fix(rich-md-editor): raw-HTML-block fragmentation + styling follow-up#5440
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub. |
PR SummaryMedium Risk Overview UI: raw HTML/footnote snippets drop the warning-tinted background in favor of the same neutral surface as inline Tests add structural assertions ( Reviewed by Cursor Bugbot for commit c043d9c. Configure here. |
Greptile SummaryThis PR fixes raw HTML handling in the rich markdown editor. The main changes are:
Confidence Score: 5/5This looks safe to merge.
Important Files Changed
Reviews (5): Last reviewed commit: "fix(rich-md-editor): mask HTML comments ..." | Re-trigger Greptile |
0269c2a to
64c7b40
Compare
|
@cursor review |
- RawHtmlBlock previously delegated to marked's built-in block-HTML tokenizer, which (per CommonMark's HTML-block-type-6 rule) stops at the first blank line. A real <details> with a paragraph inside — the common case — fragmented into a raw chip, an ordinary rendered paragraph, and a second raw chip, stranding real content in between. - Fixed with a custom block-level tokenizer that scans to the tag's matching close (reusing the balanced open/close depth-tracking already built for nested inline HTML), spanning blank lines, restricted to CommonMark's own block-tag whitelist (details, div, table, section, etc.) so tags that can legitimately start a paragraph (em, a, span, code, kbd) are left untouched. - Dropped the warning-colored tint on raw HTML/footnote blocks (color-mix with --warning read as an error state) in favor of the same neutral surface as existing code blocks — the hover badge alone now signals "this is raw, unrendered text".
…ention edge cases - CommonMark allows up to 3 leading spaces before a block-HTML opening line; the new block tokenizer required column 0, so an indented <details>/<div> still fragmented across blank lines. Fixed by splitting off the leading indent, matching against the rest, and stitching the indent back onto raw. - The open-tag and balanced-close regexes stopped at the first `>`, even inside a quoted attribute value (e.g. data-example="a > b"), producing wrong match lengths and miscounting a same-tag mention inside the quoted value as a real nested tag. Replaced with an attribute-aware pattern that treats a full quoted value (including any interior >) as one unit. - The balanced-tag scan couldn't distinguish real markup from a tag name mentioned inside inline code or a fenced code block. Now masks code regions (same-length filler, positions preserved) before scanning, so a properly-escaped mention (backticks or a fenced example) is never miscounted. A genuinely bare, unescaped mention outside code remains a known, inherent limitation of regex-based tag matching (shared by real HTML parsers given the same ambiguous input) — verified this can never lose data or hang, only reflow to a stable fixpoint on save.
…ce scan - maskCodeRegions's fenced-code regex required fence markers at column 0, so a fence quoted inside a markdown blockquote (each line prefixed with `> `) wasn't recognized, leaving tag mentions inside it visible to the balance scanner. Extended the fence pattern to tolerate an optional blockquote prefix on both the opening and closing fence line. - BLOCK_HTML_TAG_NAMES includes several void elements (link, meta, base, hr, ...) that have no closing tag at all. The block tokenizer only treated an explicit self-closing `/>` as complete, so a void tag without one would scan the rest of the document for a `</meta>` that will never appear, risking a false match on a later same-name mention. Now reuses the existing VOID_TAGS set (already used by the inline tokenizer) to treat these as complete right after the open tag. Also restored `hr` to the whitelist, which was accidentally dropped when transcribing marked's tag list.
7ef46eb to
0ea00f3
Compare
|
@cursor review |
…code masking maskCodeRegions's fence pattern handled a blockquoted fence (`> \`\`\``) but still required the fence marker at column 0 otherwise, missing CommonMark's independent up-to-3-space fence indent tolerance this codebase already relies on elsewhere (FENCE_OPEN/FENCE_CLOSE in markdown-parse.ts). An indented fence's tag-name mention could still end a whitelisted block early. Fixed by combining both allowances in one prefix pattern (zero-or-more blockquote levels, each independently followed by up to 3 more spaces of indent).
|
@cursor review |
A tag-name mention inside an HTML comment (e.g. `<!-- see </div> below -->`) inside a whitelisted raw HTML block could still be matched by the balance scan and end the block early, fragmenting it. maskCodeRegions already masked fenced/inline code the same way; extend it to mask comments too.
|
@cursor review |
There was a problem hiding this comment.
✅ Bugbot reviewed your changes and found no new issues!
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit c043d9c. Configure here.
Summary
Follow-up to #5438 (already merged) — two issues found during manual testing after merge:
RawHtmlBlockdelegated to marked's built-in block-HTML tokenizer, which (per CommonMark's HTML-block-type-6 rule) stops at the first blank line. A real<details>with a paragraph inside — the common case — fragmented into a raw chip, an ordinary rendered paragraph, and a second raw chip, stranding real content in between. Fixed with a custom block-level tokenizer that scans to the tag's matching close (reusing the balanced open/close depth-tracking already built for nested inline HTML), spanning blank lines, restricted to CommonMark's own block-tag whitelist (details,div,table,section, etc.) — tags that can legitimately start a paragraph (em,a,span,code,kbd) are left untouched.color-mixwith--warningread as an error/warning state, which isn't the right signal) in favor of the same neutral surface as existing code blocks — the hover badge alone now conveys "this is raw, unrendered text".Type of Change
Testing
Added 8 new tests specifically targeting the fragmentation fix (structural node-count assertions, not just string round-trip): multi-paragraph
<details>, multi-paragraph<div>, nested same-tag blocks across blank lines, a paragraph starting with a non-whitelisted inline tag (regression guard so<em>/<a>still parse as real marks), a stray non-whitelisted tag alone on a line, an unterminated tag (no crash), a comment spanning blank lines, and a table+code-block adjacency check. 236/236 tests passing in the directory, fulltsc --noEmitclean, biome clean.Checklist