firefox 6.0a2 consistently throws NS_ERROR_FAILURE for indeterm/state if not supported, and returns an empty string for value. Exceptions include unlink (seems to always return indeterm/state false), and styleWithCss/useCss (throw NS_ERROR_FAILURE even for value).

    Chrome 14 dev returns false for all three, and even does this for unrecognized commands. It also always defines value if state is defined: it returns the state cast to a string, either "true" or "false".

    Opera 11.11 returns false for state and "" for value (it doesn't support indeterm). Like Chrome, this is even for unrecognized commands.

    Gecko's behavior is the most useful. If the author tries querying some aspect of a command that makes no sense, they shouldn't receive a value that looks like it might make sense but is actually just a constant. Originally, I went even further than Gecko: I required exceptions even for value, since doing otherwise makes no sense. But throwing more exceptions is less compatible on the whole than throwing more exceptions, so based on discussion, I switched to a behavior more like Opera, which is more or less IE/WebKit behavior but made slightly more sane.

    If command is not supported or has no indeterminacy, return false.

  2. Return true if command is indeterminate, otherwise false.

When the queryCommandState(command) method on the HTMLDocument interface is invoked, the user agent must run the following steps:

  1. See comment on the comparable line for queryCommandIndeterm().

    If command is not supported or has no state, return false.

  2. If the state override for command is set, return it.
  3. Return true if command's state is true, otherwise false.

Firefox 6.0a2 always throws an exception when this is called. Opera 11.11 seems to return false if there's nothing editable on the page, which is unhelpful. The spec follows IE9 and Chrome 14 dev. The reason this is useful, compared to just running one of the other methods and seeing if you get a NOT_SUPPORTED_ERR, is that other methods might throw different exceptions for other reasons. It's easier to check a boolean than to check exception types, especially since as of June 2011 UAs aren't remotely consistent on what they do with unsupported commands.

Actually, correction: Firefox < 15ish throws an exception if nothing editable is on the page. Otherwise it behaves just like IE/Chrome. See Mozilla bug 742240.

When the queryCommandSupported(command) method on the HTMLDocument interface is invoked, the user agent must return true if command is supported, and false otherwise.

When the queryCommandValue(command) method on the HTMLDocument interface is invoked, the user agent must run the following steps:

  1. This is what Firefox 6.0a2 and Opera 11.11 seem to do when the command isn't enabled. Chrome 14 dev seems to return the string "false", and IE9 seems to return boolean false. For the case where there's no value, or the command isn't supported, see the comment on the comparable line for queryCommandIndeterm().

    If command is not supported or has no value, return the empty string.

  2. Yuck. This is incredibly messy, as are lots of other fontSize-related things, but I don't want to define a whole second notion of value for the sake of a single command . . .

    If command is "fontSize" and its value override is set, convert the value override to an integer number of pixels and return the legacy font size for the result.

  3. If the value override for command is set, return it.
  4. Return command's value.

All of these methods must treat their command argument ASCII case-insensitively.

The methods in this section have mostly been designed so that the following invariants hold after execCommand() is called, assuming it didn't throw an exception:

The first two points do not always hold for strikethrough or underline, because it can be impossible to unset text-decoration in CSS. Also, by design, the state of insertOrderedList and insertUnorderedList might be true both before and after calling, because they only remove one level of indentation. unlink should set the value to null. And finally, the state of the various justify commands should always be true after calling, and the value should always be the appropriate string ("center", "justify", "left", or "right"). Any other deviations from these invariants are bugs in the specification.

Common definitions

An HTML element is an Element whose namespace is the HTML namespace.

A prohibited paragraph child name is "address", "article", "aside", "blockquote", "caption", "center", "col", "colgroup", "dd", "details", "dir", "div", "dl", "dt", "fieldset", "figcaption", "figure", "footer", "form", "h1", "h2", "h3", "h4", "h5", "h6", "header", "hgroup", "hr", "li", "listing", "menu", "nav", "ol", "p", "plaintext", "pre", "section", "summary", "table", "tbody", "td", "tfoot", "th", "thead", "tr", "ul", or "xmp".

These are all the things that will close a <p> if found as a descendant. I think. Plus table stuff, since that can't be a descendant of a p either, although it won't auto-close it.

A prohibited paragraph child is an HTML element whose local name is a prohibited paragraph child name.

The block/inline node definitions are CSS-based. "Prohibited paragraph child" is conceptually similar to "block node", but based on the element name. Generally we want to use block/inline node when we're interested in the visual effect, and prohibited paragraph children when we're concerned about parsing or semantics. TODO: Audit all "block node" usages to see if they need to become "visible block node", now that block nodes can be invisible (if they descend from display: none).

A block node is either an Element whose "display" property does not have resolved value "inline" or "inline-block" or "inline-table" or "none", or a document, or a DocumentFragment.

An inline node is a node that is not a block node.

An editing host is a node that is either an HTML element with a contenteditable attribute set to the true state, or the HTML element child of a document whose designMode is enabled.

Something is editable if it is a node; it is not an editing host; it does not have a contenteditable attribute set to the false state; its parent is an editing host or editable; and either it is an HTML element, or it is an svg or math element, or it is not an Element and its parent is an HTML element.

An editable node cannot be a document or DocumentFragment, its parent cannot be null, and it must descend from either an Element or a document.

The editing host of node is null if node is neither editable nor an editing host; node itself, if node is an editing host; or the nearest ancestor of node that is an editing host, if node is editable.

Two nodes are in the same editing host if the editing host of the first is non-null and the same as the editing host of the second.

Barring bugs, the algorithms here will not alter the attributes of a non-editable element; will not remove a non-editable node from its parent (except to immediately give it a new parent in the same editing host); and will not add, remove, or reorder children of a node unless it is either editable or an editing host. An editing host is never editable, so authors are assured that editing commands will only modify the editing host's contents and not the editing host itself.

A collapsed line break is a br that begins a line box which has nothing else in it, and therefore has zero height.

Is this a good definition at all? I mean things like <p>foo<br></p>, or the second one in <p>foo<br><br></p>. The way I test it is by adding a text node after it containing a zwsp; if that changes the offsetHeight of its nearest non-inline ancestor, I deem it collapsed. But what if it happens to be display: none right now, for instance? Or its ancestor has a fixed height? Would it be better to use some DOM-based definition?

TODO: The thing about li is a not very nice hack. The issue is that an li won't collapse even if it has no children at all, but that's not true in all browsers (at least not in Opera 11.11), and also it breaks assumptions elsewhere. E.g., if it gets turned into a p.

An extraneous line break is a br that has no visual effect, in that removing it from the DOM would not change layout, except that a br that is the sole child of an li is not extraneous.

Also possibly a bad definition. Again, I test by just removing it and seeing what happens. (Actually, setting display: none, so that it doesn't mess up ranges.)

A whitespace node is either a Text node whose data is the empty string; or a Text node whose data consists only of one or more tabs (0x0009), line feeds (0x000A), carriage returns (0x000D), and/or spaces (0x0020), and whose parent is an Element whose resolved value for "white-space" is "normal" or "nowrap"; or a Text node whose data consists only of one or more tabs (0x0009), carriage returns (0x000D), and/or spaces (0x0020), and whose parent is an Element whose resolved value for "white-space" is "pre-line".

node is a collapsed whitespace node if the following algorithm returns true:

This definition is also bad. It's a crude attempt to emulate CSS2.1 16.6.1, but leaving out a ton of the subtleties. I actually don't want the exact CSS definitions, because those depend on things like where lines are broken, but I'm not sure this definition is right anyway. E.g., what about a pre-line text node consisting of a single line break that's at the end of a block? That collapses, same idea as an extraneous line break. We could also worry about nodes containing only zwsp or such if we wanted, or display: none, or . . .

  1. If node is not a whitespace node, return false.
  2. If node's data is the empty string, return true.
  3. Let ancestor be node's parent.
  4. If ancestor is null, return true.
  5. If the "display" property of some ancestor of node has resolved value "none", return true.
  6. While ancestor is not a block node and its parent is not null, set ancestor to its parent.
  7. At this point we know node consists of some whitespace, of a sort that will collapse if it's at the start or end of a line. We go backwards until we find the first block boundary, and if everything until there is invisible or whitespace, we conclude that node is collapsed. We assume a block boundary is either when we hit a line break or block node, or we hit the end of ancestor (which is the nearest ancestor block node). All this is very imprecise, of course, but it's fairly simple and will work in common cases.

    We have to avoid invoking the definition of "visible" here to avoid infinite recursion: that depends on the concept of collapsed whitespace nodes. Instead, we repeat the parts we need, which turns out to be "not much of it".

    Let reference be node.

  8. While reference is a descendant of ancestor:
    1. Let reference be the node before it in tree order.
    2. If reference is a block node or a br, return true.
    3. If reference is a Text node that is not a whitespace node, or is an img, break from this loop.
  9. We found something before our text node on (probably) the same line, so presumably it's not at the line's start. Now we need to look forward and see if we're at the line's end. If we aren't there either, then we assume we're not collapsed, so return false.

    Let reference be node.

  10. While reference is a descendant of ancestor:
    1. Let reference be the node after it in tree order, or null if there is no such node.
    2. If reference is a block node or a br, return true.
    3. If reference is a Text node that is not a whitespace node, or is an img, break from this loop.
  11. Return false.

TODO: Consider whether we really want to depend on img specifically here. It seems more likely that we want something like "any replaced content that has nonzero height and width" or such. When fixing this, make sure to audit for other occurrences of this assumption.

Something is visible if it is a node that either is a block node, or a Text node that is not a collapsed whitespace node, or an img, or a br that is not an extraneous line break, or any node with a visible descendant; excluding any node with an inclusive ancestor Element whose "display" property has resolved value "none".

Something is invisible if it is a node that is not visible.

TODO: Reconsider whether we want to lump invisible nodes in here. If we don't and change the definition, make sure to audit all callers, since then a block could have collapsed block prop descendants that aren't children.

A collapsed block prop is either a collapsed line break that is not an extraneous line break, or an Element that is an inline node and whose children are all either invisible or collapsed block props and that has at least one child that is a collapsed block prop.

A collapsed block prop is something like the <br> in <p><br></p>, or the <br> and <span> in <p><span><br></span></p>. These are necessary to stop the block from having zero height when it has no other contents, but serve no purpose and should be removed once the block has other contents that stop it from collapsing.

TODO: I say "first range" because I think that's what Gecko actually does, and Gecko is the only one that allows multiple ranges in a selection. This is keeping in mind that it stores ranges sorted by start, not by the order the user added them, and silently removes or shortens existing ranges to avoid overlap. It probably makes the most sense in the long term to have the command affect all ranges. But I'll leave this for later.

The active range is the range of the selection given by calling getSelection() on the context object. (Thus the active range may be null.)

Each HTMLDocument has a boolean CSS styling flag associated with it, which must initially be false. (The styleWithCSS command can be used to modify or query it, by means of the execCommand() and queryCommandState() methods.)

Each HTMLDocument is associated with a string known as the default single-line container name, which must initially be "p". (The defaultParagraphSeparator command can be used to modify or query it, by means of the execCommand() and queryCommandValue() methods.)

For some commands, each HTMLDocument must have a boolean state override and/or a string value override. These do not change the command's state or value, but change the way some algorithms behave, as specified in those algorithms' definitions. Initially, both must be unset for every command. Whenever the number of ranges in the selection changes to something different, and whenever a boundary point of the range at a given index in the selection changes to something different, the state override and value override must be unset for every command. The value override for the backColor command must be the same as the value override for the hiliteColor command, such that setting one sets the other to the same thing and unsetting one unsets the other.

The primary purpose of state and value overrides is that if the user runs a command like bold with a collapsed selection, then types something without moving the cursor, they expect it to have the given style (bold or such). Thus the commands like bold set state and value overrides, and insertText checks for them and applies them to the newly-inserted text. Other commands like delete also interact with overrides.

See bug 16207.

When document.open() is called and a document's singleton objects are all replaced by new instances of those objects, editing state associated with that document (including the CSS styling flag, default single-line container name, and any state overrides or value overrides) must be reset.

Of course, any action that replaces a document object entirely, such as reloading the page, will also reset any editing state associated with the document.

When this specification refers to a method or attribute that is defined in a specification, the user agent must treat the method or attribute as defined by that specification. In particular, if a script has overridden a standard property with a custom one, the user agent must only use the overridden property when a script refers to it, and must continue to use the specification-defined behavior when this specification refers to it.

When a list or set of nodes is assigned to a variable without specifying the order, they must be initially in tree order, if they share a root. (If they don't share a root, the order will be specified.) When the user agent is instructed to run particular steps for each member of a list, it must do so sequentially in the list's order.

Common algorithms

Assorted common algorithms

To move a node to a new location, preserving ranges, remove the node from its original parent (if any), then insert it in the new location. In doing so, follow these rules instead of those defined by the insert and remove algorithms:

Many of the algorithms in this specification move nodes around in the DOM. The normal rules for range mutation require that any range endpoints inside those nodes are moved to the node's parent as soon as the node is moved, which would corrupt the selection. For instance, if the user selects the text "foo" and then bolds it, first we produce <b></b>foo, then <b>foo</b>. When we move the "foo" text node into its new parent, we have to do so "preserving ranges", so that the text "foo" is still selected.

  1. Let node be the moved node, old parent and old index be the old parent (which may be null) and index, and new parent and new index be the new parent and index.
  2. This is actually implicit, but I state it anyway for completeness.

    If a boundary point's node is the same as or a descendant of node, leave it unchanged, so it moves to the new location.

  3. If a boundary point's node is new parent and its offset is greater than new index, add one to its offset.
  4. If a boundary point's node is old parent and its offset is old index or old index + 1, set its node to new parent and add new indexold index to its offset.
  5. If a boundary point's node is old parent and its offset is greater than old index + 1, subtract one from its offset.

TODO: Do we want to get rid of attributes that are no longer allowed here?

To set the tag name of an Element element to new name:

This is needed because the DOM doesn't allow any way of changing an existing element's name. Sometimes we want to, e.g., convert a markup element to a span. In that case we invoke this algorithm to create a new element, move it to the right place, copy attributes from the old element, move the old element's children, and remove the old element.

  1. If element is an HTML element with local name equal to new name, return element.
  2. If element's parent is null, return element.
  3. Let replacement element be the result of calling createElement(new name) on the ownerDocument of element.
  4. Insert replacement element into element's parent immediately before element.
  5. Copy all attributes of element to replacement element, in order.
  6. While element has children, append the first child of element as the last child of replacement element, preserving ranges.
  7. Remove element from its parent.
  8. Return replacement element.

To remove extraneous line breaks before a node node:

<br> sometimes has no effect in CSS, such as in the markup foo<br><p>bar</p>. In such cases we like to remove the extra markup to keep things tidy.

  1. Let ref be the previousSibling of node.
  2. If ref is null, abort these steps.
  3. While ref has children, set ref to its lastChild.
  4. While ref is invisible but not an extraneous line break, and ref does not equal node's parent, set ref to the node before it in tree order.
  5. If ref is an editable extraneous line break, remove it from its parent.

To remove extraneous line breaks at the end of a node node:

  1. Let ref be node.
  2. While ref has children, set ref to its lastChild.
  3. While ref is invisible but not an extraneous line break, and ref does not equal node, set ref to the node before it in tree order.
  4. If ref is an editable extraneous line break:
    1. If the block ends with <span><br></span>, for instance, we want to remove the span too.

      While ref's parent is editable and invisible, set ref to its parent.

    2. Remove ref from its parent.

To remove extraneous line breaks from a node, first remove extraneous line breaks before it, then remove extraneous line breaks at the end of it.

Wrapping a list of nodes

To wrap a list node list of consecutive sibling nodes, run the following algorithm. In addition to node list, the algorithm accepts two inputs: an algorithm sibling criteria that accepts a node as input and outputs a boolean, and an algorithm new parent instructions that accepts nothing as input and outputs a node or null. If not provided, sibling criteria returns false and new parent instructions returns null.

This algorithm basically does two things. First, it looks at the previous and next siblings of the nodes in node list. If running sibling criteria on one or both of the siblings returns true, the nodes in node list are moved into the sibling(s). Otherwise, new parent instructions is run, and the result is used to wrap node list. For instance, to wrap node list in a <b>, one might invoke this algorithm with sibling criteria returning true only for <b> elements and new parent instructions creating and returning a new <b> element.

  1. We need to treat brs as visible here even if they're not, because wrapping them might be significant even if they're invisible: it can turn an extraneous line break into a non-extraneous one.

    If every member of node list is invisible, and none is a br, return null and abort these steps.

  2. If node list's first member's parent is null, return null and abort these steps.
  3. Trailing br's like this always need to go along with their line. Otherwise they'll create an extra line if we wrap in a block element, instead of vanishing as they should.

    If node list's last member is an inline node that's not a br, and node list's last member's nextSibling is a br, append that br to node list.

  4. See bug 13811, bug 14231. If there's a non-adjacent sibling that matches the sibling criteria and only invisible nodes intervene, we want to skip over the invisible nodes. For instance, bolding <b>foo</b><!--bar-->[baz] should produce <b>foo<!--bar-->baz</b>. Similarly, and more usefully, creating an ordered list with <ol><li>foo</ol> <p>[bar]</p> should produce <ol><li>foo</li> <li>[bar]</ol>, not <ol><li>foo</ol> <ol><li>[bar]</ol>.

    While node list's first member's previousSibling is invisible, prepend it to node list.

  5. While node list's last member's nextSibling is invisible, append it to node list.
  6. If the previousSibling of the first member of node list is editable and running sibling criteria on it returns true, let new parent be the previousSibling of the first member of node list.
  7. Otherwise, if the nextSibling of the last member of node list is editable and running sibling criteria on it returns true, let new parent be the nextSibling of the last member of node list.
  8. Otherwise, run new parent instructions, and let new parent be the result.
  9. This can only happen if new parent instructions is run and it returns null. This can be used to only merge with adjacent siblings, in case you don't want to create a new parent if that fails.

    If new parent is null, abort these steps and return null.

  10. Most callers will create a new element to return in new parent instructions, whose parent will therefore be null. But they can also return an existing node if that makes sense, so the nodes will be moved to an uncle or something. The toggle lists algorithm makes use of this.

    If new parent's parent is null:

    1. Insert new parent into the parent of the first member of node list immediately before the first member of node list.
    2. Basically, we want any boundary points around the wrapped nodes to go inside the wrapper. Without this step, wrapping "{}<br>" in a blockquote would go like

      {}<br>
      -> {}<blockquote></blockquote><br>
      -> {}<blockquote><br></blockquote>.

      The second line is due to range mutation rules: a boundary point with an offset equal to the index of a newly-inserted node stays put, so it remains before it. With this step, it goes like

      {}<br>
      -> {}<blockquote></blockquote><br>
      -> <blockquote></blockquote>{}<br>
      -> <blockquote>{}<br></blockquote>.

      The difference in the final step is because we move the <br> "preserving ranges". This means that adjacent boundary points get swept along with it. Previously, the <blockquote> intervened, so a boundary point after it would get taken along but one before it would not.

      Another solution that one might be tempted to consider would be to just put the wrapper after the wrapped elements. Then the boundary points would stay put, before the wrapper, so they'd still be adjacent to the nodes to be wrapped, like:

      {<p>foo</p>}
      -> {<p>foo</p>}<blockquote></blockquote>
      -> <blockquote>{<p>foo</p>}</blockquote>.

      The problem is that this completely breaks if you're wrapping multiple things and not all are selected. It would go like this:

      <p>foo</p>{<p>bar</p>}
      -> <p>foo</p>{<p>bar</p>}<blockquote></blockquote>
      -> <p>foo</p><blockquote>{<p>bar</p>}</blockquote>
      -> <blockquote>{<p>foo</p><p>bar</p>}</blockquote>.

      The last step is again because of the range mutation rules: the boundary point stays put when a new node is inserted. They're fundamentally asymmetric.

      An alternative solution would be to define the concept of moving a list of adjacent sibling nodes while preserving ranges, and handle this explicitly at a more abstract level.

      TODO: Think about this some more. Maybe there's a better way.

      If any range has a boundary point with node equal to the parent of new parent and offset equal to the index of new parent, add one to that boundary point's offset.

  11. Let original parent be the parent of the first member of node list.
  12. If new parent is before the first member of node list in tree order:
    1. If new parent is not an inline node, but the last visible child of new parent and the first visible member of node list are both inline nodes, and the last child of new parent is not a br, call createElement("br") on the ownerDocument of new parent and append the result as the last child of new parent.
    2. For each node in node list, append node as the last child of new parent, preserving ranges.
  13. Otherwise:
    1. If new parent is not an inline node, but the first visible child of new parent and the last visible member of node list are both inline nodes, and the last member of node list is not a br, call createElement("br") on the ownerDocument of new parent and insert the result as the first child of new parent.
    2. For each node in node list, in reverse order, insert node as the first child of new parent, preserving ranges.
  14. This could happen if new parent instructions returned a node whose parent wasn't null.

    If original parent is editable and has no children, remove it from its parent.

  15. Probably because both the previous and next sibling met them. We want to merge them in this case.

    If new parent's nextSibling is editable and running sibling criteria on it returns true:

    1. If new parent is not an inline node, but new parent's last child and new parent's nextSibling's first child are both inline nodes, and new parent's last child is not a br, call createElement("br") on the ownerDocument of new parent and append the result as the last child of new parent.
    2. While new parent's nextSibling has children, append its first child as the last child of new parent, preserving ranges.
    3. Remove new parent's nextSibling from its parent.
  16. Remove extraneous line breaks from new parent.
  17. Return new parent.

Allowed children

List is mostly based on current HTML5, together with obsolete elements. I mostly got the obsolete element list by testing what Firefox 5.0a2 splits when you do insertHorizontalRule.

TODO: The definitions of prohibited paragraph children and elements with inline contents should be in the HTML spec (possibly under a different name) so they don't fall out of sync. They'll do for now.

A name of an element with inline contents is "a", "abbr", "b", "bdi", "bdo", "cite", "code", "dfn", "em", "h1", "h2", "h3", "h4", "h5", "h6", "i", "kbd", "mark", "p", "pre", "q", "rp", "rt", "ruby", "s", "samp", "small", "span", "strong", "sub", "sup", "u", "var", "acronym", "listing", "strike", "xmp", "big", "blink", "font", "marquee", "nobr", or "tt".

An element with inline contents is an HTML element whose local name is a name of an element with inline contents.

TODO: This list doesn't currently match HTML's validity requirements for a few reasons:

  1. We need to handle invalid elements, which have no conformance requirements but should be treated properly. In particular, they can interfere with serialization (e.g., center cannot descend from p).
  2. Sometimes users give instructions that have to produce invalid DOMs to get the expected effect, like indenting the first item of a list.
  3. The HTML validity requirements are sometimes quite complicated.
  4. I just haven't had bothered to be systematic about it yet. I've only covered what's come up in my tests.

I deliberately allow dt to contain headers and such, in violation of HTML. If I didn't, then when the user tried to formatBlock a dt as a header, it would break apart the whole dl, which seems worse. See bug 13201.

A node or string child is an allowed child of a node or string parent if the following algorithm returns true:

Often we move around nodes, and sometimes this can result in unreasonable things like two <p>'s nested inside one another. This algorithm checks for DOMs we never want to have, so that other algorithms can avoid creating them or fix them if they do happen. The fix disallowed ancestors algorithm is one frequently-invoked caller of this algorithm.

  1. If parent is "colgroup", "table", "tbody", "tfoot", "thead", "tr", or an HTML element with local name equal to one of those, and child is a Text node whose data does not consist solely of space characters, return false.
  2. Actually, no node can occur in the DOM after plaintext, generally. But let's not get too carried away.

    If parent is "script", "style", "plaintext", or "xmp", or an HTML element with local name equal to one of those, and child is not a Text node, return false.

  3. If child is a document, DocumentFragment, or DocumentType, return false.
  4. If child is an HTML element, set child to the local name of child.
  5. If child is not a string, return true.
  6. If parent is an HTML element:
    1. Cannot be serialized as text/html. In some cases it can, like <a>foo<table><td><a>bar</a></td></table>baz</a>, but it's invalid in those cases too, so no need for complication.

      If child is "a", and parent or some ancestor of parent is an a, return false.

    2. This generally cannot be serialized either, for p. For other elements with inline contents, this serves to prevent things like <span><p>foo</p></span>, which will parse fine but aren't supposed to happen anyway.

      If child is a prohibited paragraph child name and parent or some ancestor of parent is an element with inline contents, return false.

    3. Also can't be serialized as text/html.

      If child is "h1", "h2", "h3", "h4", "h5", or "h6", and parent or some ancestor of parent is an HTML element with local name "h1", "h2", "h3", "h4", "h5", or "h6", return false.

    4. Further requirements only care about the parent itself, not ancestors, so we don't need to know the node itself.

      Let parent be the local name of parent.

  7. If parent is an Element or DocumentFragment, return true.
  8. If parent is not a string, return false.
  9. We allow children even where some intervening nodes will be inserted, like tr as a child of table.

    If parent is on the left-hand side of an entry on the following list, then return true if child is listed on the right-hand side of that entry, and false otherwise.

  10. If child is "body", "caption", "col", "colgroup", "frame", "frameset", "head", "html", "tbody", "td", "tfoot", "th", "thead", or "tr", return false.
  11. dd/dt/li will serialize fine as the child of random stuff, but it makes no sense at all, so we want to avoid it anyway.

    If child is "dd" or "dt" and parent is not "dl", return false.

  12. If child is "li" and parent is not "ol" or "ul", return false.
  13. If parent is on the left-hand side of an entry on the following list and child is listed on the right-hand side of that entry, return false.
  14. Return true.

Inline formatting commands

Inline formatting command definitions

The difference between "contained" and "effectively contained" is basically that 1) in <b>[foo]</b>, the text node and the <b> are effectively contained but not contained; and 2) in <b>f[o]o</b>, the text node is effectively contained but not contained, and the <b> is neither effectively contained nor contained.

A node node is effectively contained in a range range if range is not collapsed, and at least one of the following holds:

A modifiable element is a b, em, i, s, span, strike, strong, sub, sup, or u element with no attributes except possibly style; or a font element with no attributes except possibly style, color, face, and/or size; or an a element with no attributes except possibly style and/or href.

A simple modifiable element is an HTML element for which at least one of the following holds:

Conceptually, a simple modifiable element is a modifiable element which specifies a value for at most one command. As the names imply, inline formatting commands will try not to modify anything other than modifiable elements. For instance, <dfn> normally creates italics, but it's not modifiable, so running the italic command will not remove it: it will nest <span style="font-style: normal"> inside.

A formattable node is an editable visible node that is either a Text node, an img, or a br.

Two quantities are equivalent values for a command if either both are null, or both are strings and they're equal and the command does not define any equivalent values, or both are strings and the command defines equivalent values and they match the definition.

Two quantities are loosely equivalent values for a command if either they are equivalent values for the command, or if the command is the fontSize command; one of the quantities is one of "x-small", "small", "medium", "large", "x-large", "xx-large", or "xxx-large"; and the other quantity is the resolved value of "font-size" on a font element whose size attribute has the corresponding value set ("1" through "7" respectively).

Loose equivalence needs to be used when comparing effective command values to other values, while regular equivalence is used in other cases. The effective command value for fontSize is converted to pixels, so comparing it to a specified value literally would produce false negatives. But a specified value in pixels is actually different from a specified value like "small" or "x-large", because there is no precise mapping from such keywords to pixels.

If a command has inline command activated values defined but nothing else defines when it is indeterminate, it is indeterminate if among formattable nodes effectively contained in the active range, there is at least one whose effective command value is one of the given values and at least one whose effective command value is not one of the given values.

For bold and similar commands, IE 9 RC seems to consider the state true or false depending on the first element. All other browsers follow the same general idea as the spec, considering a range bold only if all text in it is bold, and this seems to match at least OpenOffice.org's bold feature. Opera 11.11 seemingly doesn't take CSS into account, and only looks at whether something descends from a <b>. I couldn't properly test IE9 because it threw exceptions (Error: Unspecified error.) on most of the tests I ran. But what I have here seems to match Firefox 6.0a2 in every case, and Chrome 14 dev in all cases with a few exceptions.

If a command has inline command activated values defined, its state is true if either no formattable node is effectively contained in the active range, and the active range's start node's effective command value is one of the given values; or if there is at least one formattable node effectively contained in the active range, and all of them have an effective command value equal to one of the given values.

Testing with hiliteColor: Opera 11.11 seems to always return the effective command value of the active range's start node. Chrome 14 dev returns boolean false consistently, bizarrely enough. Firefox 6.0a2 seems to follow the same idea as the spec, but it likes to return "transparent", including sometimes when the answer really clearly should not be "transparent". IE9 throws exceptions most of the time for backColor, so I can't say for sure, but in the few cases where it doesn't throw it returns a random-looking number, so I'll assume it's crazy like for foreColor.

I decided on something that would guarantee the following invariant: whenever you execute a command with a value provided (assuming value is relevant), queryCommandValue() will always return something equivalent to what you set.

If a command is a standard inline value command, it is indeterminate if among formattable nodes that are effectively contained in the active range, there are two that have distinct effective command values. Its value is the effective command value of the first formattable node that is effectively contained in the active range; or if there is no such node, the effective command value of the active range's start node; or if that is null, the empty string.

The notions of inline command activated values and standard inline value commands are mostly just shorthand to avoid repeating the same boilerplate half a dozen times.

Assorted inline formatting command algorithms

The effective command value of a node node for a given command is returned by the following algorithm, which will return either a string or null:

This is logically somewhat like CSS computed or resolved values, and in fact for most commands it's identical to CSS resolved values (see the end of the algorithm). We need a separate concept for some commands where we can't rely on CSS for some reason: createLink and unlink aren't CSS-related at all, backColor and hiliteColor need special treatment because background-color isn't an inherited property, subscript and superscript rely on <sub>/<sup> instead of CSS vertical-align, and strikethrough and underline don't map to unique CSS properties.

  1. If neither node nor its parent is an Element, return null.
  2. If node is not an Element, return the effective command value of its parent for command.
  3. If command is "createLink" or "unlink":
    1. While node is not null, and is not an a element that has an href attribute, set node to its parent.
    2. If node is null, return null.
    3. Return the value of node's href attribute.
  4. If command is "backColor" or "hiliteColor":
    1. While the resolved value of "background-color" on node is any fully transparent value, and node's parent is an Element, set node to its parent.
    2. What happens if everything's background is fully transparent? See bug. If you do queryCommandValue() for backColor/hiliteColor when no backgrounds are set anywhere, so it goes up to the root element, no two engines agree. IE10PP2 just returns a random-looking number (16777215). Firefox 8.0a2 returns "transparent", Chrome 15 dev returns "rgba(0, 0, 0, 0)", and Opera 11.50 returns "rgb(255, 255, 255)".

      Opera's behavior is incorrect. The current page might be an iframe, in which case the background really will be transparent and the (inaccessible) background color of the parent page will show through. Also, the user might have changed their preferences, but I'm not too worried about that.

      So instead we just return the value as-is. This means that it will be fully transparent, which is perhaps somewhat useless information, but it's the best we can do. More generally, any non-opaque value is not going to tell you what you actually want, namely "what color is the user actually seeing?" We have no realistic way to work around this in the general case.

      Return the resolved value of "background-color" for node.

  5. If command is "subscript" or "superscript":
    1. Let affected by subscript and affected by superscript be two boolean variables, both initially false.
    2. While node is an inline node:
      1. Firefox 6.0a2 ignores vertical-align for this purpose, and only cares about <sub> and <sup> tags themselves. Opera 11.11 is similar, and in fact behaves like that even for commands like bold. The spec originally followed Chrome 14 dev, mainly because WebKit itself will produce spans with vertical-align sub or super, and we want to handle them correctly. However, Ryosuke informs me that WebKit's behavior here is viewed as a bug, so I changed it to match Gecko/Opera.

        If node is a sub, set affected by subscript to true.

      2. Otherwise, if node is a sup, set affected by superscript to true.
      3. Set node to its parent.
    3. If affected by subscript and affected by superscript are both true, return the string "mixed".
    4. If affected by subscript is true, return "subscript".
    5. If affected by superscript is true, return "superscript".
    6. Return null.
  6. If command is "strikethrough", and the "text-decoration" property of node or any of its ancestors has resolved value containing "line-through", return "line-through". Otherwise, return null.
  7. If command is "underline", and the "text-decoration" property of node or any of its ancestors has resolved value containing "underline", return "underline". Otherwise, return null.
  8. Return the resolved value for node of the relevant CSS property for command.

The specified command value of an Element element for a given command is returned by the following algorithm, which will return either a string or null:

This is logically somewhat like CSS inline style. In addition to the caveats for effective command value, we also treat elements like <b> and <font> as having the same meaning as <span>s with inline style set, because they're logically pretty much the same and can in fact be produced by the same command depending on the CSS styling flag.

  1. If command is "backColor" or "hiliteColor" and the Element's display property does not have resolved value "inline", return null.
  2. If command is "createLink" or "unlink":
    1. If element is an a element and has an href attribute, return the value of that attribute.
    2. Return null.
  3. If command is "subscript" or "superscript":
    1. If element is a sup, return "superscript".
    2. If element is a sub, return "subscript".
    3. Return null.
  4. If command is "strikethrough", and element has a style attribute set, and that attribute sets "text-decoration":
    1. If element's style attribute sets "text-decoration" to a value containing "line-through", return "line-through".
    2. Return null.
  5. If command is "strikethrough" and element is an s or strike element, return "line-through".
  6. If command is "underline", and element has a style attribute set, and that attribute sets "text-decoration":
    1. If element's style attribute sets "text-decoration" to a value containing "underline", return "underline".
    2. Return null.
  7. If command is "underline" and element is a u element, return "underline".
  8. Let property be the relevant CSS property for command.
  9. If property is null, return null.
  10. If element has a style attribute set, and that attribute has the effect of setting property, return the value that it sets property to.
  11. If element is a font element that has an attribute whose effect is to create a presentational hint for property, return the value that the hint sets property to. (For a size of 7, this will be the non-CSS value "xxx-large".)
  12. If element is in the following list, and property is equal to the CSS property name listed for it, return the string listed for it.
  13. Return null.

To record the values of a list of nodes node list:

When we move nodes around, we often change their parents. If their parents had any styles applied, this will make the nodes' styles change too, which often isn't what we want. For instance, if something is wrapped in <blockquote style="color: red">, and a script runs the outdent command on it, the blockquote will be removed and the style will go along with it. Recording the values of its children first, then restoring them afterward, will ensure the nodes don't change color when outdented.

  1. Let values be a list of (node, command, specified command value) triples, initially empty.
  2. As with removeFormat, we put subscript first so it doesn't interfere with fontSize, and omit superscript because it's redundant with subscript.

    For each node in node list, for each command in the list "subscript", "bold", "fontName", "fontSize", "foreColor", "hiliteColor", "italic", "strikethrough", and "underline" in that order:

    1. Let ancestor equal node.
    2. If ancestor is not an Element, set it to its parent.
    3. While ancestor is an Element and its specified command value for command is null, set it to its parent.
    4. If ancestor is an Element, add (node, command, ancestor's specified command value for command) to values. Otherwise add (node, command, null) to values.
  3. Return values.

To restore the values specified by a list values returned by the record the values algorithm:

  1. For each (node, command, value) triple in values:
    1. Let ancestor equal node.
    2. If ancestor is not an Element, set it to its parent.
    3. While ancestor is an Element and its specified command value for command is null, set it to its parent.
    4. If value is null and ancestor is an Element, push down values on node for command, with new value null.
    5. Otherwise, if ancestor is an Element and its specified command value for command is not equivalent to value, or if ancestor is not an Element and value is not null, force the value of command to value on node.

Clearing an element's value

To clear the value of an Element element:

The idea is to remove any specified command value that the element might have for the command. This might involve changing its attributes, setting its tag name, or removing it entirely while leaving its children in place. The key caller is set the selection's value, which clears the values of everything in the selection before doing anything else to keep the markup tidy.

  1. Let command be the current command.
  2. If element is not editable, return the empty list.
  3. We want to abort early so that we don't try unsetting background-color on a non-inline element.

    If element's specified command value for command is null, return the empty list.

  4. If element is a simple modifiable element:
    1. Let children be the children of element.
    2. For each child in children, insert child into element's parent immediately before element, preserving ranges.
    3. Remove element from its parent.
    4. Return children.
  5. If command is "strikethrough", and element has a style attribute that sets "text-decoration" to some value containing "line-through", delete "line-through" from the value.
  6. If command is "underline", and element has a style attribute that sets "text-decoration" to some value containing "underline", delete "underline" from the value.
  7. If the relevant CSS property for command is not null, unset that property of element.
  8. If element is a font element:
    1. If command is "foreColor", unset element's color attribute, if set.
    2. If command is "fontName", unset element's face attribute, if set.
    3. If command is "fontSize", unset element's size attribute, if set.
  9. If element is an a element and command is "createLink" or "unlink", unset the href property of element.
  10. If we get past this step, we're something like <b class=foo> where we want to keep the extra attributes, so we stick them on a span.

    If element's specified command value for command is null, return the empty list.

  11. Set the tag name of element to "span", and return the one-node list consisting of the result.

Pushing down values

This algorithm goes up to just below the nearest ancestor with the right style, then re-applies the bad styles repeatedly going down, omitting the things we want to have the new style. This is basically what WebKit does, although WebKit sometimes starts higher up and therefore makes more intrusive changes, often creating more markup. IE follows the same general approach too.

Gecko instead seems to start breaking up elements from the bottom, so that the range consists of a few consecutive siblings, and it can then break up the problematic element into a maximum of two pieces. The spec's approach seems to create fewer elements and simpler markup (or at least markup that's no more complex) in most cases I throw at it.

Gecko's approach does have the major advantage that it gets underlines right in many cases for free. E.g.,

<u>foo<font color=red>[bar]baz</font></u>
-> <u>foo</u><font color=red>bar<u>baz</u></font> (spec)
-> <u>foo</u><font color=red>bar</font><u><font color=red>baz</font></u> (Gecko)

The spec's markup here is much shorter and contains fewer elements, but is wrong: the underline under "baz" has changed color from black to red. It might be worth trying to copy Gecko's results in such cases, but that won't solve all underline problems, so perhaps it's not worth it.

Opera also seems to break up the markup surrounding the range, but even more aggressively: even if it doesn't need to pull down styles. In some cases this does actually result in shorter markup, specifically if the existing tags are short (like i or b) and we're adding tags that are long (like span with a style attribute).

To push down values to a node node, given a new value new value:

The idea here is that if an undesired value is being propagated from an ancestor, we remove that style from the ancestor and re-apply it to all the descendants other than node. This way we don't have to have nested styles, which is usually more cluttered (although not always).

  1. Let command be the current command.
  2. E.g., a text node child of a document fragment.

    If node's parent is not an Element, abort this algorithm.

  3. If the effective command value of command is loosely equivalent to new value on node, abort this algorithm.
  4. Let current ancestor be node's parent.
  5. Let ancestor list be a list of nodes, initially empty.
  6. While current ancestor is an editable Element and the effective command value of command is not loosely equivalent to new value on it, append current ancestor to ancestor list, then set current ancestor to its parent.
  7. If ancestor list is empty, abort this algorithm.
  8. Let propagated value be the specified command value of command on the last member of ancestor list.
  9. We can only remove specified values, so if the value isn't specified, give up. Unless we're actually trying to push down a null specified value, like for unlink.

    If propagated value is null and is not equal to new value, abort this algorithm.

  10. If we go all the way up to the root and still don't have the desired value, pushing down values is pointless. It will create extra markup for no purpose. Except if the value is null, which basically just means "try to get rid of anything affecting the current element but don't aim for any specific value".

    Nevertheless, Chrome 14 dev does seem to do this. Running bold on <span style=font-weight:300>f[o]o</span> breaks up the span and adds a <b> as a sibling. In IE9, Firefox 6.0a2, and Opera 11.50, it instead nests the <b> inside the <span>. It's a tradeoff: WebKit's behavior is better for things like

    <font color=red>fo[o</font><font color=blue>b]ar</font>
    -> <font color=red>fo</font><font color=green>[ob]</font><font color=blue>ar</font>

    (where the spec adds two extra font tags instead of one), but the spec is simpler for things like

    <font color=red>f[o]o</font>
    -> <font color=red>f<font color=green>[o]</font>o</font>

    (where WebKit splits the existing tag up in addition to creating a new tag). I'm not particularly sure which approach is better overall, so I'll go with the majority of browsers. If these algorithms move to use runs of consecutive siblings instead of doing everything node-by-node, it might make sense to break up the parent as long as it won't create an extra node (i.e., we're styling something that includes the first or last child).

    If the effective command value of command is not loosely equivalent to new value on the parent of the last member of ancestor list, and new value is not null, abort this algorithm.

  11. While ancestor list is not empty:
    1. Let current ancestor be the last member of ancestor list.
    2. Remove the last member from ancestor list.
    3. If the specified command value of current ancestor for command is not null, set propagated value to that value.
    4. Let children be the children of current ancestor.
    5. If the specified command value of current ancestor for command is not null, clear the value of current ancestor.
    6. For every child in children:
      1. If child is node, continue with the next child.
      2. TODO: This will be incorrect for relative font sizes. If the font size on the parent was removed and the font size on the child is in ems or percents or something, it will now change value. This isn't likely to come up, so we'll ignore it for now.

        If child is an Element whose specified command value for command is neither null nor equivalent to propagated value, continue with the next child.

      3. If child is the last member of ancestor list, continue with the next child.
      4. Force the value of child, with command as in this algorithm and new value equal to propagated value.

Forcing the value of a node

To force the value of a node node to new value:

This algorithm checks if the node has the desired value, and if not, it wraps the node (or, if that's not possible, its descendants) in a simple modifiable element. After forcing the value, descendants might still have a different value.

  1. Let command be the current command.
  2. If node's parent is null, abort this algorithm.
  3. If new value is null, abort this algorithm.
  4. If node is an allowed child of "span":
    1. Even if the value matches, we stick it in a preceding sibling if possible. This ensures "a<cite>b</cite>c" -> "<i>a<cite>b</cite>c</i>" instead of "<i>a</i><cite>b</cite><i>c</i>". While we're at it, we also handle more elaborate cases like <b>foo</b>[bar]<b>baz</b> and even <i><b>foo</b></i>[bar]<i><b>baz</b></i> (the latter becomes <b><i>foo</i>bar<i>baz</i></b>).

      Theoretically this algorithm could pointlessly reorganize the DOM in the event of unreasonable style rules, but it's not a big enough deal for us to care, since the resulting style will still be right.

      Reorder modifiable descendants of node's previousSibling.

    2. Reorder modifiable descendants of node's nextSibling.
    3. The new parent instructions we'd want are too complicated to reasonably feed into the wrap algorithm, so we just let them return null and do the wrapping ourselves if sibling criteria didn't return true.

      Wrap the one-node list consisting of node, with sibling criteria returning true for a simple modifiable element whose specified command value is equivalent to new value and whose effective command value is loosely equivalent to new value and false otherwise, and with new parent instructions returning null.

  5. Bug 13996.

    If node is invisible, abort this algorithm.

  6. If the effective command value of command is loosely equivalent to new value on node, abort this algorithm.
  7. If node is not an allowed child of "span":
    1. Let children be all children of node, omitting any that are Elements whose specified command value for command is neither null nor equivalent to new value.
    2. This means that if it has no children, we do nothing. IE9 inserts an empty wrapper element in that case, but I'm not sure what the point is, and no one else does, so I don't. WebKit seems to ignore the node if its only child consists solely of whitespace, but I don't see any grounds for that and no one else does, so I don't.

      Force the value of each node in children, with command and new value as in this invocation of the algorithm.

    3. Abort this algorithm.
  8. If the effective command value of command is loosely equivalent to new value on node, abort this algorithm.
  9. Let new parent be null.
  10. If the CSS styling flag is false:
    1. If command is "bold" and new value is "bold", let new parent be the result of calling createElement("b") on the ownerDocument of node.
    2. If command is "italic" and new value is "italic", let new parent be the result of calling createElement("i") on the ownerDocument of node.
    3. TODO: Actual UAs use strike, not s, but s is shorter and HTML5 makes strike invalid. I've gone with s for now, but maybe we want to change the spec to require strike.

      If command is "strikethrough" and new value is "line-through", let new parent be the result of calling createElement("s") on the ownerDocument of node.

    4. If command is "underline" and new value is "underline", let new parent be the result of calling createElement("u") on the ownerDocument of node.
    5. See comment for foreColor for discussion. TODO: Define more carefully what happens when things are out of range or not integers or whatever.

      If command is "foreColor", and new value is fully opaque with red, green, and blue components in the range 0 to 255:

      1. Let new parent be the result of calling createElement("font") on the ownerDocument of node.
      2. Set the color attribute of new parent to the result of applying the rules for serializing simple color values to new value (interpreted as a simple color).
    6. If command is "fontName", let new parent be the result of calling createElement("font") on the ownerDocument of node, then set the face attribute of new parent to new value.
  11. If command is "createLink" or "unlink":
    1. Let new parent be the result of calling createElement("a") on the ownerDocument of node.
    2. Set the href attribute of new parent to new value.
    3. Nested a elements are bad, because they can't be serialized to text/html. hrefs should already have been cleared in a previous step, but we might have <a name> or such lurking about.

      Let ancestor be node's parent.

    4. While ancestor is not null:
      1. TODO: This will mean any link-specific attributes will be transferred, which makes them both invalid and useless. Is that okay? I don't really want to list them all, because that sort of list is prone to bitrot.

        If ancestor is an a, set the tag name of ancestor to "span", and let ancestor be the result.

      2. Set ancestor to its parent.
  12. WebKit is the only engine that ever outputs anything but font tags for fontSize. For size=7, it uses font-size: -webkit-xxx-large. We just output a font tag no matter what for size=7.

    If command is "fontSize"; and new value is one of "x-small", "small", "medium", "large", "x-large", "xx-large", or "xxx-large"; and either the CSS styling flag is false, or new value is "xxx-large": let new parent be the result of calling createElement("font") on the ownerDocument of node, then set the size attribute of new parent to the number from the following table based on new value:

  13. We always use sup/sub elements, even in CSS mode, following Gecko and contradicting WebKit. This is because <span value="vertical-align: sub/super">, the obvious equivalent (and what WebKit uses), behaves quite differently: it doesn't reduce font-size, which is ugly. WebKit's behavior is a bug anyway.

    If command is "subscript" or "superscript" and new value is "subscript", let new parent be the result of calling createElement("sub") on the ownerDocument of node.

  14. If command is "subscript" or "superscript" and new value is "superscript", let new parent be the result of calling createElement("sup") on the ownerDocument of node.
  15. If new parent is null, let new parent be the result of calling createElement("span") on the ownerDocument of node.
  16. This preserves boundary points correctly, as usual.

    Insert new parent in node's parent before node.

  17. If the effective command value of command for new parent is not loosely equivalent to new value, and the relevant CSS property for command is not null, set that CSS property of new parent to new value (if the new value would be valid).

    Need to be explicit. I think "if the new value would be valid" means "if the new value isn't xxx-large for font-size", need to double-check.

  18. If command is "strikethrough", and new value is "line-through", and the effective command value of "strikethrough" for new parent is not "line-through", set the "text-decoration" property of new parent to "line-through".
  19. If command is "underline", and new value is "underline", and the effective command value of "underline" for new parent is not "underline", set the "text-decoration" property of new parent to "underline".
  20. Append node to new parent as its last child, preserving ranges.
  21. If node is an Element and the effective command value of command for node is not loosely equivalent to new value:
    1. Insert node into the parent of new parent before new parent, preserving ranges.
    2. Remove new parent from its parent.
    3. Let children be all children of node, omitting any that are Elements whose specified command value for command is neither null nor equivalent to new value.
    4. Force the value of each node in children, with command and new value as in this invocation of the algorithm.

To reorder modifiable descendants of a node node, given a command command and a value new value:

  1. Let candidate equal node.
  2. While candidate is a modifiable element, and candidate has exactly one child, and that child is also a modifiable element, and candidate is not a simple modifiable element or candidate's specified command value for command is not equivalent to new value, set candidate to its child.
  3. If candidate is node, or is not a simple modifiable element, or its specified command value is not equivalent to new value, or its effective command value is not loosely equivalent to new value, abort these steps.
  4. While candidate has children, insert the first child of candidate into candidate's parent immediately before candidate, preserving ranges.
  5. If candidate had no children, any boundary point inside it will get moved to its parent here, which is okay. We don't want to preserve ranges, because that would move boundary points that originally were in candidate but were moved to its parent by the last step to move to node's parent.

    We move to after node so that boundary points before and after node wind up consistently inside candidate when we move preserving ranges. If we had

    {<node>foo<candidate></candidate></node>}

    it thus becomes

    {<node>foo</node>}<candidate></candidate>

    by the range mutation rules, and then when we move preserving ranges, it becomes

    <candidate>{<node>foo</node>}</candidate>

    which is reasonable.

    If we had inserted candidate before node, instead it would go

    {<candidate></candidate><node>foo</node>}
    {<candidate><node>foo</node>}</candidate>

    because of the interaction of regular range mutation rules with preserving-ranges rules.

    Insert candidate into node's parent immediately after node.

  6. Append the node as the last child of candidate, preserving ranges.

Setting the selection's value

To set the selection's value to new value:

The effect of this algorithm is to ensure that all nodes effectively contained in the selection have the value requested, producing the simplest markup possible to achieve that effect. It's inspired by the approach WebKit takes. The only places where the algorithm should fail are when there's an !important CSS rule that conflicts with the requested style (which we don't try to override because we assume it's !important for a reason), or when it's literally impossible to succeed (such as when a text-decoration or link URL is propagated from a non-editable ancestor). Any other failures are bugs.

First, if a node has a specified command value for the command, we unset it (clear its value). This step also removes simple modifiable elements entirely, and replaces elements like b or font with spans if they aren't simple modifiable elements. This will be sufficient if the desired value is inherited from an ancestor, or if it's the default (like font-style: normal) and no conflicting value is inherited from an ancestor. Even if clearing values doesn't actually fix the style of the node we're dealing with, we do it anyway to simplify the generated markup.

If clearing values didn't work, and it looks like an ancestor has a specified command value that we're inheriting, we push the value down from that ancestor. Thus if we're unbolding the letter "r" in

<b>foo <i>bar</i> baz</b>,

we get

<b>foo </b><i><b>ba</b>r</i><b> baz</b>.

If we didn't push down values, the final step (forcing values) would instead give us

<b>foo <i>ba<span style="font-weight: normal">r</span></i> baz</b>,

which is much longer and uglier. We take care not to disturb the style or semantics of anything but the node we're dealing with.

We'll only push down values if some ancestor actually has the value we want, so we can inherit it. Otherwise, it will just create useless markup.

Finally, if neither of the above strategies worked, we have to add new markup to get the desired value (forcing the value). First we try just sticking it into its previous or next sibling, if that's a simple modifiable element (so it won't add any styles or semantics we don't want). Otherwise, we create a new simple modifiable element and wrap it in that. It's common that a previous sibling is the simple modifiable element we want, because often we'll set the value of several consecutive siblings in succession. In that case, the element created for the first can be reused for the later ones.

This last step works a bit differently if the node isn't an allowed child of "span". In that case, wrapping it in a simple modifiable element would make the document less conforming than it already was, or would cause other problems. Instead, we recursively force the value of its children. The recursion will terminate when we hit a node that's an allowed child of "span", or when there are no further descendants. (In the latter case, there are no descendants that are text nodes or such, so we don't really need to style anything.)

After all this, the node is guaranteed to have the value we want, barring bugs in the algorithm or the two exceptions noted earlier (!important style rules, and impossible cases). We then re-run the algorithm on each child recursively. Typically this means just clearing the value of each descendant, because it should then inherit the value we just set on its ancestor. In the unusual case that a descendant's value is wrong even after we clear its value, such as because of a non-inline style rule (like trying to unbold a heading), we'll repeat the above steps to ensure that the value really gets set as desired.

  1. Let command be the current command.
  2. IE9 seems to wrap the whole line instead, or something like that, although it does nothing for createLink. We follow all other browsers' general behavior: change the state/value, and then make sure that takes effect if the user types something before changing the cursor position.

    If there is no formattable node effectively contained in the active range:

    1. If command has inline command activated values, set the state override to true if new value is among them and false if it's not.
    2. If command is "subscript", unset the state override for "superscript".
    3. If command is "superscript", unset the state override for "subscript".
    4. If new value is null, unset the value override (if any).
    5. Otherwise, if command is "createLink" or it has a value specified, set the value override to new value.
    6. Abort these steps.
  3. The last sentence here just prettifies the resulting range a bit.

    If the active range's start node is an editable Text node, and its start offset is neither zero nor its start node's length, call splitText() on the active range's start node, with argument equal to the active range's start offset. Then set the active range's start node to the result, and its start offset to zero.

  4. If the active range's end node is an editable Text node, and its end offset is neither zero nor its end node's length, call splitText() on the active range's end node, with argument equal to the active range's end offset.
  5. Let element list be all editable Elements effectively contained in the active range.
  6. For each element in element list, clear the value of element.
  7. We skip non-editable nodes.

    IE9
    Allows everything to be modified by execCommand(), regardless of whether it's editable.
    Firefox 4.0
    Ignores execCommand() if the start and end of the selection are not both editable. If the start and end are editable but something in the middle is not, seems to relocate the non-editable part in the middle or something like that.
    Chrome 12 dev
    Ignores execCommand() if the start and end of the selection are not both editable. If the start and end are editable but something in the middle is not, applies the given command but skips the non-editable parts. But the state doesn't ignore the non-editable parts, so if you bold such a selection you can't unbold it, for instance, since the middle part will remain bold (so it will keep on trying to bold it instead of switching to unbold).
    Opera 11.00
    Ignores execCommand() if the start and end of the selection are not both editable. If the start and end are editable but something in the middle is not, applies the command to everything, even the non-editable part.

    I chose to go with the non-IE behavior, per discussion. Ignoring non-editable things is convenient for the common use-case of an editor, where you don't want the user to bold random parts of the UI when they hit the bold button. For cases where it's not desired, you can always turn designMode on briefly before using execCommand(), so the non-IE behavior is a lot easier to work around than the IE behavior.

    I don't see the value in ever just ignoring execCommand(). If the start and end are not editable, I'm going to say you should still style any editable nodes in between. I'm also going to ignore non-editable nodes for the purposes of determining state, so (for instance) if all the editable nodes are bolded, it will unbold instead of bolding.

    Let node list be all editable nodes effectively contained in the active range.

  8. TODO: This is inefficient. It would be most efficient to only push down values on the highest-level effectively contained nodes, and to batch operations so we handle runs of adjacent siblings at once. Should we bother fixing this?

    For each node in node list:

    1. Push down values on node.
    2. If the node isn't an allowed child of "span", forcing its value will just force its children's value, which is redundant. So don't.

      If node is an allowed child of "span", force the value of node.

The backColor command

For historical reasons, backColor and hiliteColor behave identically.

We have three behaviors to choose from for this one:

  1. Chrome 11 dev and IE 9 RC treat it the same as hiliteColor (although IE 9 RC doesn't support hiliteColor itself).
  2. Firefox 4 in non-CSS mode sets the bgcolor of the nearest td or body, or something like that. In testing, it seems to jump out of contenteditable elements to style non-editable ancestors, which is alarming.
  3. Firefox 4 in CSS mode and Opera 11 set the background of the nearest block container, although it doesn't seem to be very dependable (probably I just don't get what exactly it's doing).

(1) is obviously redundant, but has plurality support, so we could spec it that way if the other ways were useless.

(3) is incoherent from a user perspective. For instance, if you try it on paragraphs the background will have big gaps where the margins are. If you try it on an inline element that's a child of the editing host, it will do nothing or apply the background to everything or such, even though such an inline element is visually indistinguishable from one sitting inside a div. This would only make sense if we take considerable effort to ensure that block elements all have no margins, or if we wrap things in a div if they have margins, or something like that.

That leaves (2). That might be useful if it actually set the document's background color, but it seems like it sets table cell backgrounds sometimes instead, which is really confusing.

The path of least resistance is to standardize this as meaning the same thing as hiliteColor, and make up new commands if we want to do things like set the document background color. See hiliteColor for comments.

Action:

  1. If value is not a valid CSS color, prepend "#" to it.
  2. If value is still not a valid CSS color, or if it is currentColor, return false.
  3. Set the selection's value to value.
  4. Return true.

Standard inline value command

Relevant CSS property: "background-color"

Equivalent values: Either both strings are valid CSS colors and have the same red, green, blue, and alpha components, or neither string is a valid CSS color.

The bold command

If the selection is collapsed (but not if it contains nothing but is not collapsed), IE9 wraps the whole line in a <strong>. This seems bizarre and no one else does it, so I don't do it. It's a similar story for similar commands (fontName, italic, etc.). Except not for strikethrough, where it just does nothing if the selection is empty. Why strikethrough? I don't know.

Action: If queryCommandState("bold") returns true, set the selection's value to "normal". Otherwise set the selection's value to "bold". Either way, return true.

The cutoff of 600 matches Chrome 14 dev. The cutoff used by IE9 and Firefox 6.0a2 seems to be 500, and the distinction isn't relevant for Opera 11.11 (it doesn't use CSS here at all AFAICT). On my test systems with default fonts, Chrome 14 dev displays 700 and up as bold, while the other three display 600 and up as bold.

Thus in Chrome on my system, the bold command will behave a bit oddly the first time you hit it if there's anything in the range with font-weight: 600, but it will look right in other browsers. On the other hand, if I followed IE/Firefox, it would look wrong on all my browsers for font-weight: 500.

700 actually makes more sense: then you'd view 100-300 as light, 400-600 as medium, 700-900 as bold. But that's not how it seems to work in browsers, so I'll go with 600 as the cutoff.

Inline command activated values: "bold", "600", "700", "800", or "900"

Relevant CSS property: "font-weight"

Equivalent values: Either the two strings are equal, or one is "bold" and the other is "700", or one is "normal" and the other is "400".

If the selection doesn't contain anything (meaning, e.g., deleteContents() doesn't change anything), then Chrome 12 dev inserts a link at the selection start, with the text equal to the link URL. Other browsers don't do it, so I don't either.

IE10PP2, Firefox 7.0a2, Chrome 14 dev, and Opera 11.50 all do not support indeterminate, state, or value for createLink or unlink. I previously defined indeterminate and value anyway because they make sense, but then undefined them. The nontrivial thing is what value to return if there's no link, since any string can occur as a link href, in principle.

What are the use-cases for indeterm, state, or value for createLink/unlink?

Action:

  1. Firefox 4b11 and Chrome 11 dev both silently do nothing in this case. IE 9 RC and Opera 11 both treat the request literally. Gecko and WebKit probably have it right here: users who enter no URL are very unlikely to want to link to a relative URL resolving to the current document. If they really want to, they can always specify "#" for the value, or the author can rewrite it, so it's not like this makes the API less useful.

    If value is the empty string, return false.

  2. There are three approaches here. For instance, if you ask browsers to create a link to "http://example.org" on the "b" here:

    <a href=http://example.com><b>Abc</b></a>

    Chrome 10 dev produces:

    <b><a href=http://example.com>A</a><a href=http://example.org>b</a><a href=http://example.com>c</a></b>

    Firefox 4b11 produces (roughly):

    <a href=http://example.com><b>A<a href=http://example.org>b</a>c</b></a>

    (This doesn't round-trip through text/html serialization.) IE 9 RC and Opera 11 produce simply:

    <a href=http://example.org><b>Abc</b></a>

    The last behavior probably best matches user expectations. If you happen to miss out a character when selecting the link you want to change, do you really intend to only change the link of part of it?

    For each editable a element that has an href attribute and is an ancestor of some node effectively contained in the active range, set that a element's href attribute to value.

  3. Set the selection's value to value.
  4. Return true.

The fontName command

UAs differ a bit in the details here:

IE 9 RC
Empty string sets <font face="">
Firefox 4b11
Empty string does nothing
Chrome 11 dev
Empty string does nothing, '"monospace"' same as 'monospace' (i.e., cannot escape font-family keywords because quotes are stripped, clearly wrong)
Opera 11
Empty string sets <font face="">

Setting an empty font-family has the effect of inheriting the font from the parent (although I don't see where the February 24, 2011 CSS 3 Fonts draft says that). Thus it makes sense that if we special-case this, it should be to unset the font somehow.

Special-casing the empty string to do nothing doesn't make sense to me. With createLink we'd expect the user to enter the URL themselves, so it makes sense to special-case clicking OK without entering anything. But here it's very likely that the font list will be fixed by the author (how many users will understand CSS font-family syntax?), so I don't think such usability concerns apply.

Action: Set the selection's value to value, then return true.

The value is complicated.

IE 9 RC
Always the empty string. Not very useful.
Firefox 4b11
Confusing. Sometimes it returns generic family names, like "sans-serif". Sometimes it gives specific font names, like "tt" when the font is specified as "monospace". Sometimes it gives the literal font-family string. Not sure what it's doing here.
Chrome 11 dev
Gives the literal value of font-family, except if it's inherited from default values (no explicit style declarations anywhere), when it seems to return the exact font name.
Opera 11
Returns the literal value of font-family, except if it's inherited from default values, when it returns the empty string.

I'm just going to punt on this and say it should be the resolved value of font-family. I'll leave CSSOM to decide what that means if there are no applicable style rules.

Standard inline value command

Relevant CSS property: "font-family"

The fontSize command

IE 9
Parses the value as a number (allowing floating-point), rounds to the nearest integer, then clamps to the range 1 to 7. If the value is not a valid number, including if it has trailing characters (like "2em"), does nothing. Normalizes relative sizes, so "+0" is the same as "+3", etc. Treats empty string the same as "1".
Firefox 4.0
Passes the value through literally to <font size=>, so "2em" gets you <font size="2em">. Always uses <font>, even with styleWithCss true. Ignores the command if the value is the empty string.
Chrome 12 dev
Parses the value as a legacy font size, so "2em" becomes "2", then outputs a <font> with the resulting number. If there is no resulting number, like for a value of "xx-small", does nothing. In styleWithCss mode, outputs a span with corresponding CSS keywords: 1 = x-small, 2 = small, . . ., 6 = xx-large, 7 = -webkit-xxx-large. Normalizes relative sizes, so "+0" is the same as "3", etc. Ignores the command if the value is the empty string.
Opera 11
Parses the value as an integer (ignoring floating-point as trailing characters), then outputs that. This means that "+0" becomes <font size=0> instead of <font size=+0> or <font size=3>. Non-numeric values get interpreted as 0. Does not clamp, and is willing to output negative numbers. Treats empty string as "0".

What all of these have in common is that they force the author to deal with legacy font values and don't let them use CSS. This is undesirable, but to avoid it we'd really have to create a new command. If nothing else, the value returned by queryCommandValue() has to be numeric, so authors can't really use the command sanely no matter what we do. See bug 14251.

Note that 1 is the same size as x-small in browsers, not xx-small, contrary to the CSS Fonts spec.

Action:

  1. Strip leading and trailing whitespace from value.
  2. If value is not a valid floating point number, and would not be a valid floating point number if a single leading "+" character were stripped, return false.
  3. If the first character of value is "+", delete the character and let mode be "relative-plus".
  4. Otherwise, if the first character of value is "-", delete the character and let mode be "relative-minus".
  5. Otherwise, let mode be "absolute".
  6. Apply the rules for parsing non-negative integers to value, and let number be the result.
  7. If mode is "relative-plus", add three to number.
  8. If mode is "relative-minus", negate number, then add three to it.
  9. If number is less than one, let number equal 1.
  10. If number is greater than seven, let number equal 7.
  11. Set value to the string here corresponding to number:

    The entry for 7 here is an issue: there's no CSS value that corresponds to it. Even if we got one added to the drafts, it wouldn't be backward-compatible to use it. WebKit is the only engine that supports CSS output for fontSize, and it uses -webkit-xxx-large in this case, which is unworkable. Instead, we just always output a font tag for size 7. If authors want conforming markup, they'll need to give CSS sizes above size 7, not legacy sizes.

  12. Set the selection's value to value.
  13. Return true.

This follows Firefox 6.0a2. Chrome 14 dev always returns false. Note that indeterminacy here keys off the effective command value, while the value is based only on an approximation (a number from one to seven). Thus it's possible for every subrange of the selection to have the same value, but for the selection to still be indeterminate. Setting the fontSize to the value will make it determinate without changing anything's value.

Indeterminate: True if among formattable nodes that are effectively contained in the active range, there are two that have distinct effective command values. Otherwise false.

IE9
Seems to return a number based on the computed font-size, but only if it's exactly right, otherwise it returns null. Something like that.
Firefox 6.0a2
Seemingly goes up to the nearest ancestor that's a <font size> and returns the literal value of that attribute, or "" if there's no such ancestor.
Chrome 14 dev
Gets the computed font-size in pixels, and rounds to the nearest <font size> equivalent, rounding up in the event of a tie. Except that if it's small enough, it returns "0", which doesn't make sense because that behaves the same as "1".
Opera 11.11
Like Firefox, except it returns "3" if there's no <font size> ancestor, and it converts relative values to absolute ("+1" -> "4").

Chrome's behavior seems the most useful. As usual, IE returns a variable type and all other browsers return strings, and we follow other browsers.

If the selection isn't someplace editable, Chrome works like usual; some other browsers behave differently. I see no reason to behave differently.

Value:

  1. If the active range is null, return the empty string.
  2. See comment for standard inline value commands on how I decided on this choice of node.

    Let pixel size be the effective command value of the first formattable node that is effectively contained in the active range, or if there is no such node, the effective command value of the active range's start node, in either case interpreted as a number of pixels.

  3. Return the legacy font size for pixel size.

Relevant CSS property: "font-size"

The legacy font size for an integer pixel size is returned by the following algorithm:

  1. Let returned size be 1.
  2. While returned size is less than 7:
    1. Let lower bound be the resolved value of "font-size" in pixels of a font element whose size attribute is set to returned size.
    2. Let upper bound be the resolved value of "font-size" in pixels of a font element whose size attribute is set to one plus returned size.
    3. Let average be the average of upper bound and lower bound.
    4. If pixel size is less than average, return the one-code unit string consisting of the digit returned size.
    5. Add one to returned size.
  3. Return "7".

The foreColor command

Color interpretations:

                        IE10PP2       Firefox 7.0a2            Chrome 14 dev            Opera 11.50
blue                    blue          blue                     #0000ff                  #0000ff
f                       #f            -                        -                        #f00000
#f                      #f            -                        -                        #f00000
00f                     #00f          -                        #0000ff                  #00000f
#00f                    #00f          rgb(0, 0, 255)           #0000ff                  #00000f
0000ff                  #0000ff       -                        #0000ff                  #0000ff
#0000ff                 #0000ff       rgb(0, 0, 255)           #0000ff                  #0000ff
000000fff               #0000ff       -                        -                        -
#000000fff              #0000ff       -                        -                        -
rgb(0, 0, 255)          rgb(0,0,255)  rgb(0, 0, 255)           #0000ff                  #00b000
rgb(0%, 0%, 100%)       rgb(0,0,255)  rgb(0, 0, 255)           #0000ff                  #00b000
rgb( 0 ,0 ,255)         rgb(0,0,255)  rgb(0, 0, 255)           #0000ff                  #00b000
rgba(0, 0, 255, 0.0)    #ba0000       rgba(0, 0, 255, 0)       rgba(0, 0, 255, 0)       #00ba00
rgb(15, -10, 375)       rgb(15,0,255) rgb(15, 0, 255)          #0f00ff                  #00b015
rgba(0, 0, 0, 1)        #ba0010       rgb(0, 0, 0)             -                        #00ba00
rgba(255, 255, 255, 1)  #000055       rgb(255, 255, 255)       #ffffff                  #00ba02
rgba(0, 0, 255, 0.5)    #ba0000       rgba(0, 0, 255, 0.5)     rgba(0, 0, 255, 0.5)     #00ba00
hsl(240, 100%, 50%)     #000150       rgb(0, 0, 255)           #0000ff                  #000024
cornsilk                cornsilk      cornsilk                 #fff8dc                  #fff8dc
potato quiche           #0000c0       -                        -                        #000a00
transparent             transparent   -                        rgba(0, 0, 0, 0)         #00a000
currentColor            #c0e000       currentcolor             rgba(0, 0, 0, 0)         #c000e0

The interpretations given for Firefox are only in styleWithCSS mode. In non-styleWithCSS mode, it just outputs the string literally as the <font color> attribute value, which can lead to different results. The given output for Chrome is for <font>; the output in styleWithCSS mode is the same, but rgb() is used instead of hex notation, and "transparent" and "currentcolor" are passed through under those names. IE and Opera only support <font> to begin with.

Conclusions:

What I'm going to say is that it either has to be a valid CSS color, or prefixing it with # must result in a valid CSS color. For <font>, I'll say that the output color should be normalized to #xxxxxx form. If the color is not a simple color (fully opaque with all channels between 0 and 255), I'll force style="" even if styleWithCSS mode is off. Some of this disagrees with all browsers, but it's unlikely to hurt and it makes sense.

Action:

  1. TODO: Define "valid CSS color" (here and in other color places).

    If value is not a valid CSS color, prepend "#" to it.

  2. currentColor is bad for the same reason as relative font sizes. It will confuse the algorithm, and doesn't seem very useful anyway.

    If value is still not a valid CSS color, or if it is currentColor, return false.

  3. Set the selection's value to value.
  4. Return true.

Opera 11 seems to return true for the state if there's some color style applied, false otherwise, which seems fairly useless; authors want to use value here, not state. So I'll match other browsers and not define any state.

For value, the spec essentially matches Firefox 6.0a2 and Chrome 14 dev, as far as how to decide what color the node has. IE9 seems to always return the number 0 for some bizarre reason. There are some cases where Firefox returns the empty string for some reason, and it seems to select the active node a little differently. Opera uses #xxxxxx format for getComputedStyle() but rgb() here, and also drops the transparent part of the color if there is any.

Standard inline value command

Relevant CSS property: "color"

Equivalent values: Either both strings are valid CSS colors and have the same red, green, blue, and alpha components, or neither string is a valid CSS color.

The hiliteColor command

For historical reasons, backColor and hiliteColor behave identically.

IE 9 RC doesn't support this. It uses backColor instead, but Gecko and Opera treat that differently, while all non-IE browsers treat hiliteColor the same, so I'm standardizing hiliteColor as the way to highlight text.

This is slightly tricky, because background-color does different things on block and inline elements. Given the name ("hiliteColor"), we really only want to apply it to inline elements. This is how everyone but Gecko behaves, but Gecko sometimes applies it to blocks too. WebKit doesn't set it on non-inline elements, but does clear it and push it down from them.

The spec doesn't do any of these: background-color on non-inline elements is not touched by hiliteColor, neither created nor removed. If users want to remove the style, they need to use removeFormat. Adding it usually makes no sense; see the comment for backColor.

For color parsing, see the comment for foreColor.

See bug 13829.

Action:

  1. If value is not a valid CSS color, prepend "#" to it.
  2. currentColor is bad for the same reason as relative font sizes. It will confuse the algorithm, and doesn't seem very useful anyway. For hiliteColor you could conceive of it being useful, but it will still confuse the algorithm, so ban it for now anyway.

    If value is still not a valid CSS color, or if it is currentColor, return false.

  3. Set the selection's value to value.
  4. Return true.

For indeterminacy, this follows no one. Firefox 6.0a2 and Chrome 14 dev both always return false. However, the spec makes sense, since it's consistent with other commands.

Standard inline value command

Relevant CSS property: "background-color"

Equivalent values: Either both strings are valid CSS colors and have the same red, green, blue, and alpha components, or neither string is a valid CSS color.

The italic command

Action: If queryCommandState("italic") returns true, set the selection's value to "normal". Otherwise set the selection's value to "italic". Either way, return true.

Inline command activated values: "italic" or "oblique"

Relevant CSS property: "font-style"

The removeFormat command

See bug, and also research by Ryosuke for WebKit.

Tested in IE 9, Firefox 4.0, Chrome 12 dev, Opera 11.00.

All elements whose default rendering is display: block are left untouched by all browsers (although IE seems to throw an exception on <marquee> for some reason).

It's not clear to me why we should leave <a> alone, but everyone but Opera does. In OpenOffice.org 3.2.1, doing "Default Formatting (Ctrl+M)" doesn't remove links. In Microsoft Word 2007, doing "Clear Formatting" also doesn't remove links. Verdict: don't remove links. Apparently they don't logically qualify as "formatting".

Conclusion: IE/WebKit is a solid majority by market share and they're closely interoperable, since WebKit copied IE here. Also, it makes more sense to assume that unrecognized elements don't represent any kind of inline formatting, i.e., have a blacklist of elements to remove instead of a whitelist to keep. Thus I remove more or less the same things as IE/WebKit.

I remove blink because IE does it and it makes sense, although Chrome doesn't; I remove abbr although only Firefox does, for consistency with acronym; and I remove bdi and mark because they're evidently left alone only because they're unrecognized. Finally, I remove span because otherwise, something like <span style="font-variant: small-caps"> will be left intact, which isn't expected and matches no browser except IE. (Chrome doesn't remove spans in general, but it does remove spans with style attributes, or something like that.)

Browsers will split up all these inline elements if the selection is contained within them. Opera does strip unrecognized elements with display: block if they're within the selection, but doesn't split them up if they contain the selection.

Chrome 14 dev removes style attributes from every element in the range, but IE10PP2, Firefox 7.0a2, and Opera 11.50 do not, so I go with them. As noted above, this means I need to remove spans. I could conceivably change to remove only spans with style attributes, but it doesn't seem worth it: I'll just match Gecko.

TODO: This has to be kept in sync when new HTML elements are added. I need to figure out some way of coordinating this.

A removeFormat candidate is an editable HTML element with local name "abbr", "acronym", "b", "bdi", "bdo", "big", "blink", "cite", "code", "dfn", "em", "font", "i", "ins", "kbd", "mark", "nobr", "q", "s", "samp", "small", "span", "strike", "strong", "sub", "sup", "tt", "u", or "var".

Action:

  1. Let elements to remove be a list of every removeFormat candidate effectively contained in the active range.
  2. For each element in elements to remove:
    1. While element has children, insert the first child of element into the parent of element immediately before element, preserving ranges.
    2. Remove element from its parent.
  3. The last sentence just prettifies the resulting range a bit.

    If the active range's start node is an editable Text node, and its start offset is neither zero nor its start node's length, call splitText() on the active range's start node, with argument equal to the active range's start offset. Then set the active range's start node to the result, and its start offset to zero.

  4. If the active range's end node is an editable Text node, and its end offset is neither zero nor its end node's length, call splitText() on the active range's end node, with argument equal to the active range's end offset.
  5. Let node list consist of all editable nodes effectively contained in the active range.
  6. TODO: Splitting the parent is really a block algorithm. It's not clear whether it's desirable to use for inline nodes. Perhaps it's okay, but it makes me a little uneasy.

    For each node in node list, while node's parent is a removeFormat candidate in the same editing host as node, split the parent of the one-node list consisting of node.

  7. This step is for cases like <p style=font-weight:bold>foo[bar]baz</p>, where splitting/removing tags won't help. We don't need to run superscript, since subscript does the same thing here. We run subscript first so <sub>/<sup> won't upset fontSize.

    For each of the entries in the following list, in the given order, set the selection's value to null, with command as given.

    1. subscript
    2. bold
    3. fontName
    4. fontSize
    5. foreColor
    6. hiliteColor
    7. italic
    8. strikethrough
    9. underline
  8. Return true.

The strikethrough command

TODO: See underline TODO.

Action: If queryCommandState("strikethrough") returns true, set the selection's value to null. Otherwise set the selection's value to "line-through". Either way, return true.

Inline command activated values: "line-through"

The subscript command

Action:

  1. Call queryCommandState("subscript"), and let state be the result.
  2. Set the selection's value to null.
  3. If state is false, set the selection's value to "subscript".
  4. Return true.

Indeterminate: True if either among formattable nodes that are effectively contained in the active range, there is at least one with effective command value "subscript" and at least one with some other effective command value; or if there is some formattable node effectively contained in the active range with effective command value "mixed". Otherwise false.

For <sup><sub>foo</sub></sup>, Firefox 6.0a2 and Opera 11.11 say the state is true for both superscript and subscript, and indeterminate is false; Chrome 14 dev says it's true for subscript but not superscript, and indeterminate is false. We follow neither of these behaviors: we return false for both states, and say indeterminate is true. The reason is because we want to return true for a state if we'll do nothing, false if we'll do something; and if we have nesting like this, we'll always do something, namely get rid of all those ancestors and replace them with a single tag. This matches what happens in other indeterminate situations, so it's fair to consider it indeterminate.

Inline command activated values: "subscript"

The superscript command

Action:

  1. Call queryCommandState("superscript"), and let state be the result.
  2. Set the selection's value to null.
  3. If state is false, set the selection's value to "superscript".
  4. Return true.

Indeterminate: True if either among formattable nodes that are effectively contained in the active range, there is at least one with effective command value "superscript" and at least one with some other effective command value; or if there is some formattable node effectively contained in the active range with effective command value "mixed". Otherwise false.

Inline command activated values: "superscript"

The underline command

TODO: There are a lot of problems with underline color and thickness, because text-decoration in CSS is horrible. These aren't prohibitive for normal use and existing browsers don't handle them either, so fixing these problems or working around them can be put off for now.

Action: If queryCommandState("underline") returns true, set the selection's value to null. Otherwise set the selection's value to "underline". Either way, return true.

Inline command activated values: "underline"

IE 9 RC unlinks the whole link you're pointing at, while others only unlink the current text. The latter behavior seems less expected, as with createLink, although I can't articulate precisely why. Word 2007 and OpenOffice.org 3.2.1 (Ubuntu) seem to give an option to remove the whole link or none of it, which backs the spec's requirement. See also #whatwg logs starting at 2011-05-13 at 16:53 EDT (UTC-0400).

See comment for the createLink command about indeterm/state/value.

Action:

  1. Let hyperlinks be a list of every a element that has an href attribute and is contained in the active range or is an ancestor of one of its boundary points.
  2. Clear the value of each member of hyperlinks.
  3. Return true.

Block formatting commands

Block formatting command definitions

An indentation element is either a blockquote, or a div that has a style attribute that sets "margin" or some subproperty of it.

We need to allow stuff that sets border/padding because WebKit (Chrome 12 dev) sets "border: none; padding: 0px" when indenting. We need to allow stuff that sets classes because WebKit sets class="webkit-indent-blockquote". We need to allow stuff that sets dir because IE9 does. The criteria could probably be tightened up a bit to reduce false positives, but it'll do for now.

A simple indentation element is an indentation element that has no attributes except possibly

The notions of indentation element and simple indentation element parallel those of modifiable element and simple modifiable element.

listing and xmp are included because otherwise insertParagraph inside them won't work, since paragraphs aren't an allowed child.

A non-list single-line container is an HTML element with local name "address", "div", "h1", "h2", "h3", "h4", "h5", "h6", "listing", "p", "pre", or "xmp".

A single-line container is either a non-list single-line container, or an HTML element with local name "li", "dt", or "dd".

The block node of a node node is either a block node or null, as returned by the following algorithm:

  1. While node is an inline node, set node to its parent.
  2. Return node.

Bug 14062. See also Mozilla bug 590640, specifically comments 48 and on.

If a command preserves overrides, then before taking its action, the user agent must record current overrides. After taking the action, if the active range is collapsed, it must restore states and values from the recorded list.

All block commands preserve overrides except the insertText command, which treats overrides specially.

Assorted block formatting command algorithms

TODO: When breaking a non-inline element out of an inline element, like p in b or whatever, it would make sense to re-wrap the contents in the inline tag.

To fix disallowed ancestors of node:

We often run this algorithm after we move a node someplace, just in case it wound up somewhere it's not supposed to be. This avoids things like unserializable DOMs, blocks nested inside inlines, etc.

  1. If node is not editable, abort these steps.
  2. This case is really intended to handle stuff like list items or table cells that wander outside their proper place. We generally convert them into ps.

    If node is not an allowed child of any of its ancestors in the same editing host:

    1. If node is a dd or dt, wrap the one-node list consisting of node, with sibling criteria returning true for any dl with no attributes and false otherwise, and new parent instructions returning the result of calling createElement("dl") on the context object. Then abort these steps.
    2. There's no reason to change the node to a paragraph if that won't make it an allowed child anyway.

      If "p" is not an allowed child of the editing host of node, abort these steps.

    3. If node is not a prohibited paragraph child, abort these steps.
    4. Set the tag name of node to the default single-line container name, and let node be the result.
    5. Because maybe it somehow wound up as the child of a p, like via insertHTML.

      Fix disallowed ancestors of node.

    6. Let children be node's children.
    7. For each child in children, if child is a prohibited paragraph child:
      1. Record the values of the one-node list consisting of child, and let values be the result.
      2. Split the parent of the one-node list consisting of child.
      3. Restore the values from values.
    8. Abort these steps.
  3. Record the values of the one-node list consisting of node, and let values be the result.
  4. While node is not an allowed child of its parent, split the parent of the one-node list consisting of node.
  5. Restore the values from values.

This algorithm implies that we don't support a sublist in the middle of an item, only at the end. For instance,

<li>foo<ol>...</ol>bar</li>

gets transformed to

<li>foo</li><ol>...</ol><li>bar</li>

which in particular creates an extra list marker for "bar". This is okay; we don't need to expose all of HTML's markup abilities through execCommand(). Similarly, the superscript and subscript commands don't allow nesting. I didn't see any way to get a sublist in the middle of an item in Word 2007 or in OpenOffice.org 3.2.1 Ubuntu package, nor in any browser using just execCommand(), so it should be no big problem if we require that such nesting not occur. (Existing browsers behave weirdly and inconsistently when confronted with this kind of nesting.)

The reason we need this is that otherwise it gets very confusing to figure out what happens in cases like trying to outdent

<ol><li>[foo<ol><li>bar]</ol>baz</ol>

If we first normalize, then the natural answer is something like

<p>[foo<ol><li>bar]<li>baz</ol>

but if we don't, we'd have to special-case in the toggle lists and outdent algorithms. This might be worthwhile, but it's not at all clear, and what I have works okay, so I'll stick with it for now.

TODO: Investigate fixing this.

To normalize sublists in a node item:

  1. If item is not an li or it is not editable or its parent is not editable, abort these steps.
  2. Let new item be null.
  3. While item has an ol or ul child:
    1. Let child be the last child of item.
    2. If child is an ol or ul, or new item is null and child is a Text node whose data consists of zero of more space characters:
      1. Set new item to null.
      2. Insert child into the parent of item immediately following item, preserving ranges.
    3. Otherwise:
      1. If new item is null, let new item be the result of calling createElement("li") on the ownerDocument of item, then insert new item into the parent of item immediately after item.
      2. Insert child into new item as its first child, preserving ranges.

The selection's list state is returned by the following algorithm:

This is just a helper to tell the state and indeterminacy of the insertOrderedList command and the insertUnorderedList command:

ol indeterm ol state ul indeterm ul state
ol false true false false
ul false false false true
mixed true false true false
mixed ol true false false false
mixed ul false false true false
none false false false false
  1. If the active range is null, return "none".
  2. Block-extend the active range, and let new range be the result.
  3. Let node list be a list of nodes, initially empty.
  4. For each node contained in new range, append node to node list if the last member of node list (if any) is not an ancestor of node; node is editable; node is not an indentation element; and node is either an ol or ul, or the child of an ol or ul, or an allowed child of "li".
  5. If node list is empty, return "none".
  6. The child-of-child case is necessary right now because of the following:

    <ol><li>[foo<ol><li>bar]</ol>baz</ol>

    With the current (July 2011) block-extend algorithm, this will become:

    {<ol><li>foo<ol><li>bar</ol>}baz</ol>

    because of the magical li handling in block-extend. We want this to register as ol, because after normalizing sublists it will become

    {<ol><li>foo</li><ol><li>bar</ol>}<li>baz</ol>

    But the text node "foo" will wind up in node list, and is not the child of an ol. This is all very messy and has to do with questionable decisions about how to handle nested lists.

    If every member of node list is either an ol or the child of an ol or the child of an li child of an ol, and none is a ul or an ancestor of a ul, return "ol".

  7. This condition and the last are mutually exclusive, so the order is actually irrelevant. Clearly they could only both hold if no member of node list is an ol or ul, so if they both held, every member would have to be either the child of an ol and of a ul, or of an ol and an li, or a ul and an li, or of an li that's the child of both an ol and a ul. This is impossible unless the list is empty, in which case we already aborted.

    If every member of node list is either a ul or the child of a ul or the child of an li child of a ul, and none is an ol or an ancestor of an ol, return "ul".

  8. If some member of node list is either an ol or the child or ancestor of an ol or the child of an li child of an ol, and some member of node list is either a ul or the child or ancestor of a ul or the child of an li child of a ul, return "mixed".
  9. If some member of node list is either an ol or the child or ancestor of an ol or the child of an li child of an ol, return "mixed ol".
  10. If some member of node list is either a ul or the child or ancestor of a ul or the child of an li child of a ul, return "mixed ul".
  11. Return "none".

When querying the value of justify*, IE9 seems to return boolean false across the board when it doesn't throw exceptions, which it usually does in my tests. Chrome 14 dev returns the string "true" or "false" depending on state, as in other cases, which is useless. Opera 11.11 returns "" across the board. Firefox 6.0a2 behaves like with other command values: it returns "center"/"justify"/"left"/"right" depending on the active range's start node. Since this is the only behavior that's possibly useful, it's what I specced. Firefox ties the value closely to the state, returning true for the state if and only if the value matches the desired value, but this seems less useful than what I've specced for the state.

This API is based on the four-state text-align of CSS 2.1. We do some crude mapping to make it not break too badly with CSS3 values, but it's not going to work well given the design of the API.

The alignment value of a node node is returned by the following algorithm:

This is basically like the resolved value of text-align, but with two key differences. First, it only ever evaluates to center/justify/left/right, since that's the model that the justify commands work with. Second, it ignores inline elements, because text-align has no effect on them and their alignment is actually governed by their nearest block ancestor (if any).

  1. While node is neither null nor an Element, or it is an Element but its "display" property has resolved value "inline" or "none", set node to its parent.
  2. This means there's no applicable style rule, so probably it will wind up left-aligned. Of course this ignores the fact that the alignment will really be "start", so this is wrong for RTL, but it's a pretty marginal corner case anyway. (It will only happen if, e.g., everything up to and including the html and body elements have display: inline or none.)

    If node is not an Element, return "left".

  3. If node's "text-align" property has resolved value "start", return "left" if the directionality of node is "ltr", "right" if it is "rtl".
  4. If node's "text-align" property has resolved value "end", return "right" if the directionality of node is "ltr", "left" if it is "rtl".
  5. If node's "text-align" property has resolved value "center", "justify", "left", or "right", return that value.
  6. Return "left".

Sometimes one location corresponds to multiple distinct boundary points. For instance, in the DOM <p>Hello</p>, a boundary point might lie at the beginning of the text node or the beginning of the element node, but these don't logically differ much and will appear the same to the user, so we often want to treat them the same. The algorithms here allow navigating through such equivalent boundary points, for when we want to make the selection as inclusive or exclusive as possible. For deletion, we want to delete as few nodes as possible, so we move the start node forward and the end node backward. In other cases we might do the reverse, expanding the selection. In still other cases we might want to move forward or backward to try getting to a text node.

Given a boundary point (node, offset), the next equivalent point is either a boundary point or null, as returned by the following algorithm:

  1. If node's length is zero, return null.
  2. We don't want to move into or out of zero-length nodes, because that would move us straight through them. For instance, if {}<span></span> were equivalent to <span>{}</span>, it would also be equivalent to <span></span>{}. This produces very unexpected results for nodes like <br>.
  3. If offset is node's length, and node's parent is not null, and node is an inline node, return (node's parent, 1 + node's index).
  4. For instance, <span>foo[]</span> is equivalent to <span>foo{}</span>, which is equivalent to <span>foo</span>{}. However, <p>foo{}</p> is not equivalent to <p>foo</p>{} – the cursor might look like it's in a visibly different position.
  5. If node has a child with index offset, and that child's length is not zero, and that child is an inline node, return (that child, 0).
  6. For instance, {}<span>foo</span> is equivalent to <span>{}foo</span>, which is equivalent to <span>[]foo</span>. As noted before, though, we don't descend into empty nodes. And again, {}<p>foo</p> is different from <p>{}foo</p>.
  7. Return null.

Given a boundary point (node, offset), the previous equivalent point is either a boundary point or null, as returned by the following algorithm:

  1. If node's length is zero, return null.
  2. If offset is 0, and node's parent is not null, and node is an inline node, return (node's parent, node's index).
  3. If node has a child with index offset − 1, and that child's length is not zero, and that child is an inline node, return (that child, that child's length).
  4. Return null.

The first equivalent point of a boundary point (node, offset) is returned by the following algorithm:

  1. While (node, offset)'s previous equivalent point is not null, set (node, offset) to its previous equivalent point.
  2. Return (node, offset).

The last equivalent point of a boundary point (node, offset) is returned by the following algorithm:

  1. While (node, offset)'s next equivalent point is not null, set (node, offset) to its next equivalent point.
  2. Return (node, offset).

Block-extending a range

A boundary point (node, offset) is a block start point if either node's parent is null and offset is zero; or node has a child with index offset − 1, and that child is either a visible block node or a visible br.

A boundary point (node, offset) is a block end point if either node's parent is null and offset is node's length; or node has a child with index offset, and that child is a visible block node.

A boundary point is a block boundary point if it is either a block start point or a block end point.

When a user agent is to block-extend a range range, it must run the following steps:

Generally, block commands work on any block that contains part of the selection, even if the selection doesn't include the whole block. This algorithm takes an input range, copies it, stretches out the copy to contain entire blocks, and returns the result. Then the caller will normally use it instead of the range it started with. For instance, if the cursor is collapsed in a text node inside a paragraph, this will generally return a range that includes the whole paragraph.

Two bits of magic worth noting. First, <br> counts as a block delimiter here, since it looks the same as a block boundary (assuming no margin etc.) and this is a visual API. We include the <br> as part of the line that precedes it. Second, if the selection is inside an <li>, this will extend it to include the whole <li>. This latter point is weird, and I should re-examine it sometime, but it seems to work.

  1. Let start node, start offset, end node, and end offset be the start and end nodes and offsets of range.
  2. If some inclusive ancestor of start node is an li, set start offset to the index of the last such li in tree order, and set start node to that li's parent.
  3. If (start node, start offset) is not a block start point, repeat the following steps:
    1. If start offset is zero, set it to start node's index, then set start node to its parent.
    2. Otherwise, subtract one from start offset.
    3. If (start node, start offset) is a block boundary point, break from this loop.
  4. This just changes something like <div>{<p>foo]</p></div> to {<div><p>foo]</p></div>.

    While start offset is zero and start node's parent is not null, set start offset to start node's index, then set start node to its parent.

  5. If some inclusive ancestor of end node is an li, set end offset to one plus the index of the last such li in tree order, and set end node to that li's parent.
  6. If (end node, end offset) is not a block end point, repeat the following steps:
    1. If end offset is end node's length, set it to one plus end node's index, then set end node to its parent.
    2. Otherwise, add one to end offset.
    3. If (end node, end offset) is a block boundary point, break from this loop.
  7. While end offset is end node's length and end node's parent is not null, set end offset to one plus end node's index, then set end node to its parent.
  8. Let new range be a new range whose start and end nodes and offsets are start node, start offset, end node, and end offset.
  9. Return new range.

A node node follows a line break if the following algorithm returns true:

  1. Let offset be zero.
  2. While (node, offset) is not a block boundary point:
    1. If node has a visible child with index offset minus one, return false.
    2. If offset is zero or node has no children, set offset to node's index, then set node to its parent.
    3. Otherwise, set node to its child with index offset minus one, then set offset to node's length.
  3. Return true.

A node node precedes a line break if the following algorithm returns true:

  1. Let offset be node's length.
  2. While (node, offset) is not a block boundary point:
    1. If node has a visible child with index offset, return false.
    2. If offset is node's length or node has no children, set offset to one plus node's index, then set node to its parent.
    3. Otherwise, set node to its child with index offset and set offset to zero.
  3. Return true.

Recording and restoring overrides

To record current overrides:

  1. Let overrides be a list of (string, string or boolean) ordered pairs, initially empty.
  2. When restoring, some commands can interfere with others. Specifically, we want to restore createLink before foreColor and underline, and subscript and superscript before fontSize. TODO: This approach only works for default styles (although I'm not sure offhand how we could handle non-default styles in principle).

    Firefox 7.0a2 and Opera 11.50 don't honor createLink with collapsed selections. If you insert text, it's not linked. The spec follows Chrome 14 dev. IE9 also ignores createLink with collapsed selections, but its behavior in other cases for collapsed selections is totally different from all other browsers, so it's not a fair comparison.

    If there is a value override for "createLink", add ("createLink", value override for "createLink") to overrides.

  3. Firefox 7.0a2 and Opera 11.50 will honor repeated subscript/superscript commands on a collapsed selection, allowing you to nest them. The spec follows the general philosophy that we don't allow users to nest subscript/superscript, so the last one wins. Chrome 14 dev is similar to the spec.

    For each command in the list "bold", "italic", "strikethrough", "subscript", "superscript", "underline", in order: if there is a state override for command, add (command, command's state override) to overrides.

  4. For each command in the list "fontName", "fontSize", "foreColor", "hiliteColor", in order: if there is a value override for command, add (command, command's value override) to overrides.
  5. Return overrides.

To record current states and values:

  1. Let overrides be a list of (string, string or boolean) ordered pairs, initially empty.
  2. Let node be the first formattable node effectively contained in the active range, or null if there is none.
  3. If node is null, return overrides.
  4. Add ("createLink", node's effective command value for "createLink") to overrides.
  5. Thus we will set state overrides based on the first formattable node, to match values. This means that if you have <p>foo<b>[bar</b>baz]</p> and hit backspace and hit A, you'll get <p>foo<b>a[]</b></p>, although bold was previously indeterminate. This is needed to match the behavior of hitting A straight away, since innerText doesn't strip wrappers when it invokes "delete the contents".

    For each command in the list "bold", "italic", "strikethrough", "subscript", "superscript", "underline", in order: if node's effective command value for command is one of its inline command activated values, add (command, true) to overrides, and otherwise add (command, false) to overrides.

  6. For each command in the list "fontName", "foreColor", "hiliteColor", in order: add (command, command's value) to overrides.
  7. Special case for fontSize, because its values are weird.

    Add ("fontSize", node's effective command value for "fontSize") to overrides.

    This is wrong: it will convert non-pixel sizes to pixel sizes. But I don't see any way to avoid it. Hopefully it won't come up too often. font-size is a real problem, because the mapping from specified value to computed value is lossy and not fully defined (e.g., how many px is "small"?).

  8. Return overrides.

To restore states and values specified by a list overrides returned by the record current overrides or record current states and values algorithm:

  1. Let node be the first formattable node effectively contained in the active range, or null if there is none.
  2. If node is not null, then for each (command, override) pair in overrides, in order:
    1. If override is a boolean, and queryCommandState(command) returns something different from override, take the action for command, with value equal to the empty string.
    2. Otherwise, if override is a string, and command is neither "createLink" nor "fontSize", and queryCommandValue(command) returns something not equivalent to override, take the action for command, with value equal to override.
    3. This special case is needed because createLink has no value.

      Otherwise, if override is a string; and command is "createLink"; and either there is a value override for "createLink" that is not equal to override, or there is no value override for "createLink" and node's effective command value for "createLink" is not equal to override: take the action for "createLink", with value equal to override.

    4. The override will be some CSS value, so we have to convert it to a legacy font size.

      Otherwise, if override is a string; and command is "fontSize"; and either there is a value override for "fontSize" that is not equal to override, or there is no value override for "fontSize" and node's effective command value for "fontSize" is not loosely equivalent to override:

      1. Convert override to an integer number of pixels, and set override to the legacy font size for the result.
      2. Take the action for "fontSize", with value equal to override.
    5. Otherwise, continue this loop from the beginning.
    6. If we took the action for a command, we need to reset node, because it might have changed. For instance, if the selection was foo[bar]baz, the text node could have been split so that the first part is now outside the active range.

      Set node to the first formattable node effectively contained in the active range, if there is one.

  3. Otherwise, for each (command, override) pair in overrides, in order:
    1. If override is a boolean, set the state override for command to override.
    2. If override is a string, set the value override for command to override.

Deleting the selection

TODO: Consider what should happen for block merging in corner cases like display: inline-table.

To delete the selection, given a block merging flag that defaults to true, a strip wrappers flag that defaults to true, and a string direction that defaults to "forward":

The idea behind this algorithm is self-explanatory, but the details wind up being remarkably complicated.

First, any editable nodes inside the selection will be deleted, and the selection will be collapsed. By way of contrast, effectively contained tries to expand the range to include as much as possible, so <p>[foo]</p> contains the <p>. What we do here is contract the range to include as little as possible, so {<p>foo</p>} contains only foo and doesn't delete the paragraph.

After that, if the selection originally started and ended in different blocks, and the block merging flag is true, the end block will get merged into the start block. This is needed so if the user selects text on several lines and deletes it, the text immediately that was before the selection winds up on the same line as the text immediately after it. For example, <p>fo[o</p><div>b]ar</div> becomes <p>fo[]ar</p>. This procedure winds up being tricky, and takes up a large chunk of the logic.

Tables are a notable special case. If an entire table is contained in the range, it will be deleted. If it's anything less, only the contents of the cells will be deleted and the table structure will be left intact.

The strip wrappers flag controls what happens if the deletion removes all the contents of an inline element. If wrappers are being stripped, the empty inline element will be removed: this is usually what you want, because the user can't position the selection inside it. But callers like the insertText command that intend to immediately insert new contents want to leave the wrappers, so the new contents are wrapped by the same thing as the old.

Even if strip wrappers is true, the algorithm will set a state override and value override for any styles it winds up removing. This way, if the user deletes a wrapper that adds a style (or link for that matter), then types something, the new text will get the style from the old text.

  1. If the active range is null, abort these steps and do nothing.
  2. Canonicalize whitespace at the active range's start.
  3. Canonicalize whitespace at the active range's end.
  4. Let (start node, start offset) be the last equivalent point for the active range's start.
  5. Let (end node, end offset) be the first equivalent point for the active range's end.
  6. If (end node, end offset) is not after (start node, start offset):

    This is a selection like <span>foo[</span></span>]bar</span>, where the boundary points are equivalent but not identical. We just collapse it and abort, since there's nothing to delete.

    1. If direction is "forward", call collapseToStart() on the context object's selection.
    2. Otherwise, call collapseToEnd() on the context object's selection.
    3. Abort these steps.
  7. If start node is a Text node and start offset is 0, set start offset to the index of start node, then set start node to its parent.
  8. If end node is a Text node and end offset is its length, set end offset to one plus the index of end node, then set end node to its parent.

    The previous two steps are so that we won't leave empty text nodes anywhere.

  9. Call collapse(start node, start offset) on the context object's selection.
  10. Call extend(end node, end offset) on the context object's selection.
  11. When we delete a selection that spans multiple blocks, we merge the end block's contents into the start block, like

    <p>fo[o</p><pre>b]ar</pre>
    -> <p>fo[]ar</p>.

    We figure out what the start and end blocks are before we start deleting anything.

  12. Let start block be the active range's start node.
  13. While start block's parent is in the same editing host and start block is an inline node, set start block to its parent.
  14. We only merge to or from block nodes or editing hosts. (This is just in case someone makes a span into an editing host and sticks paragraphs inside it or something . . . we could probably drop that proviso.) Firefox 7.0a2 ignores the display property when merging, so it doesn't merge <span style=display:block> but does merge <p style=display:inline>. This is undesirable, because it's visually wrong. IE10PP2 and Chrome 14 dev behave more like the spec, and Opera 11.50 seems to be unable to make up its mind.

    If span isn't an allowed child, it's probably something unpleasant like a table row or a list or such. We don't want to merge to or from something like that, because we'd most likely wind up with the wrong type of child somewhere. It should be pretty hard for this to happen given the normalization we do on the selection; I'm not actually sure how it could happen at all, actually, unless you start out with a DOM that has non-allowed children someplace. So it's basically a sanity check.

    We don't let either start block or end block be a td or th. This means we'll never merge to or from a td or th. This matches Firefox 5.0a2, and reportedly Word as well. Chrome 13 dev and Opera 11.11 allow merging from a non-table cell end block to a table cell start block, but not vice versa. In IE9 the delete key just does nothing.

    If start block is neither a block node nor an editing host, or "span" is not an allowed child of start block, or start block is a td or th, set start block to null.

  15. Let end block be the active range's end node.
  16. While end block's parent is in the same editing host and end block is an inline node, set end block to its parent.
  17. If end block is neither a block node nor an editing host, or "span" is not an allowed child of end block, or end block is a td or th, set end block to null.
  18. Later on we'll restore overrides. This ensures that if we delete inline formatting elements and the user then types something, the typed text will have the same style as before.

  19. As far as I can tell, IE9 and Opera 11.50 don't do this at all. If you delete a selection and then start typing, the new text doesn't take on the styles of the old text.

    Firefox 7.0a2 seems to do it for some styles but not others. Strikethrough, superscript, subscript, and links seem to be lost, at a minimum.

    The spec goes with something like Chrome 14 dev, which tries to preserve lots of stuff.

    Record current states and values, and let overrides be the result.

  20. Now we actually begin deleting things.

  21. This whole piece of the algorithm is based on deleteContents() in DOM Range, copy-pasted and then adjusted to fit.

    If start node and end node are the same, and start node is an editable Text node:

    1. Call deleteData(start offset, end offsetstart offset) on start node.
    2. Canonicalize whitespace at (start node, start offset), with fix collapsed space false.
    3. If direction is "forward", call collapseToStart() on the context object's selection.
    4. Otherwise, call collapseToEnd() on the context object's selection.
    5. This is needed to restore any overrides that would otherwise be lost. TODO: In this and similar cases, we could optimize by saving only overrides, not the full state/value.

      Restore states and values from overrides.

    6. Abort these steps.
  22. If start node is an editable Text node, call deleteData() on it, with start offset as the first argument and (length of start nodestart offset) as the second argument.
  23. Let node list be a list of nodes, initially empty.
  24. IE9 doesn't seem to let you do any intercell deletions: the delete key does nothing if you select across multiple cells. Firefox 5.0a2 and Opera 11.11 behave as the spec says, not removing any table things. Chrome 13 dev will remove entire rows if selected. Note that IE, Firefox, Word 2007, and OpenOffice.org 3.2.1 Ubuntu all switch to a magic cell-selection mode when you try to select between cells, at least in some cases, instead of selecting letter-by-letter.

    For each node contained in the active range, append node to node list if the last member of node list (if any) is not an ancestor of node; node is editable; and node is not a thead, tbody, tfoot, tr, th, or td.

  25. For each node in node list:
    1. Let parent be the parent of node.
    2. Remove node from parent.
    3. Do this before stripping wrappers: see bug 13831.

      If the block node of parent has no visible children, and parent is editable or an editing host, call createElement("br") on the context object and append the result as the last child of parent.

    4. Taking insertText to test the case where strip wrappers is false, with value a: <p>[foo<b>bar</b>]baz becomes <p>a[]baz per spec, in IE9, and in Chrome 14 dev. Firefox 7.0a2 and Opera 11.50 make it <p>a[]<b></b>baz, with a useless wrapper. <p>foo<b>[bar</b>baz] becomes <p>foo<b>a[]</b> per spec and in IE9 and Firefox 7.0a2 and Opera 11.50; in Chrome 14 dev apparently it initially becomes <p>fooa[], but then the style is recreated. This is detectable if you do something weird like <span style=color:#aBcDeF> instead of <b>: it comes <font class=Apple-style-span color=#abcdef> or such. I follow IE9 in all cases, because it makes the most sense.

      If strip wrappers is true or parent is not an inclusive ancestor of start node, while parent is an editable inline node with length 0, let grandparent be the parent of parent, then remove parent from grandparent, then set parent to grandparent.

      Even if strip wrappers is false, we still want to strip wrappers that aren't inclusive ancestors of start node. The idea of not stripping wrappers is that we're going to insert new content right afterward, like text or an image, but that new content will be inserted at the start node. Wrappers in other places still need to be removed, because they would otherwise remain empty.

  26. If end node is an editable Text node, call deleteData(0, end offset) on it.
  27. Canonicalize whitespace at the active range's start, with fix collapsed space false.
  28. Canonicalize whitespace at the active range's end, with fix collapsed space false.
  29. Now we need to merge blocks. The simplest case is something like

    <p>fo[o</p><p>bar</p><p>b]az</p>
    -> <p>fo</p>{}<p>az</p>
    -> <p>fo{}az</p>

    where neither block descends from the other. More complicated is something like

    foo[<p>]bar</p>
    -> foo[]bar

    or

    <p>foo[</p>]bar
    -> <p>foo[]bar</p>

    where one descends from the other.

  30. If block merging is false, or start block or end block is null, or start block is not in the same editing host as end block, or start block and end block are the same:
    1. If direction is "forward", call collapseToStart() on the context object's selection.
    2. Otherwise, call collapseToEnd() on the context object's selection.
    3. Restore states and values from overrides.
    4. Abort these steps.
  31. We might have added a br to the start/end block in an earlier step. Now we're about to merge the blocks, and we don't want the br's to get in the way. The end block is being destroyed no matter what. If the start block winds up empty after merging, we'll add a new br child at the end so it doesn't collapse.

    If start block has one child, which is a collapsed block prop, remove its child from it.

  32. Just repeatedly blow up the end block in this case.

    If start block is an ancestor of end block:

    1. Let reference node be end block.
    2. While reference node is not a child of start block, set reference node to its parent.
    3. Call collapse() on the context object's selection, with first argument start block and second argument the index of reference node.
    4. If end block has no children:
      1. While end block is editable and is the only child of its parent and is not a child of start block, let parent equal end block, then remove end block from parent, then set end block to parent.
      2. If end block is editable and is not an inline node, and its previousSibling and nextSibling are both inline nodes, call createElement("br") on the context object and insert it into end block's parent immediately after end block.
      3. If end block is editable, remove it from its parent.
      4. Restore states and values from overrides.
      5. Abort these steps.
    5. If end block's firstChild is not an inline node, restore states and values from record, then abort these steps.
    6. Let children be a list of nodes, initially empty.
    7. Append the first child of end block to children.
    8. While children's last member is not a br, and children's last member's nextSibling is an inline node, append children's last member's nextSibling to children.
    9. Record the values of children, and let values be the result.
    10. While children's first member's parent is not start block, split the parent of children.
    11. If children's first member's previousSibling is an editable br, remove that br from its parent.
  33. In this case, pull in everything that comes after start block, until we hit a br or block node.

    Otherwise, if start block is a descendant of end block:

    1. Call collapse() on the context object's selection, with first argument start block and second argument start block's length.
    2. Let reference node be start block.
    3. While reference node is not a child of end block, set reference node to its parent.
    4. If reference node's nextSibling is an inline node and start block's lastChild is a br, remove start block's lastChild from it.
    5. Let nodes to move be a list of nodes, initially empty.
    6. If reference node's nextSibling is neither null nor a block node, append it to nodes to move.
    7. While nodes to move is nonempty and its last member isn't a br and its last member's nextSibling is neither null nor a block node, append its last member's nextSibling to nodes to move.
    8. Record the values of nodes to move, and let values be the result.
    9. For each node in nodes to move, append node as the last child of start block, preserving ranges.
  34. In the last case, just move all the children of the end block to the start block, and then get rid of any elements we emptied that way.

    Otherwise:

    1. Call collapse() on the context object's selection, with first argument start block and second argument start block's length.
    2. If end block's firstChild is an inline node and start block's lastChild is a br, remove start block's lastChild from it.
    3. Record the values of end block's children, and let values be the result.
    4. While end block has children, append the first child of end block to start block, preserving ranges.
    5. While end block has no children, let parent be the parent of end block, then remove end block from parent, then set end block to parent.
  35. We might have deleted the contents between two lists, in which case we should merge them. See bug 13976.

  36. Let ancestor be start block.
  37. While ancestor has an inclusive ancestor ol in the same editing host whose nextSibling is also an ol in the same editing host, or an inclusive ancestor ul in the same editing host whose nextSibling is also a ul in the same editing host:
    1. While ancestor and its nextSibling are not both ols in the same editing host, and are also not both uls in the same editing host, set ancestor to its parent.
    2. While ancestor's nextSibling has children, append ancestor's nextSibling's firstChild as the last child of ancestor, preserving ranges.
    3. Remove ancestor's nextSibling from its parent.
  38. Restore the values from values.
  39. If start block has no children, call createElement("br") on the context object and append the result as the last child of start block.
  40. Remove extraneous line breaks at the end of start block.
  41. Restore states and values from overrides.

Splitting a node list's parent

To split the parent of a list node list of consecutive sibling nodes:

This algorithm breaks up the parent of node list. If they're the only children of their parent, the parent is removed entirely. If there are preceding or following siblings, the original parent is left intact as the parent of those siblings. If there are both preceding and following siblings, the original parent is left as the parent of the following siblings and a clone is used for the parent of the preceding siblings.

We make sure not to disrupt the appearance any more than necessary. Obviously margins or such on the parent will be lost, but the children will not wind up on the same line as anything they weren't already on the same line as. E.g., if we split the parent of "bar" in foo<p>bar</p>, we get foo<br>bar, not foobar. (This is amazingly complicated and error-prone.) We don't preserve inline styles: callers that want to do that should call record the values and restore the values themselves.

All this is useful in a lot of situations, like for outdenting. For inline formatting commands, we almost always rely on pushing down values instead, since that often leads to tidier markup.

  1. Let original parent be the parent of the first member of node list.
  2. If original parent is not editable or its parent is null, do nothing and abort these steps.
  3. If the first child of original parent is in node list, remove extraneous line breaks before original parent.
  4. If the first child of original parent is in node list, and original parent follows a line break, set follows line break to true. Otherwise, set follows line break to false.
  5. If the last child of original parent is in node list, and original parent precedes a line break, set precedes line break to true. Otherwise, set precedes line break to false.
  6. TODO: We insert things after the parent. This is bad, because it will cause them to become part of any ranges that immediately follow. For instance, if we're hitting "bar" in

    <div><p>foo<p>bar</div>{<p>baz}

    it becomes

    <div><p>foo</div>{<p>bar<p>baz}

    instead of

    <div><p>foo</div><p>bar{<p>baz}

    because of how range mutation rules work. This doesn't happen if we insert before. This may or may not be important enough to bother working around.

    If the first child of original parent is not in node list, but its last child is:

    1. For each node in node list, in reverse order, insert node into the parent of original parent immediately after original parent, preserving ranges.
    2. If precedes line break is true, and the last member of node list does not precede a line break, call createElement("br") on the context object and insert the result immediately after the last member of node list.
    3. Remove extraneous line breaks at the end of original parent.
    4. Abort these steps.
  7. If the first child of original parent is not in node list:
    1. Let cloned parent be the result of calling cloneNode(false) on original parent.
    2. If original parent has an id attribute, unset it.
    3. Insert cloned parent into the parent of original parent immediately before original parent.
    4. While the previousSibling of the first member of node list is not null, append the first child of original parent as the last child of cloned parent, preserving ranges.
  8. Notice that a boundary point that was immediately before the element will now be immediately before its children, just because of the regular range mutation rules, without needing to worry about preserving ranges. Likewise for boundary points immediately after the element, if we wind up removing the element in the final step. Preserving ranges is only necessary for the sake of boundary points in the element or its descendants.

    For each node in node list, insert node into the parent of original parent immediately before original parent, preserving ranges.

  9. If follows line break is true, and the first member of node list does not follow a line break, call createElement("br") on the context object and insert the result immediately before the first member of node list.
  10. If the last member of node list is an inline node other than a br, and the first child of original parent is a br, and original parent is not an inline node, remove the first child of original parent from original parent.
  11. If original parent has no children:
    1. Remove original parent from its parent.
    2. If precedes line break is true, and the last member of node list does not precede a line break, call createElement("br") on the context object and insert the result immediately after the last member of node list.
  12. Otherwise, remove extraneous line breaks before original parent.
  13. The parent might be null if it's a br that we removed in the last step, in which case this step isn't necessary.

    If node list's last member's nextSibling is null, but its parent is not null, remove extraneous line breaks at the end of node list's last member's parent.

To remove a node node while preserving its descendants, split the parent of node's children if it has any. If it has no children, instead remove it from its parent.

Canonical space sequences

Whitespace in HTML normally collapses. However, if the user hits the space bar twice in an HTML editor, they expect to see two spaces, not one. Even if they hit the space bar once at the beginning or end of a line, it would collapse without special handling. The only good solution here is for the author to set white-space: pre-wrap on the editable area, and on everywhere the content is reproduced. But if they don't, we have to painfully hack around the problem.

This is a basically intractable problem because of the unfortunate confluence of three factors. One, our characters are Unicode, and Unicode doesn't know about whitespace collapsing, so it provides no special characters to control it. Two, HTML itself provides no features that control whitespace collapsing without undesired side effects (like inhibiting line breaks or not being allowed inside <p>). Three, we need to support user agents that don't reliably support CSS, since that includes many popular mail clients.

The upshot is we have no good way to control whitespace collapse, so we rely on the least bad way available: &nbsp;. This doesn't collapse with adjacent whitespace in browsers, which is good. But it also doesn't allow a line break opportunity, which is bad. In any run of whitespace that we don't want to collapse, any two regular spaces must be separated by an &nbsp; so they don't collapse together, but we need to carefully limit runs of consecutive &nbsp;s to minimize the damage to line-breaking behavior.

The result is an elaborate and meticulously-crafted hodgepodge of bad compromises that frankly isn't worth the effort to explain here. The saving grace is that it all gets disabled if white-space is set to pre-wrap as it should be, so authors can opt out of the insanity. Interested readers will find detailed rationale for the exact sequences required in the comments.

See long comment before insertText.

The canonical space sequence of length n, with boolean flags non-breaking start and non-breaking end, is returned by the following algorithm:

  1. If n is zero, return the empty string.
  2. If n is one and both non-breaking start and non-breaking end are false, return a single space (U+0020).
  3. If n is one, return a single non-breaking space (U+00A0).
  4. Let buffer be the empty string.
  5. If non-breaking start is true, let repeated pair be U+00A0 U+0020. Otherwise, let it be U+0020 U+00A0.
  6. While n is greater than three, append repeated pair to buffer and subtract two from n.
  7. If n is three, append a three-code unit string to buffer depending on non-breaking start and non-breaking end:
    non-breaking start and non-breaking end false
    U+0020 U+00A0 U+0020
    non-breaking start true, non-breaking end false
    U+00A0 U+00A0 U+0020
    non-breaking start false, non-breaking end true
    U+0020 U+00A0 U+00A0
    non-breaking start and non-breaking end both true
    U+00A0 U+0020 U+00A0
  8. Otherwise, append a two-code unit string to buffer depending on non-breaking start and non-breaking end:
    non-breaking start and non-breaking end false
    non-breaking start true, non-breaking end false
    U+00A0 U+0020
    non-breaking start false, non-breaking end true
    U+0020 U+00A0
    non-breaking start and non-breaking end both true
    U+00A0 U+00A0
  9. Return buffer.

To canonicalize whitespace at (node, offset), given an optional boolean argument fix collapsed space that defaults to true:

  1. If node is neither editable nor an editing host, abort these steps.
  2. Let start node equal node and let start offset equal offset.
  3. First we go to the beginning of the current whitespace run.

    Repeat the following steps:

    1. If start node has a child in the same editing host with index start offset minus one, set start node to that child, then set start offset to start node's length.
    2. TODO: Following a line break is unlikely to be the right criterion.

      Otherwise, if start offset is zero and start node does not follow a line break and start node's parent is in the same editing host, set start offset to start node's index, then set start node to its parent.

    3. Otherwise, if start node is a Text node and its parent's resolved value for "white-space" is neither "pre" nor "pre-wrap" and start offset is not zero and the (start offset − 1)st code unit of start node's data is a space (0x0020) or non-breaking space (0x00A0), subtract one from start offset.
    4. Otherwise, break from this loop.
  4. Now we collapse any consecutive spaces, if fix collapsed space is true.

    Let end node equal start node and end offset equal start offset.

  5. Let length equal zero.
  6. This tries to delete spaces at the beginning of a line (bug 14119).

    Let collapse spaces be true if start offset is zero and start node follows a line break, otherwise false.

  7. Repeat the following steps:
    1. If end node has a child in the same editing host with index end offset, set end node to that child, then set end offset to zero.
    2. TODO: Preceding a line break is unlikely to be the right criterion.

      Otherwise, if end offset is end node's length and end node does not precede a line break and end node's parent is in the same editing host, set end offset to one plus end node's index, then set end node to its parent.

    3. Otherwise, if end node is a Text node and its parent's resolved value for "white-space" is neither "pre" nor "pre-wrap" and end offset is not end node's length and the end offsetth code unit of end node's data is a space (0x0020) or non-breaking space (0x00A0):
      1. If fix collapsed space is true, and collapse spaces is true, and the end offsetth code unit of end node's data is a space (0x0020): call deleteData(end offset, 1) on end node, then continue this loop from the beginning.
      2. Set collapse spaces to true if the end offsetth code unit of end node's data is a space (0x0020), false otherwise.
      3. Add one to end offset.
      4. Add one to length.
    4. Otherwise, break from this loop.
  8. We've already stripped leading whitespace, and collapsed consecutive spaces. Now we try to strip any collapsed trailing whitespace (bug 14119 again).

    If fix collapsed space is true, then while (start node, start offset) is before (end node, end offset):

    1. If end node has a child in the same editing host with index end offset − 1, set end node to that child, then set end offset to end node's length.
    2. Otherwise, if end offset is zero and end node's parent is in the same editing host, set end offset to end node's index, then set end node to its parent.
    3. Otherwise, if end node is a Text node and its parent's resolved value for "white-space" is neither "pre" nor "pre-wrap" and end offset is end node's length and the last code unit of end node's data is a space (0x0020) and end node precedes a line break:
      1. Subtract one from end offset.
      2. Subtract one from length.
      3. Call deleteData(end offset, 1) on end node.
    4. Otherwise, break from this loop.
  9. Finally we replace with the canonical sequence.

    Let replacement whitespace be the canonical space sequence of length length. non-breaking start is true if start offset is zero and start node follows a line break, and false otherwise. non-breaking end is true if end offset is end node's length and end node precedes a line break, and false otherwise.

  10. While (start node, start offset) is before (end node, end offset):
    1. If start node has a child with index start offset, set start node to that child, then set start offset to zero.
    2. Otherwise, if start node is not a Text node or if start offset is start node's length, set start offset to one plus start node's index, then set start node to its parent.
    3. Otherwise:
      1. Remove the first code unit from replacement whitespace, and let element be that code unit.
      2. If element is not the same as the start offsetth code unit of start node's data:
        1. We need to insert then delete, so that we don't change range boundary points. TODO: switch to using "replace data" now that DOM Core has defined that.

          Call insertData(start offset, element) on start node.

        2. Call deleteData(start offset + 1, 1) on start node.
      3. Add one to start offset.

Indenting and outdenting

There are two basically different types of indent/outdent: lists, and everything else. For lists we'll wrap the item in a nested list to indent, and split its parent to outdent. For everything else we'll wrap in a <blockquote> to indent, and try breaking it out of an ancestor indentation element to outdent.

Indenting winds up being pretty simple: just add an appropriate wrapper. There's not really anything to think about here except which wrapper we want (<ol> or <ul> or <blockquote>), and establishing that is not rocket science.

Outdenting is considerably more complicated. The basic idea we follow is to first find the nearest editable ancestor that's a list or indentation element. If we succeed, and the node we're trying to outdent is the only descendant of the ancestor, of course we can just remove the ancestor and that's that. Otherwise, what we do is remove the ancestor and then indent all its other descendants, much like pushing down values.

But of course, there are complications. We don't always actually want to remove the closest ancestor that's causing indentation. For one thing, we prefer ancestors that we can remove completely, i.e., simple indentation elements. When outdenting <blockquote><blockquote id="abc">foo</blockquote></blockquote>, removing the inner tag would result in <blockquote><div id="abc">foo</div></blockquote>, since we don't want to lose the id. Thus we prefer to remove the outer tag and wind up with <blockquote id="abc">foo</blockquote>.

Also, if the node we're outdenting is itself a list, we prefer to remove an ancestor indentation element rather than the list. Otherwise, if the user selected some text, indented it, then added a list, there would be no way to remove the indentation without removing the list first. This way, the user could remove the list with the appropriate list-toggling command or remove the indentation with the outdent command.

We have to handle entire lists of siblings at once, or else we'd wind up doing something like

<ol>
  {<li>foo</li>
  <ol><li>bar</li></ol>}
</ol>
->
<ol><ol>
  <li>foo</li>
  <li>bar</li>
</ol></ol>
->
<ol><ol><ol>
  <li>foo</li>
  <li>bar</li>
</ol></ol></ol>

since by the time we got to doing the <ol> that originally contained "bar", we won't remember that we aren't supposed to indent "foo" a second time.

To indent a list node list of consecutive sibling nodes:

  1. If node list is empty, do nothing and abort these steps.
  2. Let first node be the first member of node list.
  3. If first node's parent is an ol or ul:
    1. Let tag be the local name of the parent of first node.
    2. This matches IE9, Firefox 4.0, and Chrome 12 dev. If there's a preceding <li>, Opera 11.10 instead adds the new parent to the end of that <li>, so it's not the child of another list, which is invalid. But the other browsers' way of doing things makes things simpler. E.g., if we want to indent an <li> and it has <ol>/<ul> children, we have to distinguish between the case where we want to indent the whole <li> or only the first part. It also allows things like

      <ol><li>
        foo
        <ol><li>bar</li></ol>
        baz
      </li></ol>

      in which case it's unclear what we should do if the user selects "foo" and indents. I've filed a bug on HTML5.

      Wrap node list, with sibling criteria returning true for an HTML element with local name tag and false otherwise, and new parent instructions returning the result of calling createElement(tag) on the ownerDocument of first node.

    3. Abort these steps.
  4. Firefox 4.0 respects the CSS styling flag for indent, but Chrome 12 dev does not. I always produce blockquotes, even if CSS styling is on, for two reasons. One, IE9 handles inline margin attributes badly: when outdenting, it propagates the margin to the parent, which doesn't actually remove it. Two, in CSS mode I'd want to use <div style="margin: 1em 40px"> to match non-CSS mode, but authors are very likely to want to remove the top/bottom margin, which they can't do if it's not a special tag. Authors who really want divs for indentation could always convert the blockquotes to divs themselves. But if people really want it, I could respect CSS styling mode here too.

    The top/bottom margins might be undesirable here, but no more so than for <ol>/<ul>/<p>/etc. Here as there, authors can remove them with CSS if they want.

    blockquote indents on both sides, so we don't have to worry about directionality. In theory it would be better if we indented only on the start side, but that requires care to get right in mixed-direction cases. Even once browsers start to support margin-start and so on, we can't use them because a) we have to work okay in legacy browsers and b) it doesn't help if a descendant block has different direction (so should be indented the other way). So let's not worry about it: most browsers don't, and the ones that do get it wrong. Just indent on both sides.

    Wrap node list, with sibling criteria returning true for a simple indentation element and false otherwise, and new parent instructions returning the result of calling createElement("blockquote") on the ownerDocument of first node. Let new parent be the result.

  5. Fix disallowed ancestors of new parent.

Things that are produced for indentation that we need to consider removing:

For discussion on the list-related stuff, see the comment for insertOrderedList.

Gecko in CSS mode just adds margin properties to random elements that are lying around. We don't attempt to remove those, because 1) the amount and position of the margin can vary (it increases the margin if there's a preexisting one), so it's potentially complicated, and 2) no browser removes such margins on outdent, including Gecko, except for Gecko in CSS mode. TODO: Consider removing it anyway.

To outdent a node node:

  1. If node is not editable, abort these steps.
  2. The easy case is when the whole element is indented. In this case we remove the whole thing indiscriminately. In the case of blockquotes created by IE, this might change the direction of some children, but then their direction was probably changed incorrectly in the first place, so no harm.

    If node is a simple indentation element, remove node, preserving its descendants. Then abort these steps.

  3. This might be a simple indentation element that had style added to it by Firefox in CSS mode, for instance (color, font-family, etc.).

    If node is an indentation element:

    1. Unset the dir attribute of node, if any.
    2. Unset the margin, padding, and border CSS properties of node.
    3. Set the tag name of node to "div".
    4. Abort these steps.
  4. Approximate algorithms when an ancestor is causing the indentation appear to be:

    IE9
    Go to the innermost element causing indentation. If the stuff to be outdented includes all the contents of that element, get rid of it, but if it has any attributes, change it to a <p> with those same attributes. This is an excellent idea in general, but unfortunately it preserves explicitly-specified margins in style attributes, which isn't great. In other cases, it moves the stuff to be outdented outside. Not clear on all the details, seems to be pretty confusing. Also does a bunch of seemingly arbitrary normalization like removing divs and some attributes from some things . . .
    Firefox 4.0
    Go to the innermost element causing indentation. If the stuff to be outdented includes all the contents of that element, get rid of it, even if it has arbitrary attributes. Otherwise, move the stuff to be outdented outside the indenting element. If there are any intervening elements that include stuff not to be outdented, wrap the outdented stuff in copies (which can duplicate id's, etc.).
    Chrome 12 dev
    Go to the outermost element causing indentation (even if the current element is itself causing indentation). Move the text to be outdented outside that outermost element, without regard to any intervening elements. Then recreate the original styles on the moved text, in some fashion. Something like that; it confuses me and doesn't seem to be reasonable.
    Opera 11.00
    Like Firefox, except it goes to the outermost element, not the innermost. Also seems to special-case to avoid duplicate id's, and has a few other quirks.

    Overall, all flawed, so I'll make up my own, patterned after pushing down styles. First we search ancestors for a simple indentation element, which we stand a chance of completely removing. Failing that, we look for an indentation element that's not simple, so we can't completely remove it.

    Let current ancestor be node's parent.

  5. Let ancestor list be a list of nodes, initially empty.
  6. While current ancestor is an editable Element that is neither a simple indentation element nor an ol nor a ul, append current ancestor to ancestor list and then set current ancestor to its parent.
  7. If current ancestor is not an editable simple indentation element:
    1. Let current ancestor be node's parent.
    2. Let ancestor list be the empty list.
    3. While current ancestor is an editable Element that is neither an indentation element nor an ol nor a ul, append current ancestor to ancestor list and then set current ancestor to its parent.
  8. When asked to outdent a list wrapped in a simple indentation element, Chrome 12 dev removes the list instead of the simple indentation element. Opera 11.10 seems to remove both. IE9 and Firefox 4.0 remove the simple indentation element, as does the spec.

    If node is an ol or ul and current ancestor is not an editable indentation element:

    1. Unset the reversed, start, and type attributes of node, if any are set.
    2. Let children be the children of node.
    3. We can't turn it into a div if it's the child of an ol or ul, because that's not allowed: there's no way to group li's (see HTML bug 13128).

      If node has attributes, and its parent is not an ol or ul, set the tag name of node to "div".

    4. Otherwise:
      1. Record the values of node's children, and let values be the result.
      2. Remove node, preserving its descendants.
      3. Restore the values from values.
    5. Fix disallowed ancestors of each member of children.
    6. Abort these steps.
  9. If current ancestor is not an editable indentation element, abort these steps.
  10. If we get to this point, we have an ancestor to split up.

    Append current ancestor to ancestor list.

  11. We can't outdent it yet, because we need its children to remain intact for the loop.

    Let original ancestor be current ancestor.

  12. While ancestor list is not empty:
    1. Let current ancestor be the last member of ancestor list.
    2. Remove the last member from ancestor list.
    3. Let target be the child of current ancestor that is equal to either node or the last member of ancestor list.
    4. If target is an inline node that is not a br, and its nextSibling is a br, remove target's nextSibling from its parent.
    5. Let preceding siblings be the precedings siblings of target, and let following siblings be the followings siblings of target.
    6. Indent preceding siblings.
    7. Indent following siblings.
  13. Outdent original ancestor.

Toggling lists

This is the action for the insertOrderedList command and the insertUnorderedList command, which behave identically except for which list type they target. It does several things that vary contextually.

If everything in the selection is contained in the target list type already, this more or less just outdents everything one step. This is relatively simple.

Otherwise, it's slightly more complicated:

First, any lists of the opposite list type (other tag name) get converted to the target list type (tag name). They get merged into a sibling if appropriate, otherwise we set the tag name.

Then we go through all the affected nodes, handling each run of consecutive siblings separately. Any line that's not already wrapped in an <li> gets wrapped. If the parent at this point isn't a list at all, the run gets wrapped in a list. If it's the wrong type of list, we split the parent and rewrap it in the right type of list. That's basically it, except that we have to exercise the usual care to try merging with siblings and so forth.

Research for insertOrderedList/insertUnorderedList: tested the following command sequences in IE9, Firefox 4.0, Chrome 12 dev, Opera 11.10, OpenOffice.org 3.2.1 Ubuntu package, Microsoft Office Word 2007. The commands "ol", "ul", "indent", "outdent" correspond in browsers to "insertOrderedList", "insertUnorderedList", "indent", and "outdent"; in OO.org to "Numbering On/Off", "Bullets On/Off", "Increase Indent", "Decrease Indent"; and in Word to "Numbering", "Bullets", "Increase Indent", "Decrease Indent".

Note: OO has a bunch of extra options, like "Promote One Level", "Demote One Level", "Promote One Level With Subpoints", "Demote One Level With Subpoints", "Insert Unnumbered Entry", "Restart Numbering". The regular "Increase/Decrease Indent" commands work oddly, and I assume they're not really meant to be used inside lists. Thus I also tested with "Promote One Level" and "Demote One Level". These are denoted by OO' instead of OO.

Assume that there are style rules in effect like

ol ol { list-style-type: lower-alpha }
ol ol ol { list-style-type: lower-roman }

This is the default appearance in Word, and I set OO to something similar with Bullets and Numbering → Outline in the list editing toolbox. I'm ignoring bullet style throughout, for no particular reason.

Ignoring the conceptual model of HTML, which users won't understand, here's the conceptual model I've developed for lists: text is divided up into blocks. Each block has an indentation level and a list marker type. The list marker type can be either nothing, ordered, or unordered. A list block cannot have indentation level less than one. Any given piece of text is part of only one block. A block may be visually non-contiguous, such as if a single list block is interrupted by a further-indented block.

To find the right number (or letter) for an ordered-list block, look at the immediately preceding block, but skip over any blocks of higher indentation level. If there is no immediately preceding block, or it's not an ordered-list block, or it has a lower indentation level, the number is 1 (or a, i, etc.). Otherwise, it's the number of the preceding block plus one.

ol/ul commands change the selected block to that list marker type, or remove the list marker type if it's already the chosen type. If the block has indentation level zero, it increases to one.

indent/outdent commands change the selected block's indentation level. If a list block's indentation level is reduced to zero, it's converted to a regular block.

What this means from an HTML perspective, roughly:

Sheesh, lists are complicated.

To toggle lists, given a string tag name (either "ol" or "ul"):

  1. Let mode be "disable" if the selection's list state is tag name, and "enable" otherwise.
  2. Let other tag name be "ol" if tag name is "ul", and "ul" if tag name is "ol".
  3. Let items be a list of all lis that are inclusive ancestors of the active range's start and/or end node.
  4. TODO: This overnormalizes, but it seems like the simplest solution for now.

    For each item in items, normalize sublists of item.

  5. Block-extend the active range, and let new range be the result.
  6. If mode is "enable", then let lists to convert consist of every editable HTML element with local name other tag name that is contained in new range, and for every list in lists to convert:
    1. Convert it to the right name. If possible, we want to merge with a neighboring list of the correct type. Failing that, we set the tag name.

      If list's previousSibling or nextSibling is an editable HTML element with local name tag name:

      1. Let children be list's children.
      2. Record the values of children, and let values be the result.
      3. Split the parent of children.
      4. Wrap children, with sibling criteria returning true for an HTML element with local name tag name and false otherwise.
      5. Restore the values from values.
    2. Otherwise, set the tag name of list to tag name.
  7. Let node list be a list of nodes, initially empty.
  8. We exclude indentation elements so that selecting some random text and doing indent followed by insertOrderedList will have the same result as the reverse. Specifically,

    <blockquote>[foo]</blockquote> ->
    <blockquote><ol><li>[foo]</li></ol></blockquote>

    per spec and Firefox 4.0 and (more or less) Chrome 12 dev. Opera 11.10 instead does <ol><li>foo</li></ol>, so the indentation vanishes. IE9 does <ol><ol><li>foo</li></ol></ol>, but that doesn't make semantic sense and is different from how it would work if you reversed the commands. OpenOffice.org 3.2.1 (Ubuntu) and Word 2007 both agree with the spec in this case.

    For each node node contained in new range, if node is editable; the last member of node list (if any) is not an ancestor of node; node is not an indentation element; and either node is an ol or ul, or its parent is an ol or ul, or it is an allowed child of "li"; then append node to node list.

  9. We don't want to touch these. E.g., assuming tag name is "ol",

    [foo<ol><li>bar</ol>baz]
    -> <ol><li>[foo<li>bar<li>baz]</ol>
    not <ol><li>[foo</li><ol><li>bar</ol><li>baz]</ol>.

    But

    <ul><li>foo<li>[bar</li><ol><li>baz</ol><li>quz]</ul>
    -> <ul><li>foo</ul><ol><li>[bar</li><ol><li>baz</ol><li>quz]</ol>
    not <ul><li>foo</ul><ol><li>[bar</ol><ul><ol><li>baz</ol></ul><ol><li>quz]</ol>

    If mode is "enable", remove from node list any ol or ul whose parent is not also an ol or ul.

  10. If mode is "disable", then while node list is not empty:
    1. Let sublist be an empty list of nodes.
    2. Remove the first member from node list and append it to sublist.
    3. If the first member of sublist is an HTML element with local name tag name, outdent it and continue this loop from the beginning.
    4. While node list is not empty, and the first member of node list is the nextSibling of the last member of sublist and is not an HTML element with local name tag name, remove the first member from node list and append it to sublist.
    5. Record the values of sublist, and let values be the result.
    6. Split the parent of sublist.
    7. Fix disallowed ancestors of each member of sublist.
    8. Restore the values from values.
  11. Otherwise, while node list is not empty:
    1. Let sublist be an empty list of nodes.
    2. Accumulate consecutive sibling nodes in sublist, first converting them all to li's (except if they're already lists).

      While either sublist is empty, or node list is not empty and its first member is the nextSibling of sublist's last member:

      1. Thus <p>foo</p> becomes <ol><li>foo</ol> instead of <ol><li><p>foo</ol>, and likewise for div, but other things will be put inside the <li>.

        If node list's first member is a p or div, set the tag name of node list's first member to "li", and append the result to sublist. Remove the first member from node list.

      2. Otherwise, if the first member of node list is an li or ol or ul, remove it from node list and append it to sublist.
      3. Otherwise:
        1. Let nodes to wrap be a list of nodes, initially empty.
        2. While nodes to wrap is empty, or node list is not empty and its first member is the nextSibling of nodes to wrap's last member and the first member of node list is an inline node and the last member of nodes to wrap is an inline node other than a br, remove the first member from node list and append it to nodes to wrap.
        3. Wrap nodes to wrap, with new parent instructions returning the result of calling createElement("li") on the context object. Append the result to sublist.
    3. In this case it's already wrapped properly, nothing more to do.

      If sublist's first member's parent is an HTML element with local name tag name, or if every member of sublist is an ol or ul, continue this loop from the beginning.

    4. If sublist's first member's parent is an HTML element with local name other tag name:
      1. Record the values of sublist, and let values be the result.
      2. Split the parent of sublist.
      3. Wrap sublist, with sibling criteria returning true for an HTML element with local name tag name and false otherwise, and new parent instructions returning the result of calling createElement(tag name) on the context object.
      4. Restore the values from values.
      5. Continue this loop from the beginning.
    5. Wrap sublist, with sibling criteria returning true for an HTML element with local name tag name and false otherwise, and new parent instructions being the following:
      1. Special case: something like

        <ol><li>foo</ol><blockquote>[bar]</blockquote>

        becomes

        <ol><li>foo</li><ol><li>[bar]</ol></ol>

        instead of

        <ol><li>foo</ol><blockquote><ol><li>[bar]</ol></blockquote>.

        We handle the special case in new parent instructions instead of outside because we'd prefer to wind up in a sibling if there is one. We handle only previousSibling, not nextSibling, because we really mean for this to cover something like

        [foo<blockquote>bar]</blockquote>

        which we'll handle node-by-node. TODO: Maybe we should do this differently, like just special-case simple indentation elements in an earlier part of the algorithm? This way's a bit weird.

        If sublist's first member's parent is not an editable simple indentation element, or sublist's first member's parent's previousSibling is not an editable HTML element with local name tag name, call createElement(tag name) on the context object and return the result.

      2. Let list be sublist's first member's parent's previousSibling.
      3. Normalize sublists of list's lastChild.
      4. If list's lastChild is not an editable HTML element with local name tag name, call createElement(tag name) on the context object, and append the result as the last child of list.
      5. Return the last child of list.
    6. Fix disallowed ancestors of the previous step's result.

Justifying the selection

This is the action for the four justify* commands. It's pretty straightforward, with no notable gotchas or special cases. It works more or less like a stripped-down version of set the selection's value, except it gets to be much simpler because it's much less general. (It's not similar enough to just invoke that algorithm: too many things differ between block and inline elements.)

There are two basic ways this works in browsers: using the align attribute, and using CSS text-align. IE9 and Opera 11.11 use only the align attribute, Chrome 13 dev uses only text-align, and Firefox 5.0a2 varies based on styleWithCSS. The two ways produce entirely different results, which is a real problem, so I don't think Firefox's approach is tenable. text-align is more valid, and for typical contenteditable cases it works the same. But for cases where you have fixed-width blocks, like tables or just divs with a width set, it behaves differently, and in those cases the align attribute is more useful.

TODO: text-align doesn't behave as expected if there are descendant blocks with non-100% width, like tables. The align attribute behaves a lot more nicely in such cases, but it's not valid. Not clear what to do. For now I've stuck with text-align, just because the cases where it misbehaves can't be created by any sequence of stock execCommand()s that I know of, but this needs more careful consideration. Gecko in CSS mode seems to special-case tables, adding auto margins to the table element to get it to align correctly.

TODO: We could do something along the lines of pushing down values here, although no browser does. In fact, it's very likely this can be rewritten in terms of the inline formatting command primitives, but it's not clear if it would be worth the added complexity.

To justify the selection to a string alignment (either "center", "justify", "left", or "right"):

  1. Block-extend the active range, and let new range be the result.
  2. No browser actually removes center, but it makes sense to do so.

    Let element list be a list of all editable Elements contained in new range that either has an attribute in the HTML namespace whose local name is "align", or has a style attribute that sets "text-align", or is a center.

  3. For each element in element list:
    1. If element has an attribute in the HTML namespace whose local name is "align", remove that attribute.
    2. Unset the CSS property "text-align" on element, if it's set by a style attribute.
    3. If element is a div or span or center with no attributes, remove it, preserving its descendants.
    4. If element is a center with one or more attributes, set the tag name of element to "div".
  4. This could theoretically be necessary, like if it converted "<div align=right>foo</div>bar" to "foo<br>bar". Now we need to select "foo<br>", nor just "foo".

    Block-extend the active range, and let new range be the result.

  5. Let node list be a list of nodes, initially empty.
  6. Of tested browsers, only Chrome 13 dev seems to not apply the alignment to nodes that are already aligned. Even then, it does apply it if the alignment is just inherited from the root.

    For each node node contained in new range, append node to node list if the last member of node list (if any) is not an ancestor of node; node is editable; node is an allowed child of "div"; and node's alignment value is not alignment.

  7. While node list is not empty:
    1. Let sublist be a list of nodes, initially empty.
    2. Remove the first member of node list and append it to sublist.
    3. While node list is not empty, and the first member of node list is the nextSibling of the last member of sublist, remove the first member of node list and append it to sublist.
    4. Wrap sublist. Sibling criteria returns true for any div that has one or both of the following two attributes and no other attributes, and false otherwise:

      • An align attribute whose value is an ASCII case-insensitive match for alignment.
      • A style attribute which sets exactly one CSS property (including unrecognized or invalid attributes), which is "text-align", which is set to alignment.

      As with inline formatting, I only ever create new elements, and don't ever modify existing ones. This doesn't match how any browser behaves in this case, but for inline formatting it matches everyone but Gecko's CSS mode, so I'm just being consistent.

      New parent instructions are to call createElement("div") on the context object, then set its CSS property "text-align" to alignment and return the result.

Automatic linking

Bug 13807.

When the user inserts whitespace immediately following something that looks like a URL or e-mail address, we automatically run the createLink command on it.

An autolinkable URL is a string of the following form:

  1. IE9 and LibreOffice 3.3.4 both have a whitelist of URL schemes. That would be complicated and involve political decisions, so instead, we'll just accept anything that looks like a hierarchical URL scheme. Google Docs is similar (as of November 9, 2011), but it's too lax, and allows characters in the scheme that can't be in a scheme. For non-hierarchical schemes, we just whitelist mailto:, since it's the only common one that makes sense to autolink.

    Either a string matching the scheme pattern from RFC 3986 section 3.1 followed by the literal string ://, or the literal string mailto:; followed by

  2. We don't try to enforce that the URL is anything resembling valid per spec. Too complicated for not enough gain.

    Zero or more characters other than space characters; followed by

  3. If the user types a URL followed by some punctuation, we still want to autolink, but we don't want to include the punctuation if it's probably not meant as part of the URL.

    IE9 excludes !#&()*+,-.:;<=?@[]^_`{|}~ as trailing characters from both URLs and e-mails. A trailing " or > will prevent autolinking, and a trailing $%'/\ is included in the link.

    LibreOffice 3.3.4 excludes trailing !”#'()*+,.:;<=>?[\]^_`{|}~, and prevents autolinking on $%&-@. It includes a trailing / in URLs, but it inhibits linking for e-mails.

    Google Docs (as of November 9, 2011) is complicated. Trailing ”’,-. always prevents autolinking of a URL, and trailing #%/?_ is always included in a URL. Trailing !&=$()*+:;<>@[\]^`{|}~ prevent autolinking if there's no ? before them, but are included in the URL if there is a ?. For e-mails, _ is included, and everything else prevents autolinking.

    None of these behaviors makes maximal sense. We should exclude characters if they're more likely as delimiters than actual trailing characters; include them if they're more likely as actual trailing characters; and prevent autolinking if their presence suggests that we're not actually looking at a link or e-mail address. The lists I made up for URLs are: exclude trailing !"'(),-.:;<>[]`{}, include anything else. For e-mail, exclude anything at all.

    A character that is not one of the ASCII characters !"'(),-.:;<>[]`{}.

To autolink (node, end offset):

  1. While (node, end offset)'s previous equivalent point is not null, set it to its previous equivalent point.
  2. If node is not a Text node, or has an a ancestor, do nothing and abort these steps.
  3. Let search be the largest substring of node's data whose end is end offset and that contains no space characters.
  4. If some substring of search is an autolinkable URL:
    1. While there is no substring of node's data ending at end offset that is an autolinkable URL, decrement end offset.
    2. Let start offset be the start index of the longest substring of node's data that is an autolinkable URL ending at end offset.
    3. Let href be the substring of node's data starting at start offset and ending at end offset.
  5. Otherwise, if some substring of search is a valid e-mail address:
    1. While there is no substring of node's data ending at end offset that is a valid e-mail address, decrement end offset.
    2. Let start offset be the start index of the longest substring of node's data that is a valid e-mail address ending at end offset.
    3. Let href be "mailto:" concatenated with the substring of node's data starting at start offset and ending at end offset.
  6. Otherwise, do nothing and abort these steps.
  7. Let original range be the active range.
  8. Create a new range with start (node, start offset) and end (node, end offset), and set the context object's selection's range to it.
  9. Take the action for "createLink", with value equal to href.
  10. Set the context object's selection's range to original range.

The delete command

This is the same as hitting backspace (see Additional requirements). The easy part is if the selection isn't collapsed: just delete the selection. But it turns out rich-text editors have a lot of special behaviors for hitting backspace with a collapsed selection. Most obviously, if there's a text node right before the cursor (maybe wrapped in some inline elements), we delete its last character. But some of the special cases we run into are:

Preserves overrides

For all the deletions here, Firefox 7.0a2 will remove wrapper elements like <b> only if they're selected, like {<b>foo</b>}. IE9, Chrome 14 dev, and Opera 11.50 will all remove them even if only their contents are selected, like <b>[foo]</b>. Gecko's behavior in the latter case leaves things like <b>{}</b> in the DOM, which is unhelpful, so I don't.

Action:

  1. If the active range is not collapsed, delete the selection and return true.
  2. Needed so that if there are multiple consecutive spaces we backspace over all at once.

    Canonicalize whitespace at the active range's start.

  3. Let node and offset be the active range's start node and offset.
  4. First go up as high as possible within the current block, then drill down to the lowest possible level, in the hopes that we'll wind up at the end of a text node, or maybe in a br or hr.

    Repeat the following steps:

    1. If there's an invisible node somewhere, Firefox 5.0a2 removes that node and then stops, so each backspace removes one invisible node. All others remove the invisible node and then continue on looking for something visible to remove. The spec follows the latter behavior, since it makes more sense to the user. Of course, the definition of "invisible node" is not necessarily anything like the spec's.

      If offset is zero and node's previousSibling is an editable invisible node, remove node's previousSibling from its parent.

    2. Otherwise, if node has a child with index offset − 1 and that child is an editable invisible node, remove that child from node, then subtract one from offset.
    3. Otherwise, if offset is zero and node is an inline node, or if node is an invisible node, set offset to the index of node, then set node to its parent.
    4. When backspacing a link, Firefox 7.0a2, Chrome 14 dev, Opera 11.50, and OpenOffice.org 3.2.1 Ubuntu have no special behavior. IE9 and Word 2007 remove the link instead of deleting its last character. The latter behavior seems more useful and intuitive.

      Otherwise, if node has a child with index offset − 1 and that child is an editable a, remove that child from node, preserving its descendants. Then return true.

    5. Otherwise, if node has a child with index offset − 1 and that child is not a block node or a br or an img, set node to that child, then set offset to the length of node.
    6. Otherwise, break from this loop.
  5. At this point, node cannot be an invisible node. There are three cases:

    1. offset is zero and node is a block node. Then we'll usually merge with the previous block if one exists.
    2. offset is not zero, node is not a block node, and node does not have a child with index offset − 1. The only way this is possible is if node has a length greater than zero but no children, which implies it's a text or comment or PI. Comments and PIs are invisible nodes, so it must be a text node. We delete the previous character.
    3. offset is not zero, and the child of node with index offset − 1 is a block node or a br or an img. Then we'll usually merge the offsetth child of node with the last descendant of the offset − 1st.

    Unlike forwardDelete, there's no special case for diacritics. This means backspacing will just delete the last combining diacritic typed, or the whole character if it's precomposed. This matches everything I tested (IE9, Firefox 7.0a2, Chrome 14 dev, etc.).

    If node is a Text node and offset is not zero, or if node is a block node that has a child with index offset − 1 and that child is a br or hr or img:

    1. Call collapse(node, offset) on the context object's selection.
    2. Call extend(node, offset − 1) on the context object's selection.
    3. Delete the selection.
    4. Return true.
  6. At the time of this writing, this should be impossible. Just being safe.

    If node is an inline node, return true.

  7. If we're at the beginning of a list, we want to outdent the first list item. This doesn't actually match anyone or anything. Word 2007 and OpenOffice.org 3.2.1 Ubuntu just remove the list marker, which is weird and doesn't map well to HTML. Browsers tend to just merge with the preceding block, which isn't expected.

    If node is an li or dt or dd and is the first child of its parent, and offset is zero:

    1. Let items be a list of all lis that are ancestors of node.
    2. Normalize sublists of each item in items.
    3. Record the values of the one-node list consisting of node, and let values be the result.
    4. Split the parent of the one-node list consisting of node.
    5. Restore the values from values.
    6. Annoying hack to prevent the dl from being re-added when fixing disallowed ancestors. In most cases we want a wrapper dl added, but in two cases (delete and insertParagraph) we're actually trying to outdent the list item. TODO: there might be a better way to do this.

      If node is a dd or dt, and it is not an allowed child of any of its ancestors in the same editing host, set the tag name of node to the default single-line container name and let node be the result.

    7. Fix disallowed ancestors of node.
    8. Return true.
  8. By this point, we're almost certainly going to merge something, and the only question is what.

    Let start node equal node and let start offset equal offset.

  9. Repeat the following steps:
    1. If start offset is zero, set start offset to the index of start node and then set start node to its parent.
    2. Otherwise, if start node has an editable invisible child with index start offset minus one, remove it from start node and subtract one from start offset.
    3. Otherwise, break from this loop.
  10. At the beginning of an indented block, outdent it, similar to a list item. Browsers don't do this, word processors do. Note: this copy-pastes from the outdent command action.

    If offset is zero, and node has an editable inclusive ancestor in the same editing host that's an indentation element:

    1. Block-extend the range whose start and end are both (node, 0), and let new range be the result.
    2. Let node list be a list of nodes, initially empty.
    3. For each node current node contained in new range, append current node to node list if the last member of node list (if any) is not an ancestor of current node, and current node is editable but has no editable descendants.
    4. Outdent each node in node list.
    5. Return true.
  11. This is to avoid stripping a line break from

    foo<br><br><table><tr><td>[]bar</table>

    and similarly for <hr>. We should just do nothing here.

    If the child of start node with index start offset is a table, return true.

  12. If you try backspacing into a table, select it. This doesn't match any browser; it matches the recommendation of the "behavior when typing in contentEditable elements" document. The idea is that then you can delete it with a second backspace.

    If start node has a child with index start offset − 1, and that child is a table:

    1. Call collapse(start node, start offset − 1) on the context object's selection.
    2. Call extend(start node, start offset) on the context object's selection.
    3. Return true.
  13. Special case:

    <p>foo</p><br><p>[]bar</p>
    -> <p>foo</p><p>[]bar</p>

    and likewise for <hr>. But with <img> we merge like in other cases:

    <p>foo</p><img><p>[]bar</p>
    -> <p>foo</p><img>[]bar.

    Browsers don't do this consistently. Firefox 5.0a2 doesn't seem to do it at all.

    If offset is zero; and either the child of start node with index start offset minus one is an hr, or the child is a br whose previousSibling is either a br or not an inline node:

    1. Call collapse(start node, start offset − 1) on the context object's selection.
    2. Call extend(start node, start offset) on the context object's selection.
    3. Delete the selection.
    4. Call collapse(node, offset) on the selection.
    5. Return true.
  14. If you try backspacing out of a list item, merge it with the previous item, but add a line break. Then you have to backspace again if you really want them to be on the same line. This matches Word 2007 and OpenOffice.org 3.2.1 Ubuntu, and also matches "behavior when typing in contentEditable elements", but does not match any browser.

    Note that this behavior is quite different from what happens if you actually select the linebreak in between the two lines. In that case, the blocks are merged as normal.

    Also note that hitting backspace twice will merge with the previous item. This matches OO.org, but Word will outdent the item on subsequent backspaces. Word's behavior doesn't fit well with the way lists work in HTML, and we probably don't want it.

    If the child of start node with index start offset is an li or dt or dd, and that child's firstChild is an inline node, and start offset is not zero:

    1. Let previous item be the child of start node with index start offset minus one.
    2. If the last child is already a br, we only need to append one extra br. Otherwise we need to append two, since the first will do nothing.

      If previous item's lastChild is an inline node other than a br, call createElement("br") on the context object and append the result as the last child of previous item.

    3. If previous item's lastChild is an inline node, call createElement("br") on the context object and append the result as the last child of previous item.
  15. When merging adjacent list items, make sure we only merge the items themselves, not any block children. We want <li><p>foo<li><p>[]bar to become <li><p>foo<p>[]bar, not <li><p>foo<br>[]bar or <li><p>foo[]bar. To do the deletion, we need to wipe out the current selection, so we save it as a range. Saving it as a node/offset pair isn't enough, because it might be invalid after we do the deletion. A range will update according to the range mutation rules.

    If start node's child with index start offset is an li or dt or dd, and that child's previousSibling is also an li or dt or dd:

    1. Call cloneRange() on the active range, and let original range be the result.
    2. Set start node to its child with index start offset − 1.
    3. Set start offset to start node's length.
    4. Set node to start node's nextSibling.
    5. Call collapse(start node, start offset) on the context object's selection.
    6. Call extend(node, 0) on the context object's selection.
    7. Delete the selection.
    8. Call removeAllRanges() on the context object's selection.
    9. Call addRange(original range) on the context object's selection.
    10. Return true.
  16. General block-merging case.

    While start node has a child with index start offset minus one:

    1. If start node's child with index start offset minus one is editable and invisible, remove it from start node, then subtract one from start offset.
    2. Otherwise, set start node to its child with index start offset minus one, then set start offset to the length of start node.
  17. Call collapse(start node, start offset) on the context object's selection.
  18. Call extend(node, offset) on the context object's selection.
  19. Delete the selection, with direction "backward".
  20. Return true.

The formatBlock command

This command lets you change what block element particular lines are wrapped in. It will convert an existing wrapper if one exists, and otherwise will create a new one.

A formattable block name is "address", "dd", "div", "dt", "h1", "h2", "h3", "h4", "h5", "h6", "p", or "pre".

Tested browser versions: IE9, Firefox 4.0, Chrome 13 dev, Opera 11.10.

Firefox and Chrome will replace a <blockquote> by a <p> or other given tag. IE and Opera will nest the <p> inside instead. The latter makes more sense, given that a) we don't support formatBlock with <blockquote> and b) <blockquote>s are logically different, since they can contain many lines.

Firefox will not convert other tags like <p> to <div>, it will only wrap unwrapped lines in a <div>. Firefox also won't replace <div> by things like <p>, it will nest the <p> inside. The spec follows other browsers.

If you try to convert a <dt> to a <div> or <p> or such, Firefox breaks out of the <dl> entirely, leaving ...<dt><br></dt></dl>. Chrome will convert a <dt> or <dd> to the given element, leaving a <div> or <p> or such as the child of a <dl>. I follow IE/Opera, which only affect the contents of <dt>/<dd> (Firefox behaves this way for <dd> as well, just not <dt>). This means you can get invalid DOMs like <dt><p>foo<p></dt>, but they can be serialized as text/html, so I'm not too fussy.

When it comes to <li>, IE/Opera behave like with <dt>/<dd>, which is how I behave too. Firefox apparently refuses to do anything. Chrome tries to wrap the parent list element, breaking it up if only some of the children are selected; this produces unserializable DOMs if you're wrapping with <p>.

When you're converting multiple blocks at once, Chrome replaces them all by one block with <br> stuck in, like <p>foo</p><p>bar</p> -> <div>foo<br>bar</div>. It wipes out intervening block containers too in some cases. This might make sense for <address>/<h*>/<pre>, but other browsers don't do it.

Preserves overrides

Action:

  1. IE9 requires the brackets. If they're not provided, it does nothing.

    If value begins with a "<" character and ends with a ">" character, remove the first and last characters from it.

  2. Let value be converted to ASCII lowercase.
  3. Opera 11.10 throws NOT_SUPPORTED_ERR for bad elements, all other tested browsers ignore the input. Testing in IE9, Firefox 4.0, Chrome 13 dev, and Opera 11.10, supported elements seem to be:

    Everyone
    address, div, h*, p, pre
    Everyone but IE
    blockquote
    Everyone but Opera
    dd, dt
    IE only
    dir, menu, ol, ul
    Firefox and Chrome only
    dl
    Chrome only
    article, aside, footer, header, hgroup, nav, section

    HTML5 as of May 2011 supports: address, article, aside, blockquote, div, footer, h*, header, hgroup, nav, p, pre, section, which exactly matches Chrome except minus dd/dt/dl.

    See mailing list discussion on the subject.

    If value is not a formattable block name, return false.

  4. Block-extend the active range, and let new range be the result.
  5. Let node list be an empty list of nodes.
  6. For each node node contained in new range, append node to node list if it is editable, the last member of original node list (if any) is not an ancestor of node, node is either a non-list single-line container or an allowed child of "p" or a dd or dt, and node is not the ancestor of a prohibited paragraph child.
  7. Record the values of node list, and let values be the result.
  8. This tries to avoid misnesting if only some lines of an element are selected, so <h1>[foo]<br>bar</h1> becomes <p>[foo]</p><h1>bar</h1> instead of <h1><p>[foo]</p><br>bar</h1> or such. It tries to heuristically distinguish between divs used as line-breakers and divs used as actual wrappers by checking if they have prohibited paragraph children as descendants. It works for address too, in case there are paragraphs nested inside. Thus <address>[foo]<br>bar</address> becomes <p>[foo]</p><address>bar</address>, but <address>[foo]<p>bar</p></address> becomes <address><p>[foo]</p><p>bar</p></address>. Likewise, we don't break things out of lists or tables or such if they happen to be nested in a <div>.

    For each node in node list, while node is the descendant of an editable HTML element in the same editing host, whose local name is a formattable block name, and which is not the ancestor of a prohibited paragraph child, split the parent of the one-node list consisting of node.

  9. Restore the values from values.
  10. We have two different behaviors, one for div and p and one for everything else. The basic difference is that for div and p, we assume that it should be one line per element, while for other elements, we put in multiple lines separated by <br>. So if you do formatBlock to p on

    <div>foo</div><div>bar</div>

    or

    foo<br>bar

    you get

    <p>foo</p><p>bar</p>

    but formatBlock to h1 will get you

    <h1>foo<br>bar</h1>.

    IE9 will just change the elements as they are, so it gives <p>foo</p><p>bar</p> and <h1>foo</h1><h1>bar</h1> for <div>foo</div><div>bar</div>, but <p>foo<br>bar</p> and <h1>foo<br>bar</h1> for foo<br>bar. This is unreasonable, because the two possible inputs here look identical to the user and might have been produced by identical user input.

    Firefox 5.0a2 will give results like <p>foo</p><p>bar</p> or <h1>foo</h1><h1>bar</h1> no matter what (modulo oddities in its handling of divs). Opera 11.10 is similar, except it leaves a trailing <br> in the first element.

    Chrome 13 dev will give results like <p>foo<br>bar</p> or <h1>foo<br>bar</h1> no matter what.

    The specced behavior is a compromise between the existing behaviors, predicated on the fact that <h1>foo</h1><h1>bar</h1> almost never makes sense, and <p>foo<br>bar</p> isn't usually what's wanted either.

    While node list is not empty:

    1. If the first member of node list is a single-line container:
      1. If you try to format a single-line container with no children, IE10PP2 inserts an nbsp before formatting. (It uses nbsp instead of <br> to make blocks not collapse, so the equivalent for us would be to insert a <br>.) Firefox 7.0a2 and Opera 11.50 make the element disappear. Chrome 14 dev leaves it alone and doesn't format it. I follow Firefox/Opera just because it's the simplest given how I happen to have written the spec, and it's a corner case, so exact behavior isn't important.

        For blocks that contain only a collapsed whitespace node, IE10PP2 and Firefox 7.0a2 convert them like normal. Chrome 14 dev and Opera 11.50 leave it alone and don't format it. I go with the majority, which is again simpler to spec.

        Let sublist be the children of the first member of node list.

      2. Record the values of sublist, and let values be the result.
      3. Remove the first member of node list from its parent, preserving its descendants.
      4. Restore the values from values.
      5. Remove the first member from node list.
    2. Otherwise:
      1. Let sublist be an empty list of nodes.
      2. Remove the first member of node list and append it to sublist.
      3. While node list is not empty, and the first member of node list is the nextSibling of the last member of sublist, and the first member of node list is not a single-line container, and the last member of sublist is not a br, remove the first member of node list and append it to sublist.
    3. Wrap sublist. If value is "div" or "p", sibling criteria returns false; otherwise it returns true for an HTML element with local name value and no attributes, and false otherwise. New parent instructions return the result of running createElement(value) on the context object. Then fix disallowed ancestors of the result.
  11. Return true.

Firefox 6.0a2 throws, Chrome 14 dev always returns false, Opera 11.11 doesn't support indeterm to start with, IE9 was uncooperative in testing so I'm not sure what it does. I'm speccing it just because it makes sense.

Indeterminate:

  1. If the active range is null, return the empty string.
  2. Block-extend the active range, and let new range be the result.
  3. Let node list be all visible editable nodes that are contained in new range and have no children.
  4. If node list is empty, return false.
  5. Let type be null.
  6. For each node in node list:
    1. While node's parent is editable and in the same editing host as node, and node is not an HTML element whose local name is a formattable block name, set node to its parent.
    2. Let current type be the empty string.
    3. If node is an editable HTML element whose local name is a formattable block name, and node is not the ancestor of a prohibited paragraph child, set current type to node's local name.
    4. If type is null, set type to current type.
    5. Otherwise, if type does not equal current type, return true.
  7. Return false.

IE9 returns human-readable strings like "Normal" (p/div/etc.), "Formatted" (pre), "Heading 1" (h1), etc. Firefox 6.0a2 and Chrome 14 dev both return the appropriate tag name in lowercase, or the empty string if there is no appropriate tag. Opera 11.11 behaves the same, but with uppercase.

IE9 looks like it recognizes address, h*, pre, dd, dt, ol, ul, and dir, with everything else registering as "Normal". Firefox 6.0a2 recognizes only the arguments it accepts for formatBlock, namely address, h*, p, and pre. Chrome 14 dev recognizes address, div, h*, dd, dl, dt, p, pre plus lots of random other stuff like blockquote and section. I'll go with everything that execCommand("formatblock") accepts as an argument, which at the time of this writing means what Firefox supports plus div.

Value:

  1. If the active range is null, return the empty string.
  2. Block-extend the active range, and let new range be the result.
  3. Let node be the first visible editable node that is contained in new range and has no children. If there is no such node, return the empty string.
  4. Opera 11.11 doesn't require it be editable, so it will return "DIV" instead of "" for <div contenteditable>foo</div>.

    While node's parent is editable and in the same editing host as node, and node is not an HTML element whose local name is a formattable block name, set node to its parent.

  5. Chrome 14 dev will report "div" for <div><ol><li>foo</ol></div> or such. Opera 11.11 reports "". IE and Firefox didn't cooperate with testing. Opera makes more sense, and matches the fact that formatBlock now doesn't recognize such a div as a formatBlock candidate, so Opera it is.

    We don't really need to specify "editable" here, since it has to be editable if we got to this point.

    If node is an editable HTML element whose local name is a formattable block name, and node is not the ancestor of a prohibited paragraph child, return node's local name, converted to ASCII lowercase.

  6. Return the empty string.

The forwardDelete command

This is the same as hitting the delete key (see Additional requirements). It behaves much the same as the delete command, except of course backwards. Also, some of the special cases for backspacing don't apply, as noted in the comments. The one special case you get when deleting forward but not backward is that if the cursor is before a grapheme cluster that consists of multiple characters, like a base character with combining diacritics, we delete the diacritics too. (Backspacing just deletes the last diacritic, so you have to backspace several times to remove the whole cluster.)

Preserves overrides

Copy-pasted from delete, see there for comments.

Action:

  1. If the active range is not collapsed, delete the selection and return true.
  2. Canonicalize whitespace at the active range's start.
  3. Let node and offset be the active range's start node and offset.
  4. Repeat the following steps:
    1. If offset is the length of node and node's nextSibling is an editable invisible node, remove node's nextSibling from its parent.
    2. Otherwise, if node has a child with index offset and that child is an editable invisible node, remove that child from node.
    3. Otherwise, if offset is the length of node and node is an inline node, or if node is invisible, set offset to one plus the index of node, then set node to its parent.
    4. No special link behavior for forwardDelete here, unlike delete.

      Otherwise, if node has a child with index offset and that child is neither a block node nor a br nor an img nor a collapsed block prop, set node to that child, then set offset to zero.

    5. Otherwise, break from this loop.
  5. If node is a Text node and offset is not node's length:
    1. Let end offset be offset plus one.
    2. Firefox 7.0a2, Chrome 14 dev, Word 2007, and OpenOffice.org 3.2.1 Ubuntu act as the spec says, getting rid of all diacritics on forward delete. IE9 and Opera 11.50 have no special case and just delete the next character. I go with Firefox/Chrome/Word/OO.

      However, when I actually type in the text box as opposed to running semi-automated tests, IE9 has magical behavior: it replaces the base character with something that looks like ◌ U+25CC DOTTED CIRCLE. Further strikes of the delete key remove the diacritics, and the circle vanishes along with the last of them. I wasn't able to get it to actually replace the base character, so I'm not sure what the point is. The circle doesn't seem to appear in the DOM, and apparently it disappears in some circumstances. This might be worth standardizing somehow, I don't know.

      TODO: The way we remove diacritics is probably not right. We probably want to normalize to grapheme cluster boundaries, using UAX#29 or something. We also need to handle non-BMP stuff. The idea is that if the cursor is before a character that precedes a combining mark, you need to delete the combining mark too.

      While end offset is not node's length and the end offsetth code unit of node's data has general category M when interpreted as a Unicode code point, add one to end offset.

    3. Call collapse(node, offset) on the context object's selection.
    4. Call extend(node, end offset) on the context object's selection.
    5. Delete the selection.
    6. Return true.
  6. If node is an inline node, return true.
  7. If node has a child with index offset and that child is a br or hr or img, but is not a collapsed block prop:
    1. Call collapse(node, offset) on the context object's selection.
    2. Call extend(node, offset + 1) on the context object's selection.
    3. Delete the selection.
    4. Return true.
  8. No special list-item behavior for forwardDelete here, unlike delete.

    Let end node equal node and let end offset equal offset.

  9. If end node has a child with index end offset, and that child is a collapsed block prop, add one to end offset.
  10. Repeat the following steps:
    1. If end offset is the length of end node, set end offset to one plus the index of end node and then set end node to its parent.
    2. Otherwise, if end node has an editable invisible child with index end offset, remove it from end node.
    3. Otherwise, break from this loop.
  11. No special indentation element behavior for forwardDelete here, unlike delete.

    If the child of end node with index end offset minus one is a table, return true.

  12. If the child of end node with index end offset is a table:
    1. Call collapse(end node, end offset) on the context object's selection.
    2. Call extend(end node, end offset + 1) on the context object's selection.
    3. Return true.
  13. Note, any br will do here: a br immediately after a block is always significant.

    If offset is the length of node, and the child of end node with index end offset is an hr or br:

    1. Call collapse(end node, end offset) on the context object's selection.
    2. Call extend(end node, end offset + 1) on the context object's selection.
    3. Delete the selection.
    4. Call collapse(node, offset) on the selection.
    5. Return true.
  14. No special list-item behavior for forwardDelete here, unlike delete.

    While end node has a child with index end offset:

    1. If end node's child with index end offset is editable and invisible, remove it from end node.
    2. Otherwise, set end node to its child with index end offset and set end offset to zero.
  15. Call collapse(node, offset) on the context object's selection.
  16. Call extend(end node, end offset) on the context object's selection.
  17. Delete the selection.

The indent command

IE9
Outputs <blockquote style="margin-right: 0px" dir="ltr">, or when surrounding RTL blocks, <blockquote style="margin-left: 0px" dir="rtl">. The direction seems to go by the end of the selection. The presence of the dir attribute means that any contents that were inheriting a different dir from an ancestor get their direction changed as a side effect, but if they actually have the opposite dir specified, they won't appear to be indented. It doesn't reset top or bottom margins on the blockquote, so it adds them. If it's not wrapping a block element, like if it's only wrapping up until a <br>, it adds a <p>.
Firefox 4.0
In styleWithCSS mode, adds style="margin-left: 40px" to the appropriate block container (or margin-right if it's RTL). If there's no appropriate block container, adds a div. If multiple blocks are affected, it goes by the direction of the block whose style it's changing, which winds up being wrong for descendants with different direction. In non-styleWithCSS mode, uses <blockquote>, so it indents on both sides and also adds top/bottom margins.
Chrome 12 dev
Outputs <blockquote class="webkit-indent-blockquote" style="margin: 0 0 0 40px; border: none; padding: 0px"> in both modes for both LTR and RTL (which is broken for RTL, since it indents only on the left).
Opera 11.00
Outputs <blockquote>, so it indents on both sides and on the top/bottom.

For repeated indentation, everyone except Opera that outputs <blockquote>s just puts them at the outermost possible location, which works well. Opera puts them in the innermost position, which is broken, because it will even put them inside <p> (which will not round-trip through text/html serialization).

Gecko in CSS mode messes up by adding margins even to things like <blockquote> that already have margins from CSS rules, instead of nesting a div, so it doesn't actually increase the indentation. However, if an element has an explicit left margin (assuming LTR), it will increase the margin to 80px, so it works with WebKit's blockquotes.

We have two strategies for handling directionality: always indent on both sides (Firefox non-CSS, Opera) or try to figure out heuristically which side we want (IE, Firefox CSS). The latter approach is only possible by adding extra markup and complexity, so for now we'll take the easy way out and go with just indenting on both sides.

This reasoning doesn't discuss lists. For research on lists, see the comment for insertOrderedList. List handling is more complicated and I wound up differing from all browsers in lots of ways.

Preserves overrides

Action:

  1. Let items be a list of all lis that are inclusive ancestors of the active range's start and/or end node.
  2. TODO: This overnormalizes, but it seems like the simplest solution for now.

    For each item in items, normalize sublists of item.

  3. Block-extend the active range, and let new range be the result.
  4. Let node list be a list of nodes, initially empty.
  5. For each node node contained in new range, if node is editable and is an allowed child of "div" or "ol" and if the last member of node list (if any) is not an ancestor of node, append node to node list.
  6. Without this step, the last child of the previous sibling might be a list, which the li wouldn't get appended to.

    If the first visible member of node list is an li whose parent is an ol or ul:

    1. Let sibling be node list's first visible member's previousSibling.
    2. While sibling is invisible, set sibling to its previousSibling.
    3. If sibling is an li, normalize sublists of sibling.
  7. While node list is not empty:
    1. Let sublist be a list of nodes, initially empty.
    2. Remove the first member of node list and append it to sublist.
    3. While the first member of node list is the nextSibling of the last member of sublist, remove the first member of node list and append it to sublist.
    4. Indent sublist.

The insertHorizontalRule command

Preserves overrides

You'd think interop here would be simple, right? Nope: we have three different behaviors across four browsers. Opera 11.00 is the only one that acts more or less like the spec. IE9 and Chrome 12 dev treat the value as an id, which is weird and probably useless, so I don't do it. Firefox 4.0 produces <hr size=2 width=100%> instead of <hr>, which is also weird and almost definitely useless, so I don't do it. Then you have the varying behavior in splitting up parents to ensure validity . . .

Action:

  1. Let start node, start offset, end node, and end offset be the active range's start and end nodes and offsets.
  2. While start offset is 0 and start node's parent is not null, set start offset to start node's index, then set start node to its parent.
  3. While end offset is end node's length, and end node's parent is not null, set end offset to one plus end node's index, then set end node to its parent.
  4. Call collapse(start node, start offset) on the context object's selection.
  5. Call extend(end node, end offset) on the context object's selection.
  6. Delete the selection, with block merging false.
  7. If the active range's start node is neither editable nor an editing host, return true.
  8. We don't want to call insertNode at the start or end of a text node, because that will leave an empty text node.

    If the active range's start node is a Text node and its start offset is zero, call collapse() on the context object's selection, with first argument the active range's start node's parent and second argument the active range's start node's index.

  9. If the active range's start node is a Text node and its start offset is the length of its start node, call collapse() on the context object's selection, with first argument the active range's start node's parent, and the second argument one plus the active range's start node's index.
  10. Let hr be the result of calling createElement("hr") on the context object.
  11. Run insertNode(hr) on the active range.
  12. IE9 and Chrome 13 dev seem to never break up any ancestors, which can lead to unserializable DOMs like <hr> inside <p>. Opera 11.11 seems to always break up parents going all the way up to the contenteditable root, even ones like <div> that can contain <hr>. Firefox 5.0a2 acts the most sensibly: it only breaks up things like <p> or <b> that shouldn't contain <hr>. The spec goes with Firefox here (although the list of what to break up isn't precisely identical).

    Fix disallowed ancestors of hr.

  13. Run collapse() on the context object's selection, with first argument hr's parent and the second argument equal to one plus hr's index.
  14. Return true.

The insertHTML command

Preserves overrides

Not supported by IE9. Handling of disallowed children is interesting:

Firefox 5.0a2
Will allow <dt> inside <dt> (doesn't serialize). If you try inserting dir/ol/ul inside an existing dir/ol/ul, it will strip the list element and leave only the li's, so inserting <ul><li>abc</ul> into <ol><li>f[o]o</ol> creates <ol><li>f<li>abc<li>o</ol>. <dt>/<dd>/<li> that don't descend from a list will be left alone, not converted to <p>. Empty elements seem not to be inserted. <li> will get put inside <p>, which breaks serialization. Nothing is allowed inside <xmp>, not even text.
Chrome 13 dev
Inserting a <p> into a <p> or <li> or such will remove the child <p>, adding its contents to the parent instead. Adding an <li> or <hr> as the child of a <p> works, as does an <a> inside an <a>, <h2> inside <h1>, <li> inside <li>, <nobr> inside <nobr>, <b> inside <xmp>, etc. (all unserializable). But <dt> and <dd> seem to get converted to their contents like <p>. <ol><li>abc</ol> inside <ol><li>f[o]o</ol> becomes <ol><li>f<li>abc<li><li>o</ol>, interestingly (note the empty <li>). I don't understand how it works, but it doesn't seem to make much sense.
Opera 11.11
Seems to do almost no validity or serialization checks, except that it prevents <a> inside <a>, <nobr> inside <nobr>, and block elements inside inline elements. Interestingly, most of the places where it's non-serializable per HTML parsing are actually serializable in Opera's own parser.

Action:

  1. Chrome 14 dev and Opera 11.11 do this even if the value is empty. Firefox 5.0a2 throws an exception.

    Firefox 7.0a2 and Chrome 14 dev do strip wrappers here, so inserting HTML in the place of <b>[foo]</b> will remove the <b>. Opera 11.50 keeps the wrappers. I follow the majority and leave the "strip wrappers" flag true.

    Delete the selection.

  2. If the active range's start node is neither editable nor an editing host, return true.
  3. TODO: This has some interesting consequences. For instance, table cells and similar will just vanish if they're not in an appropriate place; and inside a script or style or xmp or such, the argument will effectively be HTML-escaped before use. Some of these consequences might be undesirable.

    Let frag be the result of calling createContextualFragment(value) on the active range.

  4. Let last child be the lastChild of frag.
  5. Firefox 5.0a2 also seems to not add empty elements like <b></b>, but Chrome 13 dev and Opera 11.11 do.

    If last child is null, return true.

  6. Let descendants be all descendants of frag.
  7. This is so we don't get something like

    <div>[foo]</div>
    -> <div>{}<br></div>
    -> <div><p>Some HTML{}</p><br></div>

    with an extra bogus line break at the end.

    If the active range's start node is a block node:

    1. Let collapsed block props be all editable collapsed block prop children of the active range's start node that have index greater than or equal to the active range's start offset.
    2. For each node in collapsed block props, remove node from its parent.
  8. We could canonicalize whitespace after inserting the fragment, but let's not. If the author wants HTML, give them HTML behavior. When asked to replace a paragraph's contents with a single space, Firefox 7.0a2 does so but inserts a <br> before it (not after); Chrome 14 dev does so and doesn't insert a <br>, so the paragraph collapses; Opera 11.50 doesn't insert the space at all, and just inserts a <br>. Correct behavior is to insert a space, then insert a <br> after it in the next step because it's an invisible node.

    Call insertNode(frag) on the active range.

  9. In case we remove all the contents, then remove the extra <br>, then only add a comment or something. In that case we want to re-add the extra <br>. We don't try fixing the actual inserted content: that's the author's lookout.

    If the active range's start node is a block node with no visible children, call createElement("br") on the context object and append the result as the last child of the active range's start node.

  10. Need to do this before fixing disallowed ancestors, since otherwise the last child might have been removed (e.g., it's an li).

    Call collapse() on the context object's selection, with last child's parent as the first argument and one plus its index as the second.

  11. We want to fix all descendants, not just children. Consider <div><li>foo</li></div>, for example.

    Fix disallowed ancestors of each member of descendants.

  12. Return true.

The insertImage command

Preserves overrides

Action:

  1. Similar logic to createLink, except even more compelling, since an HTML document linking to itself as an image is just silly. In fact, the current HTML spec instructs UAs to not even try displaying the image, and just fail immediately if the URL is empty. Firefox 4b11 silently does nothing on an empty string, but the other three browsers I tested stick in the <img> anyway.

    If value is the empty string, return false.

  2. Firefox 7.0a2 seems to strip the wrapper or not depending on the exact positioning of the selection: <b>{foo}</b> yes, <b>[foo]</b> no. Chrome 14 dev seems to strip the wrapper regardless. Opera 11.50 seems to keep the wrapper, but place the image outside it. I didn't get IE to cooperate with my tests. I chose to leave wrappers across the board because they might be meaningful: e.g., a background-color when the image is small or not fully opaque.

    Delete the selection, with strip wrappers false.

  3. Let range be the active range.
  4. If the active range's start node is neither editable nor an editing host, return true.
  5. Same logic as with insertHTML.

    If range's start node is a block node whose sole child is a br, and its start offset is 0, remove its start node's child from it.

  6. Let img be the result of calling createElement("img") on the context object.
  7. No alt text, so it's probably invalid. This matches all browsers.

    Run setAttribute("src", value) on img.

  8. This winds up putting it at the original start point of the active range, as currently specced. This matches IE9 and Firefox 5.0a2. Chrome 13 dev puts it at the end point, and Opera 11.11 puts it in between (where the range would collapse if you called deleteContents()).

    Run insertNode(img) on range.

  9. Let selection be the result of calling getSelection() on the context object.
  10. Run collapse() on selection, with first argument equal to the parent of img and the second argument equal to one plus the index of img.
  11. Return true.

The insertLineBreak command

This is the same as hitting Shift-Enter or such (see Additional requirements). It deletes the selection, and replaces it with a <br>. No real complications.

Only implemented in WebKit (Chrome 14 dev). Other tests are entirely manual (i.e., actually hitting Shift-Enter instead of running a command). There's a surprisingly large amount of interop.

IE9 is tripped up by <xmp>, and also often doesn't add an extra <br> when the one it just inserted is extraneous.

Firefox 6.0a2 doesn't notice if you're trying to put the <br> in a bad place, which can result in unserializable DOMs.

Chrome 14 dev inserts a literal linebreak for pre and xmp and maybe other similar elements. This doesn't seem very useful, so I don't bother.

Opera 11.11 isn't heedful of <xmp>, and treats <pre> somewhat oddly.

Preserves overrides

Action:

  1. IE9 doesn't strip wrappers (IE10PP2 didn't work in tests). Firefox 7.0a2 strips wrappers inconsistently depending on the exact selection endpoints. Chrome 14 dev strips wrappers but recreates any styles using new wrappers. Opera 11.50 strips all wrappers.

    Delete the selection, with strip wrappers false.

  2. If the active range's start node is neither editable nor an editing host, return true.
  3. script, xmp, table, . . .

    If the active range's start node is an Element, and "br" is not an allowed child of it, return true.

  4. If the active range's start node is not an Element, and "br" is not an allowed child of the active range's start node's parent, return true.
  5. We don't want to call insertNode at the start or end of a text node, because that will leave an empty text node.

    If the active range's start node is a Text node and its start offset is zero, call collapse() on the context object's selection, with first argument equal to the active range's start node's parent and second argument equal to the active range's start node's index.

  6. If the active range's start node is a Text node and its start offset is the length of its start node, call collapse() on the context object's selection, with first argument equal to the active range's start node's parent and second argument equal to one plus the active range's start node's index.
  7. Let br be the result of calling createElement("br") on the context object.
  8. Call insertNode(br) on the active range.
  9. Call collapse() on the context object's selection, with br's parent as the first argument and one plus br's index as the second argument.
  10. If br is a collapsed line break, call createElement("br") on the context object and let extra br be the result, then call insertNode(extra br) on the active range.
  11. Return true.

The insertOrderedList command

Preserves overrides

Action: Toggle lists with tag name "ol", then return true.

Firefox 6.0a2 sort of supports this, but it throws exceptions most of the time. It has the quirk that even if there are no ol's around, it will say it's indeterminate if there are some things that are ul's and some that are not lists at all, but this doesn't make sense and I don't duplicate it. No one else supports indeterminate for insert*List, but it makes sense if you support state.

Indeterminate: True if the selection's list state is "mixed" or "mixed ol", false otherwise.

IE9 throws exceptions in most cases, Firefox 6.0a2 in some cases as well, for no apparent reason. Ignoring those, the spec basically matches all browsers, except with a few weird random mismatches that looked like browser bugs to me.

State: True if the selection's list state is "ol", false otherwise.

The insertParagraph command

This is the same as hitting enter (see Additional requirements). The general rule is that we find the nearest single-line container ancestor, clone it and insert the clone after it, and then move all the contents after the cursor (along with the cursor itself) to the clone. But there are a few special cases:

Preserves overrides

There are three major behaviors here. Firefox 5.0a2 behaves identically to execCommand("formatBlock", false, "p"), which is not really useful. IE9 actually just overwrites the selection with an empty paragraph element, which seems not very useful either. Chrome 13 dev and Opera 11.10 behave basically the same as if the user hit the Return key. This latter behavior seems much more useful, even though it's horribly misnamed, so it's what I'll spec. Comments about IE/Firefox are based on manual tests, where I hit Enter instead of running commands.

(Actually, Opera doesn't behave quite the same for insertParagraph and line breaks. But it's pretty close, and I expect the differences are bugs.)

Then, of course, we have several flavors of line-breaking behavior to choose from. Firefox prefers <br>s, unless it's in the middle of a <p> or something. Opera and IE like <p>. Chrome prefers <div>. And there are lots of subtleties besides. I go with IE/Opera-style behavior, as discussed in a whatwg thread.

Action:

  1. Delete the selection.
  2. If the active range's start node is neither editable nor an editing host, return true.
  3. Let node and offset be the active range's start node and offset.
  4. If node is a Text node, and offset is neither 0 nor the length of node, call splitText(offset) on node.
  5. If node is a Text node and offset is its length, set offset to one plus the index of node, then set node to its parent.
  6. If node is a Text or Comment node, set offset to the index of node, then set node to its parent.
  7. Call collapse(node, offset) on the context object's selection.
  8. Let container equal node.
  9. While container is not a single-line container, and container's parent is editable and in the same editing host as node, set container to its parent.
  10. Bug 13841.

    If container is an editable single-line container in the same editing host as node, and its local name is "p" or "div":

    1. Let outer container equal container.
    2. While outer container is not a dd or dt or li, and outer container's parent is editable, set outer container to its parent.
    3. If outer container is a dd or dt or li, set container to outer container.
  11. We add the default wrapper in this case.

    If container is not editable or not in the same editing host as node or is not a single-line container:

    1. Let tag be the default single-line container name.
    2. Block-extend the active range, and let new range be the result.
    3. Let node list be a list of nodes, initially empty.
    4. Append to node list the first node in tree order that is contained in new range and is an allowed child of "p", if any.
    5. If node list is empty:
      1. Ideally, we should normalize things so that the cursor is never in a weird place after deletion, but let's be safe and bail out if we do hit this scenario. It's not clear if we need this line in the long term, but at the time of this writing there's at least one corner case where deleting can leave the cursor inside a <tr>.

        If tag is not an allowed child of the active range's start node, return true.

      2. Set container to the result of calling createElement(tag) on the context object.
      3. Call insertNode(container) on the active range.
      4. Call createElement("br") on the context object, and append the result as the last child of container.
      5. Call collapse(container, 0) on the context object's selection.
      6. Return true.
    6. TODO: It is not at all obvious that this is the correct list of nodes in all cases. It should probably work because of how the block-extend algorithm works, but further thought would be good.

      While the nextSibling of the last member of node list is not null and is an allowed child of "p", append it to node list.

    7. Wrap node list, with sibling criteria returning false and new parent instructions returning the result of calling createElement(tag) on the context object. Set container to the result.
  12. IE9 and Chrome 13 dev just break <pre> up into multiple <pre>s. Firefox 5.0a2 and Opera 11.10 insert a <br> instead, treating it differently from <p>. The latter makes more sense. What might make the most sense is to just insert an actual newline character, though, since this is a pre after all . . .

    IE9 and Chrome 13 dev also break <address> up into multiple <address>es. Firefox 5.0a2 inserts <br> instead. Opera 11.10 nests <p>s inside. I don't like Opera's behavior, because it means we nest formatBlock candidates inside one another, so I'll go with Firefox.

    listing and xmp work the same as pre in all browsers. For Firefox and Opera, this results in trying to put a br inside an xmp, so I go with IE/Chrome for xmp.

    TODO: In cases where hitting enter in a header doesn't break out of the header, we should probably follow this code path too, instead of creating an adjoining header. No browser does this, though, so we don't.

    If container's local name is "address", "listing", or "pre":

    1. Let br be the result of calling createElement("br") on the context object.
    2. Call insertNode(br) on the active range.
    3. Call collapse(node, offset + 1) on the context object's selection.
    4. Necessary because adding a br to the end of a block element does nothing if there wasn't one there already. A single newline immediately preceding a block boundary does nothing.

      If br is the last descendant of container, let br be the result of calling createElement("br") on the context object, then call insertNode(br) on the active range.

    5. Return true.
  13. Including dt/dd here follows Firefox 5.0a2, as with the special dt/dd handling below.

    If container's local name is "li", "dt", or "dd"; and either it has no children or it has a single child and that child is a br:

    1. Split the parent of the one-node list consisting of container.
    2. If container has no children, call createElement("br") on the context object and append the result as the last child of container.
    3. Annoying hack to prevent the dl from being re-added when fixing disallowed ancestors. In most cases we want a wrapper dl added, but in two cases (delete and insertParagraph) we're actually trying to outdent the list item. TODO: there might be a better way to do this.

      If container is a dd or dt, and it is not an allowed child of any of its ancestors in the same editing host, set the tag name of container to the default single-line container name and let container be the result.

    4. Fix disallowed ancestors of container.
    5. Return true.
  14. Let new line range be a new range whose start is the same as the active range's, and whose end is (container, length of container).
  15. We don't want the start to be just inside a node, because if it is, we'll leave behind an empty element either in the new or old container. Empty block nodes are fine, and we'll add a br later, but empty inline nodes are bad, since the user can't interact with them.

    While new line range's start offset is zero and its start node is not a prohibited paragraph child, set its start to (parent of start node, index of start node).

  16. While new line range's start offset is the length of its start node and its start node is not a prohibited paragraph child, set its start to (parent of start node, 1 + index of start node).
  17. Let end of line be true if new line range contains either nothing or a single br, and false otherwise.
  18. IE9 makes a new header if there's a trailing <br>. Firefox 5.0a2, Chrome 13 dev, and Opera 11.10 do not, and I follow them, since it makes more sense (such a <br> is invisible).

    If the local name of container is "h1", "h2", "h3", "h4", "h5", or "h6", and end of line is true, let new container name be the default single-line container name.

  19. This step and the next follow Firefox 5.0a2. IE9 and Chrome 13 dev act as though these two lines were not present (they clone the existing element). Opera 11.10 nests a <p> inside. Firefox is the most useful, assuming a definition list somehow winds up inside the content (like via formatBlock).

    Otherwise, if the local name of container is "dt" and end of line is true, let new container name be "dd".

  20. Otherwise, if the local name of container is "dd" and end of line is true, let new container name be "dt".
  21. Otherwise, let new container name be the local name of container.
  22. Let new container be the result of calling createElement(new container name) on the context object.
  23. Copy all attributes of container to new container.
  24. If new container has an id attribute, unset it.
  25. Insert new container into the parent of container immediately after container.
  26. Let contained nodes be all nodes contained in new line range.
  27. TODO: This blows up any ranges (other than the selection, which we reset), and can alter non-editable nodes, and maybe other bad stuff. May or may not be the best solution. The intermediate fragment is also possibly black-box detectable by DOM mutation events, but I like to pretend those don't exist.

    Let frag be the result of calling extractContents() on new line range.

  28. Unset the id attribute (if any) of each Element descendant of frag that is not in contained nodes.
  29. Call appendChild(frag) on new container.
  30. Needed in case we have something like <ol><li><p>[]foo</ol>, which becomes <ol><li><p><li><p>foo</ol>. In this case we want to add the br to the p, not the li. Likewise for <ol><li><p>foo[]</ol>.

    While container's lastChild is a prohibited paragraph child, set container to its lastChild.

  31. While new container's lastChild is a prohibited paragraph child, set new container to its lastChild.
  32. If container has no visible children, call createElement("br") on the context object, and append the result as the last child of container.
  33. These two steps follow Firefox 5.0a2, Chrome 13 dev, and Opera 11.10. IE9 instead inserts an &nbsp; which magically does not appear in innerHTML. In all cases, the reason is that an empty block box in CSS will have zero height, so the user won't be able to put the selection cursor inside it.

    If new container has no visible children, call createElement("br") on the context object, and append the result as the last child of new container.

  34. Call collapse(new container, 0) on the context object's selection.
  35. Return true.

The insertText command

This is the same as typing text (see Additional requirements). If the input string is more than one character, first we split it up into one execution per character, for simplicity. The general rule is then that we record each command's state override or value override, insert the new character and select it, restore any overrides that we saved, and collapse the selection to its end.

The idea of the override business is that the user might run a command like bold when the selection is collapsed. There's nothing to bold in that case, but if the user runs bold and then types a character, they expect it to be bold. Thus we save the requested state in an override and restore it when the user types. Deleting things can also set overrides, if the deleted text was styled.

Whitespace is a special case. The note for canonical space sequences gives some of the background. If the user tries typing a space, we make it non-breaking so it doesn't collapse with anything, then canonicalize whitespace around the insertion point so line breaking isn't adversely affected.

One last special case is a newline. For that we just pass off to the insertParagraph command.

Supported only by WebKit. Tests in other browsers were manual.

TODO: This doesn't work well if the input contains things that aren't supposed to appear in HTML, like carriage returns or nulls. Nor is it going to work well if the current cursor position is in between two halves of a non-BMP character. This will result in unserializability. The current spec disregards this, as Chrome 14 dev does. (It's not relevant to other browsers, since they don't support this as a command.)

Important issue: non-breaking space fun! The issue: if the user hits space twice, they expect it to create two spaces, not collapse. Also, if they're at the beginning or end of a line and hit space, again, they expect it not to collapse. Since we don't want to require that all contenteditable element contents always be used only with white-space: pre-wrap, we need to convert to and from non-breaking spaces.

But there's a catch: you can't just make spaces non-breaking willy-nilly, because that doesn't just stop the space from collapsing, it also prevents breaking. (Chrome 14 dev actually cheats here: in contenteditable, it doesn't collapse nbsp, but breaks after it like a regular space.) The upshot of this is that any nbsp needs to be followed by a space, or else it might end up at the beginning of a line and be visible there; and it needs to be preceded by a space, or else it might break a line prematurely. How to achieve both of these goals when there are an even number of spaces to display is left as an exercise for the reader.

Browsers vary greatly in how they handle all this, of course!

The basic philosophy of IE9 is that if you're inserting a space, and one or both of the neighboring characters is a space, change the neighboring characters to non-breaking spaces. This breaks if one of the neighboring characters is part of a run of collapsed whitespace: "foo []bar" becomes "foo &nbsp; []bar", which converts one visible space to three.

Firefox 6.0a2 will sometimes convert the space you're inserting to an nbsp, sometimes convert neighboring spaces to nbsps, and sometimes convert neighboring nbsps to spaces. I cannot discern any clear reason to when it chooses what, except that it seems to prefer runs of nbsp's followed by a single space (although not always). I didn't find any outright bugs, except the inevitable ones like nbsp's sometimes being right after letters.

Chrome 14 dev tries to normalize everything to look like " &nbsp; &nbsp; ...", alternating with space then nbsp. Unfortunately, it does so buggily, because it converts collapsed spaces to nbsp's, so inserting a space before " " makes it into " &nbsp; &nbsp;", which changes one visible space to four (or arbitrarily many).

Opera 11.11 has varying behavior, like Firefox and Chrome. Like Firefox, I didn't discern an obvious pattern.

This was all discussed.

Unfortunately, we're stuck with this nbsp stuff, because of 1) legacy reasons, 2) mail clients might not support CSS equivalents, 3) authors might not know to apply any CSS to wherever the content is eventually used. The behavior I decided on to minimize the evil is as follows:

This avoids nbsp at the end of a run except where it's needed, so words won't appear indented if they wrap to the next line. It avoids more than two nbsp's in a row, so there won't be huge chunks of space that get wrapped all at once. And it avoids nbsp at the beginning of a run except where it's needed or if there are only two spaces in the run, so words won't have to wrap unnecessarily.

This is still a huge headache, though.

Action:

  1. Chrome 14 dev does the deletion even if the value is empty. Of course, other browsers don't expose this as an execCommand(), so no one else has any defined behavior in this case at all, so I follow Chrome.

    IE9, Firefox 7.0a2, Chrome 14 dev, and Opera 11.50 all don't strip wrappers, except that as usual, Gecko does if you select the whole wrapper, like {<b>foo</b>}. Also, Chrome 14 dev seems to strip the wrapper and try recreating the style in cases like <b>[foo</b>bar], where it starts in a wrapper but ends after it; this doesn't always work so well, so I don't do it. Firefox 7.0a2 also has the deletion set overrides for indeterminate state commands, so if you run insertText on [foo<b>bar</b>baz] it will make the result bold.

    These things don't make any sense to me, so I don't do them. I set overrides based on the first editable text node in the range when deleting; preserve any wrappers at the start of the range; and restore the overrides in case preserving the wrappers isn't enough (like if they weren't set by deletion at all). This behavior seems to closely match IE9.

    Delete the selection, with strip wrappers false.

  2. If the active range's start node is neither editable nor an editing host, return true.
  3. If value's length is greater than one:
    1. For each code unit el in value, take the action for the insertText command, with value equal to el.
    2. Return true.
  4. If value is the empty string, return true.
  5. TODO: WebKit also does magic for tabs, wrapping them in a whitespace-preserving span. Should we?

    If value is a newline (U+00A0), take the action for the insertParagraph command and return true.

  6. Let node and offset be the active range's start node and offset.
  7. Just to be tidy, add to an existing text node if there is one. Firefox 5.0a2 only adds to an existing one if the range is in a text node. IE9, Chrome 14 dev, and Opera 11.11 also add to an existing text node if the range is in an element adjacent to a text node. If there are two text nodes and it's in between, like foo{}bar, IE and Opera add to the first, Chrome adds to the second, although it probably doesn't matter in practice exactly which we choose.

    If node has a child whose index is offset − 1, and that child is a Text node, set node to that child, then set offset to node's length.

  8. If node has a child whose index is offset, and that child is a Text node, set node to that child, then set offset to zero.
  9. Record current overrides, and let overrides be the result.
  10. Call collapse(node, offset) on the context object's selection.
  11. Canonicalize whitespace at (node, offset).
  12. Let (node, offset) be the active range's start.
  13. If node is a Text node:
    1. Call insertData(offset, value) on node.
    2. Call collapse(node, offset) on the context object's selection.
    3. Call extend(node, offset + 1) on the context object's selection.
  14. Otherwise:
    1. If some text is inserted into <p><br></p> or similar, we no longer need the <br>.

      If node has only one child, which is a collapsed line break, remove its child from it.

    2. Let text be the result of calling createTextNode(value) on the context object.
    3. Call insertNode(text) on the active range.
    4. Call collapse(text, 0) on the context object's selection.
    5. Call extend(text, 1) on the context object's selection.
  15. Restore states and values from overrides.
  16. Canonicalize whitespace at the active range's start, with fix collapsed space false.
  17. Canonicalize whitespace at the active range's end, with fix collapsed space false.
  18. If value is a space character, autolink the active range's start.
  19. Call collapseToEnd() on the context object's selection.
  20. Return true.

The insertUnorderedList command

Preserves overrides

See comments for insertOrderedList.

Action: Toggle lists with tag name "ul", then return true.

Indeterminate: True if the selection's list state is "mixed" or "mixed ul", false otherwise.

State: True if the selection's list state is "ul", false otherwise.

The justifyCenter command

Preserves overrides

Action: Justify the selection with alignment "center", then return true.

This roughly matches Chrome 14 dev, although not exactly. Firefox 6.0a2 always returns false.

As a general rule, ignoring nodes with children saves us from treating <div align=left><div align=center>foo</div></div> as though it's indeterminate. Chrome 14 dev seems to only pay attention to text nodes, instead, or something like that. At any rate, it fails on images. Firefox 6.0a2 (for state and value) gets tripped up by examples like the one given.

If we ever support centering of tables and similar, we'd want to pay attention even to some nodes that do have children.

Indeterminate: Return false if the active range is null. Otherwise, block-extend the active range. Return true if among visible editable nodes that are contained in the result and have no children, at least one has alignment value "center" and at least one does not. Otherwise return false.

IE9 throws exceptions in almost every case when querying the state of justify*, and Opera 11.11 returns false in every case except some seemingly random crazy ones.

Firefox 6.0a2 returns true for the state of justify* if anything in the range has the right alignment, not if everything does. This isn't consistent with how state works for the inline commands, nor with WebKit.

Chrome 14 dev counts text-align on inline elements, which is wrong, because the property has no effect. It also counts it on non-editable elements, which is wrong, because then the state for justify* wouldn't necessarily be true after executing it. (Chrome actually does align the non-editable elements, but that's just a bug.) Chrome further returns false for justify* if the justification is just the default inherited justification, e.g., left for LTR. This doesn't seem to make sense either.

State is kind of redundant here, because it's true if and only if indeterminate is false and the value is equal to the desired value. However, I'll support it anyway, since Gecko/WebKit do.

State: Return false if the active range is null. Otherwise, block-extend the active range. Return true if there is at least one visible editable node that is contained in the result and has no children, and all such nodes have alignment value "center". Otherwise return false.

Not bidi-safe, but it's a pretty marginal corner case where it fails. Firefox 6.0a2 behaves weirdly here: it keys off the start node of the active range, even if that's not contained. Thus {<div align=center>foo</div>} has value "left" and indeterminate false, which would suggest that the whole selection is aligned left, but that's not the case. Chrome 14 dev returns the state cast to a string, as usual. Opera 11.11 always returns the empty string.

Value: Return the empty string if the active range is null. Otherwise, block-extend the active range, and return the alignment value of the first visible editable node that is contained in the result and has no children. If there is no such node, return "left".

The justifyFull command

Preserves overrides

Action: Justify the selection with alignment "justify", then return true.

Indeterminate: Return false if the active range is null. Otherwise, block-extend the active range. Return true if among visible editable nodes that are contained in the result and have no children, at least one has alignment value "justify" and at least one does not. Otherwise return false.

State: Return false if the active range is null. Otherwise, block-extend the active range. Return true if there is at least one visible editable node that is contained in the result and has no children, and all such nodes have alignment value "justify". Otherwise return false.

Value: Return the empty string if the active range is null. Otherwise, block-extend the active range, and return the alignment value of the first visible editable node that is contained in the result and has no children. If there is no such node, return "left".

The justifyLeft command

Preserves overrides

Action: Justify the selection with alignment "left", then return true.

Indeterminate: Return false if the active range is null. Otherwise, block-extend the active range. Return true if among visible editable nodes that are contained in the result and have no children, at least one has alignment value "left" and at least one does not. Otherwise return false.

State: Return false if the active range is null. Otherwise, block-extend the active range. Return true if there is at least one visible editable node that is contained in the result and has no children, and all such nodes have alignment value "left". Otherwise return false.

Value: Return the empty string if the active range is null. Otherwise, block-extend the active range, and return the alignment value of the first visible editable node that is contained in the result and has no children. If there is no such node, return "left".

The justifyRight command

Preserves overrides

Action: Justify the selection with alignment "right", then return true.

Indeterminate: Return false if the active range is null. Otherwise, block-extend the active range. Return true if among visible editable nodes that are contained in the result and have no children, at least one has alignment value "right" and at least one does not. Otherwise return false.

State: Return false if the active range is null. Otherwise, block-extend the active range. Return true if there is at least one visible editable node that is contained in the result and has no children, and all such nodes have alignment value "right". Otherwise return false.

Value: Return the empty string if the active range is null. Otherwise, block-extend the active range, and return the alignment value of the first visible editable node that is contained in the result and has no children. If there is no such node, return "left".

The outdent command

Preserves overrides

Action:

  1. Let items be a list of all lis that are inclusive ancestors of the active range's start and/or end node.
  2. TODO: This overnormalizes, but it seems like the simplest solution for now.

    For each item in items, normalize sublists of item.

  3. Block-extend the active range, and let new range be the result.
  4. Let node list be a list of nodes, initially empty.
  5. This step is kind of weird. For regular outdenting, we start at the inside and outdent going out, so that we remove the innermost indentation, on the theory that that will produce the cleanest markup (remove the most nodes). For lists, we remove the outermost indentation, because it makes a difference whether we remove inner or outer indentation, and logically we want to remove outer. E.g.,

    <ol><li>foo</li><ul><li>bar</li></ul></ol>

    should become

    foo<ul><li>bar</li></ul>

    not

    foo<ol><li>bar</li></ol>.

    But this is a bit weird and I'm wondering if it's really correct. TODO: Reexamine this.

    For each node node contained in new range, append node to node list if the last member of node list (if any) is not an ancestor of node; node is editable; and either node has no editable descendants, or is an ol or ul, or is an li whose parent is an ol or ul.

  6. While node list is not empty:
    1. While the first member of node list is an ol or ul or is not the child of an ol or ul, outdent it and remove it from node list.
    2. If node list is empty, break from these substeps.
    3. Let sublist be a list of nodes, initially empty.
    4. Remove the first member of node list and append it to sublist.
    5. While the first member of node list is the nextSibling of the last member of sublist, and the first member of node list is not an ol or ul, remove the first member of node list and append it to sublist.
    6. Record the values of sublist, and let values be the result.
    7. Split the parent of sublist.
    8. Fix disallowed ancestors of each member of sublist.
    9. Restore the values from values.
  7. Return true.

Miscellaneous commands

The copy command

IE9 supports copy/cut/paste with a security warning. Firefox reportedly only supports it if you set a pref. I didn't find info on other browsers, but in my tests it didn't do anything. I'm not going to try speccing it unless implementers are interested in working out the security problems and trying to get interop. It seems like as of June 2011, everyone just uses Flash for this. So it would be nice if we could work out a more secure standardized substitute.

Action: The user agent must either copy the current selection to the clipboard as though the user had requested it, or throw a SecurityError exception. This specification does not define exactly how the selection is to be copied to the clipboard, but the Clipboard API and events specification might be useful.

User agents should exercise caution in respecting this command, because sites could abuse it to confuse and annoy the user by overwriting the clipboard with extremely long, obscene, or otherwise objectionable content.

The idea is sites might catch the SECURITY_ERR and treat it differently from NOT_SUPPORTED_ERR, like encouraging users to reconfigure their browser. However, browsers might not want to encourage authors to tell users to reconfigure their browser insecurely.

User agents may choose not to support this command at all. If a user agent will only honor the command for some whitelisted sites depending on configuration, it may either throw a SecurityError exception for non-whitelisted sites, or it may act as though the command is unsupported on those sites.

The cut command

See comment for copy.

Action: The user agent must either copy the current selection to the clipboard and then delete it, as though the user had requested it, or throw a SecurityError exception. This specification does not define exactly how the selection is to be deleted or copied to the clipboard, but the Clipboard API and events specification might be useful.

User agents should exercise caution in respecting this command, because sites could abuse it to confuse and annoy the user by overwriting the clipboard with extremely long, obscene, or otherwise objectionable content.

User agents may choose not to support this command at all. If a user agent will only honor the command for some whitelisted sites depending on configuration, it may either throw a SecurityError exception for non-whitelisted sites, or it may act as though the command is unsupported on those sites.

The defaultParagraphSeparator command

This is a new feature, added by request in bug 15527. (Opera already had an o-defaultblock command that worked similarly.) If you are implementing this, please make sure to file any feedback as bugs. The spec is not finalized yet and can still be easily changed.

Action: Let value be converted to ASCII lowercase. If value is then equal to "p" or "div", set the context object's default single-line container name to value, then return true. Otherwise, return false.

Value: Return the context object's default single-line container name.

The paste command

See comment for copy.

Action: The user agent must either delete the selection and then paste the clipboard's contents to the current cursor position, as though the user had requested it, or throw a SecurityError exception. This specification does not define exactly how the clipboard is to be converted to HTML for pasting, but the Clipboard API and events specification might be useful.

User agents should exercise caution in respecting this command, because sites could abuse it to read private information from the clipboard.

User agents may choose not to support this command at all. If a user agent will only honor the command for some whitelisted sites depending on configuration, it may either throw a SecurityError exception for non-whitelisted sites, or it may act as though the command is unsupported on those sites.

The redo command

Action: As defined by the UndoManager specification.

The selectAll command

This is totally broken: if executed inside an editing host, it has to select the editing host contents, not the whole document. See bug.

Tested using roughly this.

IE9
A bit confusing. The gist seems to be that it does selectAllChildren() on the body, except sometimes it doesn't.
Firefox 5.0a2
Throws an exception if nothing in the document is editable, which apparently it always does for execCommand(). If there's a body, it does selectAllChildren() on that, and otherwise it does selectAllChildren() on the root element. If there's no root element, throws an exception.
Chrome 13 dev
If there's no root element, removes the selection. If there's a root element but no body, collapses the selection at (document, 0). If there's a body, it selects all the contents of the body, although that doesn't mean the resulting anchor or focus actually are the body node (they're usually text nodes). But it seems to *avoid* selecting contenteditable stuff: if all the visible things in the body are contenteditable, it removes all ranges from the selection, and if some are, it freaks out and behaves oddly. But designMode doesn't trouble it.
Opera 11.11
Was characteristically uncooperative in my tests, and I didn't try to investigate further.

The behavior here is relatively simple and largely matches implementations.

Action:

  1. TODO: Is this right even for framesets?

    Let target be the body element of the context object.

    TODO: Is this right even for documents whose root element is not an HTML element?

    If target is null, let target be the context object's documentElement.

  2. If target is null, call getSelection() on the context object, and call removeAllRanges() on the result.
  3. Otherwise, call getSelection() on the context object, and call selectAllChildren(target) on the result.
  4. Return true.

The styleWithCSS command

IE9 and Opera 11.00 don't support this command. By and large, they act the way Gecko and WebKit do when styleWithCSS is off. Gecko invented it, and WebKit also supports it. The default in Firefox 4.0 is off, while all other browsers behave like the default is on (and IE/Opera give no way to turn it off), so I default it to on.

Handling of value matches Firefox 5.0a2. Chrome 13 dev treats the case-sensitive string "true" as true, the case-sensitive string "false" as false, and does nothing for any other string. I went with Gecko because this way there are only two possible effects, not three, which makes it easier to reason about and debug. Also, Gecko made up the command, so this is probably more web-compatible. Cursory searches of Google Code and GitHub suggest that authors almost always pass a boolean as the third argument when using styleWithCSS, in which case the two behaviors work the same.

Action: If value is an ASCII case-insensitive match for the string "false", set the CSS styling flag to false. Otherwise, set the CSS styling flag to true. Either way, return true.

This follows Chrome 13 dev. Firefox 5.0a2 doesn't support queryCommandState() for styleWithCSS.

State: True if the CSS styling flag is true, otherwise false.

The undo command

Action: As defined by the UndoManager specification.

The useCSS command

Supported by Firefox 4.0, but not IE9 or Opera 11.00 (which don't support styleWithCSS either), nor by Chrome 12 dev (which does support styleWithCSS). useCSS was the original feature in Mozilla 1.3, but the meaning is backward, so Gecko added styleWithCSS as a replacement. No state is defined, since only Gecko supports useCSS at all, and as of Firefox 6.0a2, it doesn't support queryCommandState() for it.

As of September 2011, WebKit is adding support for this command, so it can no longer be considered deprecated.

Action: If value is an ASCII case-insensitive match for the string "false", set the CSS styling flag to true. Otherwise, set the CSS styling flag to false. Either way, return true.

Since the effect of this command is the opposite of what one would expect, user agents are encouraged to point authors to styleWithCSS when useCSS is used, such as by logging a warning to an error console.

No state is defined for this command, so queryCommandState("useCSS") must always return false. To query the CSS styling flag's current value, authors have to use the styleWithCSS command.

Additional requirements

It has been suggested that some things here need to be platform-dependent, not fully standardized. For now I'm standardizing them anyway, because the large majority of behavior should be platform-agnostic. If anyone has suggestions as to particular things that should be left up to platform behavior, please say so.

When the user instructs the user agent to insert a line break inside an editing host, such as by pressing the Enter key while the cursor is in an editable node, the user agent must call execCommand("insertparagraph") on the relevant document.

When the user instructs the user agent to insert a line break inside an editing host without breaking out of the current block, such as by pressing Shift-Enter or Option-Enter while the cursor is in an editable node, the user agent must call execCommand("insertlinebreak") on the relevant document.

When the user instructs the user agent to delete the previous character inside an editing host, such as by pressing the Backspace key while the cursor is in an editable node, the user agent must call execCommand("delete") on the relevant document.

When the user instructs the user agent to delete the next character inside an editing host, such as by pressing the Delete key while the cursor is in an editable node, the user agent must call execCommand("forwarddelete") on the relevant document.

When the user instructs the user agent to insert text inside an editing host, such as by typing on the keyboard while the cursor is in an editable node, the user agent must call execCommand("inserttext", false, value) on the relevant document, with value equal to the text the user provided. If the user inserts multiple characters at once or in quick succession, this specification does not define whether it is treated as one insertion or several consecutive insertions.

Bug 13938.

The user agent may allow the user to make other changes to editable content, such as causing Ctrl-B to call execCommand("bold"). Any such change must be accomplished by calling execCommand(), so that in particular, "beforeinput" and "input" events fire as appropriate. The command invoked should be one of the standard commands defined in this specification if possible, and otherwise must be a vendor-specific command.

Acknowledgements

This section is not normative.

Thanks to: