Djot

Djot is a light markup syntax. It derives most of its features from commonmark, but it fixes a few things that make commonmark's syntax complex and difficult to parse efficiently. It is also much fuller-featured than commonmark, with support for definition lists, footnotes, tables, several new kinds of inline formatting (insert, delete, highlight, superscript, subscript), math, smart punctuation, attributes that can be applied to any element, and generic containers for block-level, inline-level, and raw content.

The project began as an attempt to implement some of the ideas I suggested in my essay Beyond Markdown. (See Rationale, below.)

This repository contains a reference implementation, written in Lua, and a Syntax Description. There is also a Cheatsheet and a Quick Start for Markdown Users that outlines the main differences between djot and Markdown.

Despite being written in an interpreted language, the reference implementation is very fast (converting a 260K test document in 141 ms on an M1 mac using the standard lua interpreter). It can produce an AST, rendered HTML, or a stream of match tokens that identify elements by source position, which could be used for syntax highlighting or a linting tool.

We also provide a custom pandoc writer for djot (djot-writer.lua), so that documents in other formats can be converted to djot format, and a custom pandoc reader (djot-reader.lua), so that djot documents can be converted to any format pandoc supports. To use these, just put them in your working directory and use pandoc -f djot-reader.lua to convert from djot, and pandoc -t djot-writer.lua to convert to djot. (You'll need pandoc version 2.18 or higher.)

Rationale

Here are some design goals:

It should be possible to parse djot markup in linear time, with no backtracking.
Parsing of inline elements should be "local" and not depend on what references are defined later. This is not the case in commonmark: [foo][bar] might be "[foo]" followed by a link with text "bar", or "[foo][bar]", or a link with text "foo", or a link with text "foo" followed by "[bar]", depending on whether the references [foo] and [bar] are defined elsewhere (perhaps later) in the document. This non-locality makes accurate syntax highlighting nearly impossible.
Rules for emphasis should be simpler. The fact that doubled characters are used for strong emphasis in commonmark leads to many potential ambiguities, which are resolved by a daunting list of 17 rules. It is hard to form a good mental model of these rules. Most of the time they interpret things the way a human would most naturally interpret them---but not always.
Expressive blind spots should be avoided. In commonmark, you're out of luck if you want to produce the HTML a<emph>?</emph>b, because the flanking rules classify the first asterisk in a*?*b as right-flanking. There is a way around this, but it's ugly (using a numerical entity instead of a). In djot there should not be expressive blind spots of this kind.
Rules for what content belongs to a list item should be simple. In commonmark, content under a list item must be indented as far as the first non-space content after the list marker (or five spaces after the marker, in case the list item begins with indented code). Many people get confused when their indented content is not indented far enough and does not get included in the list item.
Parsers should not be forced to recognize unicode character classes, HTML tags, or entities, or perform unicode case folding. That adds a lot of complexity.
The syntax should be friendly to hard-wrapping: hard-wrapping a paragraph should not lead to different interpretations, e.g. when a number followed by a period ends up at the beginning of a line. (I anticipate that many will ask, why hard-wrap at all? Answer: so that your document is readable just as it is, without conversion to HTML and without special editor modes that soft-wrap long lines. Remember that source readability was one of the prime goals of Markdown and Commonmark.)
The syntax should compose uniformly, in the following sense: if a sequence of lines has a certain meaning outside a list item or block quote, it should have the same meaning inside it. This principle is articulated in the commonmark spec, but the spec doesn't completely abide by it (see commonmark/commonmark-spec#634).
It should be possible to attach arbitrary attributes to any element.
There should be generic containers for text, inline content, and block-level content, to which arbitrary attributes can be applied. This allows for extensibility using AST transformations.
The syntax should be kept as simple as possible, consistent with these goals. Thus, for example, we don't need two different styles of headings or code blocks.

These goals motivated the following decisions:

Block-level elements can't interrupt paragraphs (or headings), because of goal 7. So in djot the following is a single paragraph, not (as commonmark sees it) a paragraph followed by an ordered list followed by a block quote followed by a section heading:
```
My favorite number is probably the number
1. It's the smallest natural number that is
> 0. With pencils, though, I prefer a
# 2.
```
Commonmark does make some concessions to goal 7, by forbidding lists beginning with markers other than 1. to interrupt paragraphs. But this is a compromise and a sacrifice of regularity and predictability in the syntax. Better just to have a general rule.
An implication of the last decision is that, although "tight" lists are still possible (without blank lines between items), a sublist must always be preceded by a blank line. Thus, instead of
```
- Fruits
  - apple
  - orange
```
you must write
```
- Fruits

  - apple
  - orange
```
(This blank line doesn't count against "tightness.") reStructuredText makes the same design decision.
Also to promote goal 7, we allow headings to "lazily" span multiple lines:
```
## My excessively long section heading is too
long to fit on one line.
```
While we're at it, we'll simplify by removing setext-style (underlined) headings. We don't really need two heading syntaxes (goal 11).
To meet goal 5, we have a very simple rule: anything that is indented beyond the start of the list marker belongs in the list item.
```
1. list item

  > block quote inside item 1

2. second item
```
In commonmark, this would be parsed as two separate lists with a block quote between them, because the block quote is not indented far enough. What kept us from using this simple rule in commonmark was indented code blocks. If list items are going to contain an indented code block, we need to know at what column to start counting the indentation, so we fixed on the column that makes the list look best (the first column of non-space content after the marker):
```
1.  A commonmark list item with an indented code block in it.

        code!
```
In djot, we just get rid of indented code blocks. Most people prefer fenced code blocks anyway, and we don't need two different ways of writing code blocks (goal 11).
To meet goal 6 and to avoid the complex rules commonmark adopted for handling raw HTML, we simply do not allow raw HTML, except in explicitly marked contexts, e.g. `<a id="foo">`{=html} or
```
``` =html
<table>
<tr><td>foo</td></tr>
</table>
```
```
Unlike Markdown, djot is not HTML-centric. Djot documents might be rendered to a variety of different formats, so although we want to provide the flexibility to include raw content in any output format, there is no reason to privilege HTML. For similar reasons we do not interpret HTML entities, as commonmark does.
To meet goal 2, we make reference link parsing local. Anything that looks like [foo][bar] or [foo][] gets treated as a reference link, regardless of whether [foo] is defined later in the document. A corollary is that we must get rid of shortcut link syntax, with just a single bracket pair, [like this]. It must always be clear what is a link without needing to know the surrounding context.
In support of goal 6, reference links are no longer case-insensitive. Supporting this beyond an ASCII context would require building in unicode case folding to every implementation, and it doesn't seem necessary.
A space or newline is required after > in block quotes, to avoid the violations of the principle of uniformity noted in goal 8:
```
>This is not a
>block quote in djot.
```
To meet goal 3, we avoid using doubled characters for strong emphasis. Instead, we use _ for emphasis and * for strong emphasis. Emphasis can begin with one of these characters, as long as it is not followed by a space, and will end when a similar character is encountered, as long as it is not preceded by a space and some different characters have occurred in between. In the case of overlap, the first one to be closed takes precedence. (This simple rule also avoids the need we had in commonmark to determine unicode character classes---goal 6.)
Taken just by itself, this last change would introduce a number of expressive blind spots. For example, given the simple rule,
```
_(_foo_)_
```
parses as
```
<em>(</em>foo<em>)</em>
```
rather than
```
<em>(<em>foo</em>)</em>
```
If you want the latter interpretation, djot allows you to use the syntax
```
_({_foo_})_
```
The {_ is a _ that can only open emphasis, and the _} is a _ that can only close emphasis. The same can be done with * or any other inline formatting marker that is ambiguous between an opener and closer. These curly braces are required for certain inline markup, e.g. {=highlighting=}, {+insert+}, and {-delete-}, since the characters =, +, and - are found often in ordinary text.
In support of goal 1, code span parsing does not backtrack. So if you open a code span and don't close it, it extends to the end of the paragraph. That is similar to the way fenced code blocks work in commonmark.
```
This is `inline code.
```
In support of goal 9, a generic attribute syntax is introduced. Attributes can be attached to any block-level element by putting them on the line before it, and to any inline-level element by putting them directly after it.
```
{#introduction}
This is the introductory paragraph, with
an identifier `introduction`.

                           {.important color="blue" #heading}
## heading

The word *atelier*{weight="600"} is French.
```
Since we are going to have generic attributes, we no longer support quoted titles in links. One can add a title attribute if needed, but this isn't very common, so we don't need a special syntax for it:
```
[Link text](url){title="Click me!"}
```
Fenced divs and bracketed spans are introduced in order to allow attributes to be attached to arbitrary sequences of block-level or inline-level elements. For example,
```
{#warning .sidebar}
::: Warning
This is a warning.
Here is a word in [français]{lang=fr}.
:::
```

Syntax

For a full syntax reference, see the syntax description.

A vim syntax highlighting definition for djot is provided in editors/vim/.

Playground

You can play with djot on the live djot playground, designed by @dtinth.

Installing

To install djot using luarocks, just

luarocks install djot

This will install both the library and the executable djot.

Using the Lua library

Quick start

If you just want to parse some input and produce HTML:

local djot = require("djot")
local input = "This is *djot*"
local parser = djot.Parser:new(input)
parser:parse()
local html = parser:render_html()

In more depth

local djot = require("djot")

-- Create a parser with input string and options:
local parser = djot.Parser:new(input, options)
-- Parse the document:
parser:parse()

-- 'options' may be omitted, but if present it should be
-- a table with the following boolean fields (both defaulting
-- to false):
--
-- verbose: if true, warnings are emitted to stderr.
-- sourcepos: if true, source positions are tracked.

-- At this point we have a list of "matches"
-- (including start/end -- position and an annotation like '+blockquote').
local matches = parser:get_matches()

-- 'matches' is an array of match objects; to deal with them, you'll need
local match = require("djot.match")
local startpos, endpos, annotation = match.unpack_match(matches[1])

-- Here 'startpos' and 'endpos' are integer byte positions in the
-- input string, and 'annotation' is a string starting with '+'
-- or '-', e.g. '+blockquote'.

-- You can use 'matches_pattern' to check the match's annotation
-- against a Lua pattern without unpacking it:
if match.matches_pattern(matches[1], 'blockquote') then
  -- second argument is a Lua pattern
  print ("It's either +blockquote or -blockquote!")
end

-- To render the stream of matches in a human-readable form:
parser:render_matches(io.stdout)
-- or as a string
local matches_string = parser:render_matches()

-- After parsing, warnings can also be obtained as a Lua table:
local warnings = parser.warnings

-- To build an AST from the list of matches:
parser:build_ast()

-- The AST is now available as a Lua table, parser.ast
-- It can be modified programatically and any changes
-- you make will be reflected in rendered HTML.

-- To render the AST in human-readable form:
parser:render_ast(io.stdout)
-- or as a string:
local ast_string = parser:render_ast()
-- This function will call 'build_ast' automatically if the
-- AST is not built.

-- To render the AST as an HTML string:
local html = parser:render_html()
-- Or, to send the HTML output to a handle:
parser:render_html(io.stdout)
-- This function will call 'build_ast' automatically if the
-- AST is not built.

The code

The code for djot (excluding the test suite) is standard Lua, compatible with lua 5.1--5.4 and with luajit. It has no external dependencies.

make install will build the rockspec and install the library and executable using luarocks. Once installed, the library can be used by Lua programs, and the executable can be run using djot. djot -h will give help output.

If you can't assume that lua or luajit will be installed on the target machine, you can use make djotbin to create a portable binary that includes luajit and the necessary scripts. (This assumes you have luastatic installed, and libluajit in the /usr/local tree.)

make test will run the tests, and make testall will also run some tests of pathological cases.

License

The code and documentation are released under the MIT license.

matklad / djot