rsharpier

A Rust port of CSharpier, the opinionated C# formatter. It parses with tree-sitter (a patched copy of tree-sitter-c-sharp) instead of Roslyn and produces the same output as CSharpier for the formatting tests of the .NET version.

uvx rsharpier format .           # or: uv tool install rsharpier / pip install rsharpier

rsharpier format [paths...]      # format files and directories
rsharpier check [paths...]       # exit with 1 if something is not formatted
echo 'class  A{}' | rsharpier format
rsharpier format --write-stdout --stdin-path src/File.cs < src/File.cs
rsharpier pipe-files             # editor integration, files piped through stdin
rsharpier server --server-port N # editor integration, POST /format

Supported: .cs, .csx and the XML flavours (.csproj, .props, .targets, .config, .slnx, .xaml, .axaml, .xml). Options come from .csharpierrc / .csharpierrc.json / .yaml / .yml (with overrides) or .editorconfig; .csharpierignore and .gitignore files are used the same way the .NET version uses them.

Status

The ported formatting tests of CSharpier, which live in a checkout of the .NET repository next to this one (../csharpier, or set CSHARPIER_TEST_FILES), all pass:

cargo test --release
cargo run --release --example conformance [dir] [-v] [name]   # per file results and diffs

Against real code, cargo run --release --example selfcheck <dir> formats every .cs file below a directory (IDEM=1 formats the output a second time). On CSharpier's own sources 370 of 380 files come out byte for byte the same (the rest were formatted with an older version of the .NET formatter). On more than 25,000 files from other repositories (about 4.5 million lines) formatting never panicked, kept all tokens, and was idempotent for all but two files.

How it works

Trivia

Roslyn attaches comments, directives and disabled text to tokens, and CSharpier's printers are written against that. tree-sitter gives comments as free floating "extras" and parses all #if branches as code. trivia.rs

1. finds directive lines (ignoring # inside strings and comments),
2. evaluates #if/#elif/#else for a symbol set, so inactive branches become disabled text,
3. blanks directives and disabled text with spaces (offsets stay stable) and parses that,
4. takes the leaves of the tree as tokens, lexes the gaps between them in the original text into trivia and splits them into trailing (up to the first newline) and leading, like Roslyn.

format() then formats its own output once for every symbol set from symbols::get_sets, which is what the C# version does so code in every branch gets formatted.

Validation

format.rs::validate compares the multiset of tokens before and after formatting (commas are excluded, the formatter adds and removes them; multi line strings only by their content). It is the counterpart of SyntaxNodeComparer and catches printers that drop or invent code.

The grammar

tree-sitter-c-sharp 0.23.5 mis-parses or rejects several things CSharpier needs, so the crate is vendored in grammar/tree-sitter-c-sharp. grammar/patch_grammar.py applies the fixes to the pristine grammar.js and regenerates the parser (it needs the tree-sitter CLI):

python3 grammar/patch_grammar.py <path to upstream grammar.js> <tree-sitter binary>

Fixes: x is null || y (the || ended up in the pattern), a ?? b?.c, a.b is X or Y and member access on patterns, list patterns starting with .., params parameters, and/or/not as identifiers, null conditional assignment, unions, extension blocks, allows ref struct, closed, labeled break/continue, destructors with modifiers, checked(...) and plain names/member accesses as statements, extern alias in namespaces and generic names in new List<(int a, int b)>.

Known differences and gaps

• Sorting of usings uses a lowercase comparison instead of InvariantCultureIgnoreCase.
• There is no formatting cache (--no-cache/--use-cache are accepted and ignored) and no CSharpier.MsBuild version check (--no-msbuild-check is accepted).
• Files in an encoding other than UTF-8/UTF-16 with a byte order mark are read as UTF-8.
• Syntax errors are reported by tree-sitter positions and not with Roslyn's messages.
• Code that tree-sitter cannot parse (preview syntax that is not in the grammar yet) is reported as a syntax error and left alone.