Repository navigation
Glob ? and [...] never match an emoji or other non-BMP character: "icon?.png" misses "icon😀.png", and "x[😀🎉]" misses "x😀" #176
Description
What's wrong
IsGlobMatch (TextFilter/TextFilter.cs:443-447) and ResolveGlob (TextFilter/TextFilter.cs:637-663) hand the pattern and the text to DotNet.Glob as UTF-16 code units. A character outside the Basic Multilingual Plane (emoji, CJK Extension B such as 𠮷, and so on) is a surrogate pair, so DotNet.Glob sees two characters where the user sees one.
Reproduced at f681291 with TextFilter.IsMatch(text, pattern, TextFilterType.Glob, TextFilterMatchOptions.ByWordAny):
| Pattern | Text | Actual | Expected |
|---|---|---|---|
icon?.png |
icon😀.png |
false | true |
icon??.png |
icon😀.png |
true | false |
??家 |
𠮷野家 |
false | true |
???家 |
𠮷野家 |
true | false |
??家 |
吉野家 (BMP) |
true | true |
x[😀🎉] |
x😀 |
false | true. The class lists the character literally. |
x[!a] |
x😀 |
false | true. 😀 is not a. |
Why it matters
- Type-ahead filtering over file names or labels that contain emoji or rare CJK characters silently drops matching items, and the user has no way to see why.
- A negated class such as
[!a]excludes every supplementary character. - FuzzySearch, which this library also uses, was fixed to match whole code points (FuzzySearch Bump the ktsu group with 9 updates #71/Bump the ktsu group with 11 updates #73, Bump MSTest.Sdk from 4.4.0 to 4.4.1 #90). The glob path of the same filter now disagrees with the fuzzy path.
Suggested fix
Do with surrogate pairs what MaskPathSeparators already does with / and \: map them to private-use characters before DotNet.Glob sees either string.
- Build a per-pattern map from each distinct supplementary code point that appears in the pattern to its own private-use BMP character (for example from U+E001 upward, avoiding U+E000, which is the masked separator). Apply the map to both the pattern and the text.
- In the text, replace any supplementary code point that is not in the map with one shared private-use placeholder.
?,*and negated classes then count it as one character. - Lone surrogates can be left as they are.
Acceptance criteria
- The seven cases in the table above return the expected value.
- The existing glob tests, including those for path-separator masking and
**/expansion, still pass.
Activity
matt-edmondson commentedon Oct 31, 2026
Triage
[!a]excluding every such character is the most surprising case. The glob path now disagrees with the fuzzy path, which FuzzySearch Bump the ktsu group with 9 updates #71/Bump the ktsu group with 11 updates #73/Bump MSTest.Sdk from 4.4.0 to 4.4.1 #90 already fixed.TextFilter.cs(IsGlobMatch,ResolveGlob). The proposed fix extends the existingMaskPathSeparatorsapproach.1/2also matches1\2, and excluding-*/*also hidesC:\temp:/and\are masked to the same character #132). That PR changes which private-use characters the separator masking uses, and this fix has to allocate around them. Land Mask / and \ to different characters in glob matching [patch] #175 first and build the code-point map on top of its masking scheme, so the two do not collide in the U+E000 range.Next step: once #175 merges, add the per-pattern code-point map and the shared placeholder for unmapped supplementary characters. Cover the seven table cases with tests.
Generated by Claude Code