
CSharp.Refactor
io.github.ThoriumPublished by GitHub account Thoriumv0.1.20Updated Oct 24, 2026
Offline analyzers that find and fix intent-hiding, slow or noisy C#, including agent-written code.
Overview
An offline C# refactoring analyzer that, run with --mcp, exposes analyze and list_rules to an assistant over stdio.
- What you can do
- Analyze C# projects, solutions, source files, scripts or globs and report refactoring suggestions with file and position, optionally applying them in bulk with build verification and dry-run support. List the rule catalog and restrict runs by rule code, category or preprocessor symbols. Produce reports as SARIF, HTML or CSV, with baseline ratcheting and fail-on-findings exit codes.
- When to use it
- Useful when an assistant should review or clean up C# code for correctness, performance and idiomatic style, including agent-written code, and needs a safe, offline analyzer that only edits when told to.
- Requirements
- Runs as a local process over stdio, started with the --mcp flag of the csharp-refactor .NET tool (installed via the dotnet tool install command). Requires the .NET SDK and MSBuild to load projects; no accounts, API keys or network access are declared.
Installation
In SourceWeft
- Open CSharp.Refactor in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
CSharp.Refactor
The C# little-sister of FSharp.Refactor
Functional refactoring suggestions for C#, implemented in F# on Roslyn. The rules care about correctness, measured performance and clear idiomatic C# — and deliberately not about naming, layout or conventions.
- light bulb quick fixes in your editor (Visual Studio, VS Code, Rider, anything that hosts Roslyn analyzers)
- a command-line tool that applies them in bulk, build-verified
Suggestions are Info severity: they mark an opportunity, not a defect, and
never gate your build.
Using it
The tool is fully offline: AI has been used to make it, but it doesn't use AI.
Quick start
Nothing to configure — the tool reads your project, reports what it would change, and only edits when you tell it to:
That prints every fix it would make, with file and position, and writes nothing. When the list looks right, drop the flag to apply them:
If you don't want change public methods / API, drop --api-changes:
It refuses a compilation that does not already build, recompiles in memory after every pass, and builds the project for real at the end, putting the fixes back if that fails.
For light bulbs while you type, see VS Code / Ionide and Visual Studio IDE-plugin instructions below.
[image]For agentic scenarios the dotnet tool supports MCP.
Using the tool:
Point it at whatever you have — the kind is read off the path:
Editor and CI setup
The analyzers ship as
CSharp.Refactor.Analyzers,
an ordinary Roslyn analyzer package. Reference it from the project you want
analysed and every host — dotnet build, Visual Studio, VS Code's C#
extension, Rider — loads it by itself:
The package is a development dependency: it only produces hints and quick
fixes, and nothing from it reaches your compiled output or your bin.
Visual Studio 2022 / 2026
Install the CSharp.Refactor extension:
every C# project opened in the IDE gets the suggestions and Ctrl+. fixes,
with no project change and no effect on builds. Its Tools > CSharp.Refactor
menu runs the csharp-refactor tool on the open solution (apply or dry run,
with or without --api-changes), writes the SARIF report, opens the notes
page and the .editorconfig block — the VS Code extension's commands, in
Visual Studio. Or use the package reference
above; both work, and a project with both sees each finding once.
VS Code
Install the CSharp.Refactor extension.
It bundles the analyzers and, with your consent, wires them into every C#
project on the machine through MSBuild's user ImportBefore hook (one props
file under your profile, undone by a command). Or use the package reference.
CI
The analyzers run inside the compiler, so a build can write the findings as SARIF with no extra tool:
The tool's own report carries more — fixes as suggested changes, context
snippets, stable fingerprints — and pairs with --dry-run for a lint gate:
A dry run exits 0 whether or not it found anything; --fail-on-findings
makes it exit 3 when a finding survives the filters, and --baseline turns
that into a ratchet: findings already in an earlier report are neither
reported nor fixed, only new ones surface.
Applying fixes from the command line
Projects are loaded through MSBuildWorkspace — SDK-style projects of any
target, net48 included, in every framework flavour they list. A legacy
project (<Project ToolsVersion="4.0" …> with explicit <Compile> items)
is read from its project file when the .NET Framework build host does not
answer: compile items, defines, HintPath references that exist, the
targeting pack's assemblies and project references. Without the build tree
(a reference a custom .targets adds, a sibling bin) some types stay
unresolved; the run says how many, the typed rules stand down where a type
is unknown, the in-memory check holds the error count, and the final build
runs through Visual Studio's MSBuild.exe where a baseline build succeeds
— otherwise the run is verified in memory only, and says so. An executable
another project of the solution references (its tests) is not a leaf: its
public shape stays unless --api-changes says otherwise, and every
dependent of a changed project is built too.
A C# script (.csx) is a compilation of its own: the file parsed as a
script, the running runtime as its references, #r resolved against the
script's directory, the runtime and the NuGet cache (#r "nuget: Name, Version"), #load read by the compiler for context (a loaded file is
edited when it is swept as a target of its own, and a directory sweeps
every loose script). Nothing builds a script, so the run is verified in
memory: the error count may not grow, and a script host's globals (Args)
count as errors on both sides.
Code under #if LOCAL_BUILD
Code behind a preprocessor symbol nothing defines is not in the parse tree
at all: the tool neither analyses it nor sees what a fix elsewhere does to
it, and the verification build compiles without it too. A repository that
keeps, say, a local-development path under #if LOCAL_BUILD names the
symbol for the run:
or once for every run, in .editorconfig (csharp_refactor.defines = LOCAL_BUILD, below). The run uses both, and says at the start which symbols
are active and where each came from. The symbols are added to what each
project defines, exactly as dotnet build would see them with
DefineConstants set in the environment: the SDK's DEBUG/TRACE, the
target framework's NET8_0_OR_GREATER and the project's own
$(DefineConstants);FOO all stay. (-p:DefineConstants=LOCAL_BUILD would
not do: a global property replaces every one of them.) Every MSBuild the run
starts — the workspace's design-time build, restores, the verification
builds — gets them that way, and the loaded C# projects, legacy projects and
.csx scripts get them in their parse options as well, so a project that
assigns DefineConstants outright is still analysed with them. A script
whose #r cannot be resolved while another #r sits under #if LOCAL_BUILD
gets the hint to pass --define LOCAL_BUILD.
Configuration
Rules are configured the way every Roslyn analyzer is, in .editorconfig,
nearest section winning:
csharp_refactor.defines is a run-level key: it is read once per run, from
the .editorconfig nearest the target (the walk stops at root = true),
and adds to --define, never replacing it.
csharp-refactor --create-config writes the block for you, every value at
this build's default, so it changes nothing until you edit a line. Each
rule's knobs follow its severity line, at their defaults; a knob whose
default is not one value (CR0023's measured floor, a rule's wrap_column,
which falls back to the file's max_line_length) is written commented out.
A generated file is never touched, in the editor or by the tool: one under
bin, obj, node_modules or .git, one named *.g.cs, *.designer.cs,
*.generated.cs or *.g.i.cs, one listed in ignore_paths, and one whose
header says <auto-generated> (a T4 template's, a resx designer's, the
SDK's AssemblyInfo) or "This code was generated by".
Individual findings are silenced with Roslyn's own means — #pragma warning disable CR0090, [SuppressMessage], a severity of none — and editors honour
those natively. The suppressions policy governs the tool: no-correctness
reports a suppressed correctness finding anyway (and never auto-fixes it),
none reports every suppressed finding, --honor-suppressions overrides
either for one run. The run summary counts what suppressions silenced.
A rule that shadows a Microsoft analyzer rule (CA2213, CA1031, CA2000 …) stands down for that rule's shapes wherever the Microsoft rule is enabled in the file's effective config, so nothing is reported twice and nothing oscillates. Rules.md names each rule's twins.
Warnings as errors: The tool uses the same settings as your CI build (TreatWarningsAsErrors, WarningsAsErrors, .editorconfig severities). If a fix would trigger an analyzer error, the tool leaves that file alone, names the analyzer and the rule, and tells you what to change. Use to take those fixes anyway (the warning stays visible), or to silence the analyzer for everyone.
Improving it
Building and testing
Before committing:
The test harness compiles C# string literals against the running framework,
runs the pure rules, applies every fix as text and recompiles; the host tests
drive the real DiagnosticAnalyzer and CodeFixProvider through Roslyn;
the tool tests sweep a synthetic project through MSBuildWorkspace.
tests/run-tests.ps1 runs the lot. benchmarks/PerfClaims holds the
before/after pair behind every performance claim.
A multi-line input is an indented block through the csharp helper
(tests/CSharp.Refactor.Tests/TestSource.fs): the closing quotes'
indentation is cut from every line, so the rule sees the class at column 0,
and Fantomas can move the block without changing it.
fires 3 "CR0046" source returns a rule's suggestions and asserts their
count, and assertFired does the same for a list filtered by hand. A
mismatch lists the text each suggestion fired on, so the failure names the
site that was missed, where a bare count would not.
A literal that must be a constant (an attribute argument, a [<Literal>],
a printf format) or holds a tab, a \r or trailing whitespace stays an
escaped "...\n..." string.
Beside the example-based suite, tests/CSharp.Refactor.PropertyTests is an
FsCheck suite over generated programs: a class of members each shaped for a
rule, and method bodies around boolean terms over three integers, with an
interpreter over the Roslyn tree as the oracle. Its properties are the
invariants the rules promise - every fix, editor-only alternatives included,
keeps the program compiling; no rule throws on a compilation recovered from
damaged text; applying fixes one at a time reaches a fixed point; and the
boolean and control-flow rewrites (CR0001, CR0005, CR0007, CR0008, CR0011)
keep the method's truth table at every step. The generators reach 121 of
the 124 rules (CR0017 is subsumed by CR0160 wherever both read a shape;
CR0156 and CR0158 need C# 15, which the harness cannot parse), and a
coverage test tries every shape on its own and all of them together: if a
shape stops firing the rule it was written for, it says which. Two more
properties dress the programs the way real files come — #regions around
members, a #if and a comment inside a member, CRLF, tabs, a #pragma —
and require every fix to compile and to leave what was there (no swallowed
directive or comment, no bare LF in a CRLF file); and compile them at every
language version from C# 7.3 to 13, requiring a fix to use no syntax newer
than the file's own. Both run deterministically over every shape as well
as over random programs. A rewrite that would compile and mean something else is
out of the compile invariant's reach, so a shape carries a decoy where it
can: the CR0100/CR0101 shape calls a method whose FormattableString
overload is [Obsolete(error)], so an interpolation that re-binds the call
is a compile error; the CR0082 shape hands one tuple to object, where a
retype would not compile. What no decoy can express (a record made of a
type with behaviour, a record dumped into an exception message) stays with
the example suite.
Design principles
- Never break user code. A fix is only offered when it is provably safe; the borderline case produces no suggestion. A fix must be right without a compile: the same fix runs in the editor, where nothing verifies it.
- Guards over removal. An unsafe rewrite gets a typed proof that the shape is safe and stands down where the proof fails; the rule is never deleted or blanket-disabled.
- Minimal edits. Fixes are range-based text edits; formatting outside the edited range is untouched.
- Pure core, thin adapters. Each rule is a pure function from a syntax tree and semantic model to suggestions; the Roslyn analyzer, the fix provider and the tool are one-line adapters over it.
- Hints point toward idiomatic C# only. Reversible matters of taste belong to the IDE's refactorings, not to a diagnostic.
See DESIGN.md for the full safety model, the yields-to gate, the planned catalog and the relationship to Microsoft's analyzers.
Source: README.md at commit 4999c74
Tools
0Version history
1- v0.1.20LatestOct 14, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts
