[PhilippeCollin]

What are you adding in this PR?

This PR makes liquid-html-parser about 2.6× faster on real themes, and brings it closer to how Liquid and browsers read a template.

Shopify keeps a second copy of the parser in another codebase, which picked up both the speedups and a set of fixes. This PR ports them, adapted to this repo's AST, specs and consumers (prettier, theme-check, the language server). Each commit is one change and reads on its own. The changesets describe each change for the changelog.

Faster parsing

The first five commits are #1329, #1333, #1330 and #1331, adapted to the current main, where the tokenizer loop has since moved into createTokenizer. Later commits build on them. Those PRs have the full profiling, the per-change benchmarks and the differential checks. They can close if this one lands, or land first and this branch drops them.

Commits	Change	PR
34dae62, f11f802	Tokenize plain text runs in one step	#1329
5b634d9	Check character codes before regexes in the tokenizers	#1333
320fa1f	Trim text around Liquid tags in linear time	#1330
a73b174	Parse Object.prototype names as ordinary names	#1331

Measured on this branch with Apple M3 Pro and Node 24.22. The input is the parser's 429 fixture files (Dawn, Horizon and the base theme, 3.4 MB). Builds ran in alternating processes, 5 rounds × 30 passes. Each value is the median.

All 429 files	main	after a73b174	this PR	Speedup vs main
toLiquidHtmlAST	112.7 ms	37.2 ms	43.2 ms	2.6×
toLiquidAST	113.3 ms	36.2 ms	42.2 ms	2.7×
toTolerantLiquidHtmlAST (prettier's parse step)	112.2 ms	36.6 ms	43.5 ms	2.6×

The later ports cost about 6 ms of that back. A per-commit run puts about 4 ms on 56e78a4, which adds the tokenizer's HTML comment mode and makes raw tag bodies opaque to it. The other commits are within noise. Every build parses all 429 files without an error.

Liquid and HTML fixes, and who wrote them

Commits	Change	Original work by
8445f5a, df9074a, e0a1152, 0d65a38	Legacy {% render 'x' with key: value %} and {% include %} arguments	Alok Swamy
c57e318	reversed anywhere among for and tablerow arguments	Guilherme Carreiro
5492ff9, 4ce802f	One BlockArgument type for {% block %} and {% section %} arguments	Natacha Levesque, Charles-P. Clermont
f2cfbee	Tests for parse error locations at the end of input	Iain Wilson
4e85163, 7244211	Each raw tag ends its body by the rule of the Liquid class that parses it	River, with Martin Morissette and Mbarak Bujra
7368046, 11cc0f9	pairDanglingCloseTags for elements opened in each branch of a conditional	River
c80848c	Tests that block and section markup keep the full source	Claudéric Demers
b7c9102	<script> and <style> bodies are raw text	River, with Martin Morissette
56e78a4	An HTML comment mode in the tokenizer	River, with Mbarak Bujra
dcd9e3f	Parse the Liquid inside HTML comments	River, with Martin Morissette and Mbarak Bujra

Behaviour changes worth a look

• Liquid in HTML comments is parsed. HtmlComment has children and blockStartPosition/blockEndPosition. A block can open on one side of a comment delimiter and close on the other, so {% if a %}<!-- {% endif %} --> parses. A comment that crosses a block boundary is split into several HtmlComment nodes, and a comment that never closes ends at the end of the file. opaqueHtmlComments: true keeps the old opaque body.
• Checks see the Liquid in comments. An end tag in a comment that closes no block is now a syntax error, and checks such as UndefinedObject and MissingTemplate report Liquid in comments. Liquid renders that Liquid, so this matches the output, but themes with broken Liquid commented out will get new offenses.
• HTML comments tokenize as in a browser. Only Liquid and the close count inside a comment. --!> closes one, <!--> and <!---> are empty comments, and --> outside a comment is text.
• <script> and <style> bodies are raw text. A < in them no longer opens an element, and a </script> inside Liquid or a {% comment %} doesn't close the script. <script>a<script>b</script>c</script> now fails with "Attempting to close HtmlElement 'script' before it was opened", so prettier won't format such a file and theme-check reports it. Nested <svg> still pairs.
• Raw tags close where Liquid closes them. {% endcomment xyz %}, {% endraw xyz %} and {% enddoc foo %} close their tags. A nested {% doc %} is rejected, and {% style %} nests. A comment whose body Liquid's tokenizer runs past the {% endcomment %} now fails, as it does in Liquid. The same rules apply inside {% liquid %}.
• Error reports moved for two raw tag cases. In liquid-html-syntax-error.yml, LiquidSyntaxError now reports the later unclosed tag, or the comment's open tag, with the parser's message (7244211).
• {% for x in c limit: 5 reversed %} parses instead of falling back to string markup.
• BlockArgument replaces the LiquidNamedArgument | BlockArrayArgument union for block and section arguments. BlockArrayArgument is still exported, as a subtype.
• pairDanglingCloseTags is opt-in and off by default.
• Prettier prints a split or unclosed comment as written. A split piece that holds only whitespace prints as an empty line.
• The language server completes inside HTML comments. After an error inside a comment, tolerant recovery resumes at the comment's -->, so the markup after it stays HTML.

Open questions

1. A nested <script> is now a parse error, so prettier refuses the file. Is that acceptable, or should the parser keep pairing nested scripts?
2. The new offenses for Liquid in HTML comments: should any check skip comments? Liquid renders that Liquid, so I lean no.

Not ported

• The other copy ends an <svg> at its first </svg>. This tokenizer counts nested SVGs, and I kept that.
• Hybrid tags: this repo removed them on purpose in af297ba, so section stays a plain tag.
• Empty filter argument lists ({{ 'key' | t: }}) and whitespace around = in attributes: this parser already handles both.
• The end-of-input error fix: already fixed here, so only the tests came over.
• Dead code removal, benchmarks, an unused isBareBlockArgumentName helper, and test helpers whose tests are specs here.

What's next? Any followup issues?

Decide the open questions above. If #1329, #1330, #1331 and #1333 merge first, rebase this branch to drop them.

Tophatting

• pnpm test: 8110 passed, 5 skipped, 0 failed. pnpm type-check is clean.
• Parser specs were re-recorded with UPDATE_SPECS=1 where the AST changed. New cases in specs/parser/html.yml cover split, unclosed and reopened comments.
• The new prettier fixtures for split and unclosed comments pass with TEST_IDEMPOTENCE=true.

Before you deploy

• I included a minor bump changeset (liquid-html-parser gains HtmlComment.children, BlockArgument and pairDanglingCloseTags)
• I included patch bump changesets for the fixes and the consumers
• My feature is backward compatible: no. HtmlComment gains children, and the stricter parsing above rejects some templates it used to accept.

🤖 Generated with Claude Code