tree-sitter-qmd - allow multi-line attribute lists on inline images/spans

[rundel]

Not sure if this is even something you want to support or not - if the latter feel free to close this. Otherwise this is an attempt at a semi-minimally invasive fix for multi-line attributes.

Multi-line {...} attribute lists on inline images, spans, and other inline constructs that share _pandoc_attr_specifier were rejected at the first attribute. The same form rendered fine in TS Quarto and was common in quarto-web sources, leaving q2 as the lone holdout.

Root cause was twofold in tree-sitter-markdown/grammar.js:

• attribute_specifier / _pandoc_attr_specifier had no tolerance for whitespace between { and the first specifier (or between _commonmark_specifier_start_with_class and } — also broke the single-line { .cls } form).
• _inline_whitespace is choice($._whitespace, $._soft_line_break), a single token, so a \n followed by indent on the next line couldn't match as one inter-attribute separator.

Introduce _attr_ws: prec(-1, repeat1(choice($._whitespace, $._soft_line_break))) and use it where attribute lists need to span lines:

• optional($._attr_ws) immediately after { in both attribute_specifier and _pandoc_attr_specifier.
• Add trailing optional($._attr_ws) inside _commonmark_specifier_start_with_class (already present on _commonmark_specifier_start_with_kv). Keeping the trailer inside the specifier (rather than around the choice at the wrapper) avoids an LR conflict with language_specifier.
• Swap inner _inline_whitespace → _attr_ws at the four sites that join successive classes, classes-to-kv, and successive kvs.

commonmark_specifier's leading optional($._inline_whitespace) is left untouched — changing it triggered the language_specifier conflict, and the wrapper-level _attr_ws already absorbs anything that would have leaked through.

Regenerated artifacts (parser.c, grammar.json, node-types.json) are produced by tree-sitter generate; do not hand-edit. Also regenerated crates/pampa/resources/error-corpus/_autogen-table.json via deno run -A scripts/build_error_table.ts because parser-state IDs shifted — without this regen, 7 apostrophe-quotes tests in qmd-syntax-helper fail (the rule keys off Q-2-10 diagnostics which map through (lr_state, sym) pairs).

Tests:

• 6 new tree-sitter corpus cases in inline-multiline-attrs.txt cover multi-line class-only, multi-line class + key=value, { .cls } symmetry, span multi-line class, span multi-line kv, and # Heading { .cls } for the block-level path.
• 1 new pampa integration test in test_attr_source_parsing.rs verifies that the AST and attr_source byte offsets are correct for the quarto-web-style multi-line form.

Full tree-sitter corpus: 492/492 pass (was 486). Full workspace: 8943/8943 pass. cargo xtask verify Rust legs all green.

End-to-end check:

$ pampa repro.qmd
[ Header 1 ( "test" , [] , [] ) [Str "Test"]
, Para [Image ( "" , ["hero-banner", "img-fluid"]
, [("fig-align", "center"), ("width", "600px")] )
[] ("featured.png" , "")]
, Para [Str "Done."] ]

[cscheid]

I'm not against this fix, but anything involving line breaks in Markdown gives me a case of the heebie jeebies. I think it's prudent to add tests that exercise those spans inside bulleted lists and block quotes before declaring success.

(Operational annoyance: every time a grammar fix makes it in, we need to regenerate the error corpus table and parser.c autogen file. There's instructions for Claude to do it in the repo but you'll need to install Deno.)

[rundel]

Good call - bulleted lists look like they work fine and block quotes are a nightmare. For now I punted with it just throwing an error on multiline attributes w/in a block quote

I'm not sure if the Q-2-2 upgrade to Q-2-38 behavior based on context is acceptable or not.

Claude's summary is below.

---

Builds on the earlier multi-line attribute-list fix by adding edge-case coverage and a new diagnostic for the one shape the grammar fix can't reach.

List-context tests

Multi-line {...} attribute lists already work inside list items thanks to list-continuation handling — these tests lock that in against regression.

Tree-sitter corpus (inline-multiline-attrs.txt, 6 new cases):

• Bulleted list item with multi-line span attrs.
• Ordered list item with multi-line span attrs.
• Nested bulleted list with multi-line span attrs in the inner item.
• Bulleted list item with multi-line image attrs (class + key=value).
• Top-level bullets with sibling items before and after the multi-line attrs at the same indent.
• Nested bullets with sibling inner items before and after at the same indent.

Pampa integration tests (test_attr_source_parsing.rs, 5 new cases): AST-level mirrors of the above, verifying classes, key/value pairs, and attr_source byte offsets survive the block→inline boundary for each shape.

Q-2-38 — blockquote-specific error

Multi-line attribute lists inside a > blockquote don't work: the tree-sitter external scanner short-circuits SOFT_LINE_ENDING when the next line begins with > (scanner.c:2380-2407), so the inline parser sees only the first physical line and the same (state=2587, sym="_close_block") fires as a top-level unclosed { — indistinguishable at the lookup level, but distinguishable in the source text.

qmd_error_messages.rs::upgrade_q22_to_q238_if_in_blockquote post-processes each Q-2-2 diagnostic: if the error line begins with > (after optional leading whitespace), the diagnostic is rewritten to Q-2-38:

Error: [Q-2-38] Multi-line inline attribute list inside blockquote
 1 │ > ![](img.png){
   │                ╰── Inside a blockquote, an inline `{...}` attribute list cannot span multiple lines.
ℹ Put the attribute list on a single line, or move this construct out of the blockquote.

• resources/error-corpus/Q-2-38.json documents the new code (cases: [] — manually emitted, not state-mapped).
• tests/test_q_2_38_blockquote_multiline_attrs.rs: 4 tests — image-in-bq, span-in-bq, blockquote with leading indent, and a top-level negative control confirming [attr]{[ still maps to Q-2-2.

[cscheid]

I think I want to leave this as an open issue until we can handle it uniformly. Creating these syntax exceptions is sort of opening the door for future trouble. I'd rather us reject those attributes uniformly even if it's annoying.