Meziantou.Framework.Language.Shell provides an immutable shell script concrete syntax tree (CST) with roundtrip-safe parsing, diagnostics, source locations, trivia (comments/whitespace), and editing helpers.
It is modelled on Roslyn. If you have used Microsoft.CodeAnalysis, everything here will look familiar: a
SyntaxTree over a SourceText, nodes and tokens with spans, trivia, SyntaxKind, visitors, rewriters, and
annotations.
- parse a script in the dialect you choose, without reformatting untouched text
- keep every character, including comments, blank lines, and line continuations
- report syntax issues through diagnostics (parsing never throws, whatever the input)
- edit nodes, tokens and trivia, rebuilding only what changed, and serialize back with
ToFullString() - walk or rewrite the tree with visitors
ShellDialect |
Family | Notes |
|---|---|---|
Sh |
POSIX | strict POSIX baseline |
Bash |
POSIX | [[ ]], (( )), <<<, arrays, function, <(…), coproc, select, arithmetic ** |
Zsh |
POSIX | the bash set plus foreach/end, repeat, always, anonymous functions, =(…), glob qualifiers, and brace groups that close without a separator |
PowerShell |
PowerShell | Windows PowerShell 5.1 |
PowerShellCore |
PowerShell | pwsh 7+: &&/||, ternary ? :, ??/??=, clean blocks |
Cmd |
Cmd | cmd.exe batch |
Dialects within a family share a parser; ShellDialect.Features records what each one supports. POSIX defines $((1+2)), so every dialect in the family parses it as an arithmetic expansion, while the ((1+2)) arithmetic command is bash and zsh only and stays plain text in sh.
using Meziantou.Framework.Language.Shell;
const string Script = """
# deploy the app
set -euo pipefail
for target in web api; do
if [[ -d "src/$target" ]]; then
dotnet publish "src/$target" -c Release | tee "logs/$target.log"
fi
done
""";
var tree = ShellSyntaxTree.ParseText(Script, ShellDialect.Bash);
// Nothing is lost: the tree reproduces the input byte for byte.
Console.WriteLine(tree.GetRoot().ToFullString() == Script); // True
// Invalid input produces diagnostics instead of exceptions.
foreach (var diagnostic in tree.GetDiagnostics())
{
Console.WriteLine($"{diagnostic.Id} at {diagnostic.Location}: {diagnostic.Message}");
}To read a single command rather than a whole script:
var command = (ShellCommandSyntax)ShellSyntaxTree.ParseCommand("git commit -m 'wip'", ShellDialect.Bash);
Console.WriteLine(command.NameValue); // git
Console.WriteLine(command.Arguments[2].Value); // wipParseCommand is the entry point for a single command; there is no separate expression entry point, because what a
shell expression is differs per dialect. Expressions are reached through the nodes that contain them:
PosixArithmeticExpansionSyntax.Expression for $(( )), PosixDelimitedExpressionStatementSyntax.Expression for
(( )) and [[ ]], and PowerShellExpressionStatementSyntax.Expression for a PowerShell expression statement.
var command = (ShellCommandSyntax)ShellSyntaxTree.ParseCommand("ls /root", ShellDialect.Bash);
var arithmetic = ShellSyntaxTree.ParseText("echo $((1 + 2 * 3))", ShellDialect.Bash)
.GetRoot().DescendantNodes().OfType<PosixArithmeticExpansionSyntax>().Single();
// 1 + (2 * 3): precedence is in the tree, not left to the caller.
var binary = (ShellBinaryExpressionSyntax)arithmetic.Expression;
Console.WriteLine(binary.OperatorText); // +
Console.WriteLine(binary.Right is ShellBinaryExpressionSyntax); // TrueEvery node exposes its Kind(), its Span (excluding trivia) and FullSpan (including it), its Parent, and the usual traversal methods: ChildNodes(), ChildNodesAndTokens(), DescendantNodes(), DescendantTokens(), DescendantTrivia(), Ancestors(), plus FindToken(position) and FindNode(span). Traversal is in source order.
The separators belong to the list, not to what they follow. A statement list, a pipeline, and a command list are
separated lists: the statements and the ;, | or && between them share one sequence, so Count counts statements
and SeparatorCount counts separators.
Comments are trivia, so they never interrupt the node structure but still round-trip:
foreach (var comment in tree.GetRoot().DescendantComments())
{
Console.WriteLine($"{comment.Span.Start}: {comment}");
}An edit rebuilds only the path from the changed node up to the root. Everything else is carried over as-is, so the rest of the script is untouched — and the result is the same type you started with, so no cast is needed.
Nothing is carried over onto the replacement, including the trivia in front of the node being replaced. Ask for it
with WithTriviaFrom when you want it:
var tree = ShellSyntaxTree.ParseText("echo old # keep this", ShellDialect.Bash);
var command = (ShellCommandSyntax)tree.GetRoot().Statements.Statements[0];
var argument = command.Arguments[0];
var updated = tree.GetRoot().ReplaceNode(argument, SyntaxFactory.Word("new", ShellDialect.Bash).WithTriviaFrom(argument));
Console.WriteLine(updated.ToFullString()); // echo new # keep thisReplaceToken, ReplaceTrivia, RemoveNode and the WithX method on every slot of every node work the same way.
For text-based edits, use WithChanges, which reparses:
var tree = ShellSyntaxTree.ParseText("echo old", ShellDialect.Bash);
var updated = tree.WithChanges(new TextChange(new TextSpan(5, 3), "new"));
Console.WriteLine(updated.GetRoot().ToFullString()); // echo newGetChanges reports what actually differs between two trees, with the common prefix and suffix trimmed, and
IsEquivalentTo compares them structurally, so two scripts that differ only in whitespace or comments are equivalent:
var a = ShellSyntaxTree.ParseText("echo a # note", ShellDialect.Bash);
var b = ShellSyntaxTree.ParseText("echo a", ShellDialect.Bash);
Console.WriteLine(a.IsEquivalentTo(b)); // TrueAn annotation is a marker you attach to a node and find again in the tree an edit produced, wherever it ended up:
var marker = new SyntaxAnnotation();
var marked = root.ReplaceNode(command, command.WithAdditionalAnnotations(marker));
var edited = marked.ReplaceNode(/* something else entirely */);
var found = edited.GetAnnotatedNodes(marker).Single();SyntaxFactory creates nodes programmatically and quotes for the target dialect only when needed:
var command = SyntaxFactory.Command(ShellDialect.Bash, "echo", "two words", "plain");
Console.WriteLine(command.ToFullString()); // echo 'two words' plainShellSyntaxVisitor, ShellSyntaxVisitor<TResult>, and ShellSyntaxRewriter cover every node type across all dialects, so one walker handles any tree. A rewriter descends into every node whatever its type, returns the original instance when nothing changed, and keeps the exact text of everything it did not touch:
sealed class RenameCommand(string oldName, string newName) : ShellSyntaxRewriter
{
public override SyntaxNode? VisitCommand(ShellCommandSyntax node)
{
if (node.NameValue != oldName || node.Name is null)
return base.VisitCommand(node);
// WithText keeps the word's own leading trivia, so the comment and indentation in front of the
// command are not lost. A node built from scratch carries no trivia and would drop them.
var renamed = node.Name.WithText(newName);
return node.WithElements(new SyntaxList<ShellSyntaxNode>(
node.Elements.Select(child => ReferenceEquals(child, node.Name) ? renamed : child)));
}
}rewriter.Visit(tree.GetRoot()) returns a new ShellScriptSyntax; visiting a node further down scopes the rewrite to
that subtree. ShellSyntaxWalker visits a node and everything below it without producing anything, and
ShellSyntaxVisitor dispatches on kind without recursing at all.
var options = new ShellParseOptions(ShellDialect.Bash) { MaxRecursionDepth = 64 };
var tree = ShellSyntaxTree.ParseText(script, options);MaxRecursionDepth bounds how deeply the parser descends. Input that nests beyond it reports SHELL0100 and keeps the remainder as skipped text, so deeply nested input cannot overflow the stack.
The depth limit bounds nesting, not length. An operator or member chain such as $x.a.b.c... is built by a loop rather than by recursion, so it is accepted at any length and produces a tree as deep as the chain is long. Building, walking, and editing such a tree uses no recursion either.
Version 3 moved the tree onto the same model Roslyn uses, which changed most of the API.
| Version 2 | Version 3 |
|---|---|
ShellSyntaxKind |
SyntaxKind |
node.Kind |
node.Kind() |
ShellSyntaxToken, ShellSyntaxTrivia, ShellSyntaxNodeOrToken (classes) |
SyntaxToken, SyntaxTrivia, SyntaxNodeOrToken from Meziantou.Framework.Language (structs) |
tree.Root, tree.Text, tree.SourceText, tree.Diagnostics |
tree.GetRoot(), tree.GetText(), tree.GetDiagnostics() |
node.ChildNodes (a property) |
node.ChildNodes(), or the node's own named slots |
command.ChildNodes / WithChildNodes |
command.Elements / WithElements |
list.SeparatorTokens alongside the nodes |
one SeparatedSyntaxList; the named property still reads the separators |
token.Text on an absent optional token being null |
the default token, whose text is "" — ask token.IsPresent() |
trivia.Text |
trivia.ToString() |
root.ReplaceNode(...) re-parsing, and keeping the old node's trivia |
root.ReplaceNode(...) rebuilding only what changed, and keeping nothing — use WithTriviaFrom |
root.ReplaceNode(foreign, …) quietly doing nothing |
it throws, because the node is not in the tree |
ShellSyntaxRewriter overrides returning ShellSyntaxNode? |
returning SyntaxNode? |
ShellSyntaxVisitor.DefaultVisit recursing |
it does nothing; derive from ShellSyntaxWalker to walk a tree |
redirection.HereDocument set by the parser |
derived from the tree, so it survives an edit |
New in version 3: WithX on every slot, SyntaxAnnotation, RemoveNode/SyntaxRemoveOptions, ShellSyntaxWalker,
FindToken/FindNode, and structural sharing — an edit keeps every node it did not touch.
Replacing a node, token or trivium rebuilds only the spine from it to the root; WithChanges reparses, because an
edit expressed as text can change how everything after it reads.
Nodes are shared between the trees an edit produces, so holding several versions of a script costs little more than holding one.