Use this skill for ast-grep project setup, rule authoring, rule debugging, and CLI workflows that go beyond a single structural query.
Routing
Prefer unified_search with action="structural" and workflow="scan" for read-only project scans.
Prefer unified_search with action="structural" and workflow="test" for read-only ast-grep rule tests.
Prefer structural debug_query on the public tool surface before falling back to raw ast-grep run --debug-query.
Stay on the public structural surface first when the task is only running project checks and reporting findings.
Use unified_exec only when the public structural surface cannot express the requested ast-grep flow.
Quick Start
In VT Code, prefer vtcode dependencies install ast-grep before suggesting system package managers.
External install routes such as Homebrew, Cargo, npm, pip, MacPorts, or Nix are fallback options when the user explicitly wants a system-managed install.
After installation, validate availability with ast-grep --help.
On Linux, prefer the full ast-grep binary name over sg because sg may already refer to setgroups.
When running CLI patterns with shell metavariables like $PROP, use single quotes so the shell does not expand them before ast-grep sees the pattern.
A good first rewrite example is optional chaining, for example rewriting $PROP && $PROP() to $PROP?.().
Command Overview
ast-grep run: ad-hoc query execution and one-off rewrites.
ast-grep scan: project rule scanning.
ast-grep new: scaffold and rule generation.
ast-grep test: rule-test execution.
ast-grep lsp: editor integration via language server.
Built-In Languages
ast-grep ships many built-in languages. Common aliases include bash, c, cc / cpp, cs, css, ex, go / golang, html, java, js / javascript / jsx, json, kt, lua, php, py / python, rb, rs / rust, swift, ts / typescript, tsx, and yml.
--lang <alias> and YAML language: <alias> use those built-in aliases. File-system scans infer language from built-in extensions unless the project overrides them.
In VT Code, public structural lang is passed through to ast-grep. VT Code also normalizes and infers a local subset it can pre-parse itself: Rust, Python, JavaScript, TypeScript, TSX, Go, and Java.
That local subset includes common ast-grep aliases and extensions such as golang, jsx, cjs, mjs, cts, mts, py3, and pyi.
Use languageGlobs when the repository needs a different extension mapping than ast-grep’s built-in defaults.
How Ast-Grep Works
ast-grep accepts several query formats: pattern queries, YAML rules, and programmatic API usage.
The core pipeline is parse first, match second. Tree-Sitter builds the syntax tree, then ast-grep’s Rust matcher finds the target nodes.
The main usage scenarios are search, rewrite, lint, and analyze.
ast-grep processes many files in parallel and is built to use multiple CPU cores on larger codebases.
In VT Code, the public structural surface is the read-only entry point for query, scan, and test. Use the bundled skill when the task is about YAML authoring, rewrite/apply flows, or API-level ast-grep work.
Project Scaffolding
A scan-ready ast-grep project needs workspace sgconfig.yml plus at least one rule directory, usually rules/.
rule-tests/ and utils/ are optional scaffolding that ast-grep new can create for rule tests and reusable utility rules.
If the repository already has sgconfig.yml and rules/, prefer working with the existing layout instead of recreating scaffolding.
Use ast-grep new when the repository does not have ast-grep scaffolding yet.
Use ast-grep new rule when the scaffold exists and the task is creating a new rule plus optional test case.
sgconfig.yml
sgconfig.yml is the project-level ast-grep config file, not a rule file. Treat it like the repository root for rule discovery, tests, parser overrides, and embedded-language behavior.
ruleDirs is required and is resolved relative to the directory containing sgconfig.yml.
testConfigs is optional and configures ast-grep test discovery. Each entry needs testDir; snapshotDir is optional and otherwise defaults to __snapshots__ under that testDir.
utilDirs declares directories for global utility rules shared across multiple rule files.
languageGlobs remaps files to parsers and takes precedence over ast-grep’s default extension mapping, which is useful for similar-language reuse like TS -> TSX or C -> Cpp.
customLanguages registers project-local parsers. libraryPath can be one relative library path or a target-triple map, extensions is required, expandoChar is optional, and languageSymbol defaults to tree_sitter_{name}.
languageInjections is experimental. Each entry needs hostLanguage, rule, and injected.
Use dynamic injected candidates when the rule captures $LANG and the embedded language must be chosen from a list such as css, scss, or less.
Raw ast-grep project discovery walks upward from the current working directory until it finds sgconfig.yml, and --config <file> overrides that discovery with an explicit root config path.
ast-grep scan requires project config and will error if no sgconfig.yml is found. ast-grep run can still search without project config, though it also benefits from discovered config for things like customLanguages and languageGlobs.
ast-grep scan --inspect summary is the quickest way to confirm which project directory and config file ast-grep actually selected during discovery.
ast-grep also recognizes a home-directory sgconfig.yml as a global fallback config. XDG config directories are not part of this behavior.
Rule Catalog
Use the ast-grep catalog as inspiration when the user wants existing example rules, not as something to copy blindly.
Start from examples in the same language family when possible.
Read catalog markers as hints about rule complexity:
simple pattern examples are good starting points
Fix means the example includes a rewrite path
constraints, labels, utils, transform, and rewriters mean the example depends on more advanced rule features
When adapting a catalog example, translate it to the current repository’s language, style, and safety constraints instead of preserving the example verbatim.
Prefer the bundled skill workflow when the user asks to explain, adapt, or combine catalog examples.
Rust Catalog Highlights
Avoid duplicated exports: a Rust lint-style rule can detect pub use foo::Bar; in the same source file that already exposes pub mod foo;. Treat this as API-surface cleanup, not a mechanical rewrite.
Beware chars().enumerate(): the Rust catalog rewrite from $A.chars().enumerate() to $A.char_indices() is valid when the code needs byte offsets instead of character indexes.
Count usize digits without allocation: the catalog rewrite from $NUM.to_string().chars().count() to $NUM.checked_ilog10().unwrap_or(0) + 1 is a good Rust-specific performance cleanup when the target is known to be an integer digit count.
Unsafe function without unsafe block: the Rust catalog’s function_item rule that requires unsafe modifiers but rejects bodies containing unsafe_block is a good review rule for redundant unsafe markers. It is diagnostic-oriented and should usually stay a scan rule, not an automatic rewrite.
Rewrite indoc! macro: the catalog example that removes indoc! { r#"..."# } wrappers is a rewrite-oriented example. Keep it on the CLI skill path because the replacement is formatting-sensitive and should be reviewed interactively before broad apply.
Adapt these rules to the repository’s Rust style before using them directly. In VT Code, preserve existing lint policy, public API conventions, and the project’s bias against unnecessary rewrites.
TypeScript Catalog Highlights
TypeScript vs TSX matters: keep .ts and .tsx rules separate unless the repository intentionally parses .ts as TSX through languageGlobs. Do not assume one pattern works unchanged across both parsers.
Find import file without extension: good scan rule for ESM codebases that require explicit local file extensions on static or dynamic imports. It is policy-dependent, so only use it where the runtime or bundler actually requires explicit extensions.
XState v4 to v5 migration: strong example of multi-rule YAML with utils, transform, and multi-document configs. Keep this sort of migration on the CLI skill path and review the generated diff instead of treating it as a one-line rewrite.
No await inside Promise.all([...]): good rewrite rule when the awaited expression is directly inside the array literal. Keep the rewrite narrow so it does not change intentionally sequential logic hidden behind helper calls.
No console except allowed cases: good scan rule for client-facing TypeScript, but it is repository-policy dependent. Adapt the allowed methods and environments before enabling it broadly.
Find import usage or identifiers: these examples are useful for repository analysis and dependency cleanup, not just linting. They are often better treated as search/report rules than rewrite rules.
Switch Chai should to expect: a useful migration example, but it is test-framework-specific and should be applied only where Chai is actually in use.
Speed up barrel imports: strong rewriters / transform.rewrite example for splitting one import into many direct imports. Keep it on the CLI skill path because path conventions, default-vs-named exports, and formatting policy vary by repository.
Missing Angular @Component() decorator: good example of labels plus pattern-object context and selector. Keep framework-specific rules tied to actual framework usage in the repository.
Logical assignment operators: a compact rewrite example for $A = $A || $B to $A ||= $B, but only apply it where the project’s JS target and lint policy allow ES2021 operators.
Adapt TypeScript catalog rules to the repository’s module system, framework stack, transpilation target, and lint policy before using them directly.
TSX Catalog Highlights
TSX vs TypeScript matters for parsing: JSX-bearing patterns should stay on the TSX parser unless the repository intentionally routes .ts through TSX with languageGlobs.
Unnecessary useState<T> primitives: good cleanup rewrite for useState<string|number|boolean>($A) when the initializer already gives TypeScript enough information to infer the state type.
Avoid && short-circuit in JSX: good React-facing rewrite from {cond && <View />} to {cond ? <View /> : null} when the left side can evaluate to renderable falsy values like 0.
Rewrite MobX component style: useful migration example when observer(() => ...) hides React hook linting from tooling. Keep it on the CLI skill path because naming, export shape, and component conventions vary by repository.
Avoid unnecessary React hooks: good diagnostic rule for use* functions that do not actually call hooks. Treat it as a review rule first, because renaming or de-hooking can be API-affecting.
Reverse React Compiler: clearly rewrite-oriented and intentionally opinionated. Keep it on the CLI skill path and only use it when the user explicitly wants that de-memoization behavior.
Avoid nested links: good accessibility and correctness scan rule for JSX trees.
Rename SVG attributes: strong TSX rewrite example for hyphenated SVG attribute names such as stroke-linecap to strokeLinecap. Keep it reviewable because generated markup can be formatting-sensitive.
Adapt TSX catalog rules to the repository’s React version, JSX runtime, lint rules, framework conventions, and browser-support target before using them directly.
YAML Catalog Highlights
YAML scan rules are useful for configuration-policy checks where the repository needs to flag specific keys or values rather than rewrite source code.
The catalog host/port example is a simple message-oriented rule that matches either host: $HOST or port: $PORT and attaches a diagnostic. Treat it as a starting point for config validation, not a complete policy by itself.
For YAML rules, be explicit about whether the repository cares about the key name, the value, or both. If both matter together, move from separate any patterns to a more structured rule before relying on the result.
Keep YAML config checks repository-specific. Hard-coded values like 8000 are only useful when they reflect an actual project policy.
Ruby Catalog Highlights
Rails *_filter to *_action: useful migration rewrite for older Rails controllers. Keep it on the CLI skill path because framework version, controller style, and review expectations vary by repository.
Prefer symbol over proc: good Ruby cleanup rewrite for cases like .select { |v| v.even? } to .select(&:even?), but only where the shorthand remains readable and matches local Ruby style.
Path traversal detection in Rails: good security-oriented scan rule for Rails.root.join, File.join, or send_file fed by variables. Treat it as a review hint, not a proof of exploitability, because the surrounding validation path still matters.
Adapt Ruby catalog rules to the repository’s Rails version, Ruby style guide, and security posture before using them directly.
Python Catalog Highlights
OpenAI SDK migration: useful multi-rule migration example for legacy openai Python client code, but keep it on the CLI skill path because imports, client lifetime, response shapes, and surrounding application logic often need repository-specific review.
Prefer generator expressions: good example of narrowing a rewrite to contexts like any(...), all(...), or sum(...) where generator expressions are clearly valid. Do not generalize it to every list comprehension.
Walrus operator in if statements: useful paired-rule rewrite example, but only apply it where the repository targets Python 3.8+ and the style guide accepts assignment expressions.
Remove async function: strong rewriters example for stripping async and inner await, but treat it as high-risk migration work because it changes call semantics and often requires broader control-flow review.
Pytest fixture refactors: good example of utils-driven context matching for fixture rename or type-hint updates. Keep it tied to real pytest usage so similarly named non-test code is not swept in.
Optional[T] to T | None and recursive union rewrites: useful typing-modernization examples, but only where the repository targets Python 3.10+ and static typing policy actually prefers PEP 604 unions.
SQLAlchemy mapped_column to annotated Mapped[...]: useful ORM migration example, but keep it on the CLI skill path because ORM version, model style, and nullable semantics need review.
Adapt Python catalog rules to the repository’s Python version floor, framework stack, typing policy, async model, and migration scope before using them directly.
Kotlin Catalog Highlights
Clean-architecture import checks: good scan-rule example for enforcing architectural boundaries with files plus import-path constraints. Treat it as repository-policy enforcement rather than a universal Kotlin rule.
The Kotlin catalog example is diagnostic-oriented, not rewrite-oriented. Keep it on the scan path because import-boundary violations usually need design review instead of blind mutation.
File-scoped package constraints are the point of the example: adapt the files glob and package regexes to the repository’s actual module layout before relying on the result.
Adapt Kotlin catalog rules to the repository’s package naming, architecture boundaries, Android-vs-server structure, and lint ownership before using them directly.
Java Catalog Highlights
Unused local variable detection: useful educational example for has plus ordered all plus precedes, but prefer the project’s established linter or IDE for real unused-variable enforcement because Java variable scopes are broader than the sample rule covers.
Field declarations of type String: good structural scan example showing why field_declaration plus has: { field: type } is more robust than a naive pattern when modifiers and annotations are present.
The Java catalog examples are primarily search/diagnostic material, not high-confidence autofix rules. Keep them review-oriented unless the repository explicitly wants ast-grep-based cleanup instead of compiler or linter diagnostics.
Adapt Java catalog rules to the repository’s package conventions, annotation usage, style tooling, and existing static-analysis stack before using them directly.
HTML Catalog Highlights
HTML parser reuse for framework templates: useful when Vue, Svelte, Astro, or similar files are mostly HTML, but keep parser caveats in mind because framework-specific control flow or frontmatter may require a custom language instead.
Ant Design Vue visible to open: good framework-specific attribute rewrite using enclosing-tag checks plus constraints. Keep it on the CLI skill path because framework version and component set must be confirmed first.
i18n key extraction: useful template rewrite example for wrapping static text while skipping mustache expressions, but keep it reviewable because real projects usually need key naming, dictionary updates, and whitespace policy beyond the raw rewrite.
Adapt HTML catalog rules to the repository’s template framework, parser limitations, i18n workflow, and component-library version before using them directly.
Go Catalog Highlights
Problematic defer with nested function calls: strong Go-specific scan example for catching cases where deferred arguments are evaluated immediately. Treat it as a correctness and test-reliability review rule, and prefer a manual rewrite to a closure-wrapped defer func() { ... }() when the repository agrees with that style.
Function declarations by name pattern: good example of using kind plus has plus regex when a meta-variable pattern cannot express the naming constraint directly, such as Test.*.
Contextual matching for function calls: good illustration of Go parser ambiguity between function calls and type conversions. Use contextual patterns with selector: call_expression when naive call patterns do not behave as expected.
Package-import detection: useful search/scan rule for dependency auditing, compliance checks, or migration prep. Adapt the import regex to the repository’s actual dependency boundaries instead of hard-coding example packages.
Problematic JSON tags with -,: high-signal security-oriented scan rule for Go struct tags. Treat matches as actionable review items because the example represents a real unmarshaling footgun, not just a style preference.
Adapt Go catalog rules to the repository’s Go version, test conventions, package layout, security posture, and existing static-analysis tooling before using them directly.
Cpp Catalog Highlights
Reuse Cpp rules for C only when the repository intentionally parses C sources as Cpp via languageGlobs; do not assume mixed C/C++ projects want that parser tradeoff by default.
Format-string vulnerability rewrite: strong security-oriented example for fprintf/sprintf-style calls missing an explicit format string. Keep it reviewable because real C/C++ codebases may prefer safer API migrations over mechanical "%s" insertion in some contexts.
Struct inheritance matching: useful example of AST-shaped pattern authoring in C++, especially when a shorter surface pattern produces an ERROR node. Use the full structural form or fall back to a YAML rule when the syntax is too incomplete to parse reliably.
Adapt Cpp catalog rules to the repository’s C-vs-C++ parser choice, security posture, libc usage, and coding-standard expectations before using them directly.
C Catalog Highlights
Parsing C as Cpp can reduce duplicated rule authoring, but only use that route when the repository intentionally opts into the parser tradeoff with languageGlobs; do not blur C and C++ semantics by default.
Match function calls in C with contextual patterns: good example of tree-sitter-c ambiguity around fragments like test($A). Use context plus selector: call_expression when a plain call pattern under-parses or turns into macro_type_specifier.
Rewrite method-style calls to function calls: useful migration example for C codebases that emulate methods with structs or function pointers, but keep it on the CLI skill path because it changes calling conventions and may affect ownership, pointer semantics, or naming policy.
Yoda-condition rewrite: clearly style-driven and repository-policy-sensitive. Treat it as optional rewrite material only where the project explicitly prefers constant-on-the-left comparisons.
Adapt C catalog rules to the repository’s parser choice, macro usage, pointer conventions, coding style, and safety policy before using them directly.
JavaScript API Highlights
Use @ast-grep/napi only when rule YAML or VT Code’s public structural tool is not enough. The programmatic API is the right escalation path for computed replacements, ordered-match logic, cross-node inspection, or edit orchestration that would be awkward in pure rule syntax.
Core objects are SgRoot and SgNode: parse(Lang.<X>, source) creates the tree, root() returns the root node, and find / findAll / traversal / refinement / edit APIs live on SgNode.
Matcher inputs can be pattern strings, numeric kind ids, or NapiConfig objects. Prefer patterns or NapiConfig unless there is a concrete reason to drop to raw kind ids.
getMatch and getMultipleMatches expose captured metavariables, but replace does not interpolate metavariables for you. Build replacement strings explicitly in JavaScript from matched nodes before calling commitEdits.
Keep VT Code’s boundary clear: prefer the public structural tool or CLI path for ordinary query/scan/test/rewrite flows, and only drop to NAPI when the task is genuinely programmatic.
registerDynamicLanguage and extra language packages exist, but that path is still experimental. Prefer established parsers and repo-native tooling unless dynamic-language support is actually needed.
Python API Highlights
Use ast-grep-py when the task needs programmatic AST traversal or computed edits but a Python host environment is a better fit than JavaScript or Rust. As with NAPI, prefer it only after rule YAML or VT Code’s structural/CLI path stops being a good fit.
Core objects are again SgRoot and SgNode: SgRoot(source, language) parses the source, root() returns the root node, and search, refinement, traversal, and edit APIs live on SgNode.
find and find_all support either direct rule keyword arguments or a config object. Prefer keyword-rule searches for simple cases and config objects when constraints or utility rules make the query more expressive.
get_match, get_multiple_matches, and __getitem__ expose captured metavariables. __getitem__ is useful when you want a stricter access pattern and are willing to let missing captures raise instead of returning None.
replace and commit_edits generate source edits, but they do not interpolate metavariables for you. Build replacement text explicitly from matched nodes before applying edits.
Keep VT Code’s boundary clear here too: use Python API only for genuinely programmatic transformations, not as a default substitute for public structural queries or ordinary CLI rewrites.
NAPI Performance Highlights
NAPI is not automatically faster than host-language traversal. Performance usually comes from reducing Rust↔JavaScript FFI crossings and letting ast-grep do more work per boundary crossing.
Prefer parseAsync over parse when many parse jobs can benefit from Node’s libuv thread pool and the main JS thread is already busy.
Prefer findAll over manual recursive traversal in JavaScript. One bulk Rust-side search is usually cheaper than repeated kind(), children(), and recursion calls across the FFI boundary.
Prefer findInFiles when scanning many files and you can use its file-path-oriented search model. It avoids unnecessary round-tripping source strings through JavaScript and can parallelize work in Rust threads.
findInFiles has a callback-completion caveat: its returned promise can resolve before all callbacks run. If completion ordering matters, track callback counts explicitly before treating the scan as finished.
Apply these performance tips only when scale justifies them. For small inputs or one-off transformations, simpler synchronous code is often the better tradeoff.
Rule Essentials
Start rule files with id, language, and root rule.
Treat the root rule as a rule object that matches one target AST node per result.
A rule object still needs a positive anchor. In practice, start with pattern or kind; regex is a filter, not a sufficient root rule by itself.
Atomic rules are pattern, kind, regex, nthChild, and range.
Use atomic fields such as pattern, kind, regex, nthChild, and range for direct node checks.
Use relational fields such as inside, has, follows, and precedes when the match depends on surrounding nodes.
Use composite fields such as all, any, not, and matches to combine sub-rules or reuse utility rules.
Rule object fields are effectively unordered and conjunctive; if matching becomes order-sensitive, rewrite the logic with an explicit all sequence instead of assuming YAML key order matters.
language controls how patterns parse. Syntax that is valid in one language can fail in another.
Rule Cheat Sheet
Atomic rules check properties of one node. Start here when a single syntax shape is enough.
pattern, kind, and regex are the common atomic fields. pattern can also be an object with context, selector, and optional strictness.
Pattern objects are for invalid, incomplete, or ambiguous snippets. context is required; selector picks the real target node inside that context; strictness tunes how literally the pattern matches.
Use pattern objects when the bare snippet would parse as the wrong node kind, such as JavaScript class fields or Go/C call expressions inside ambiguous fragments.
kind is usually a plain node kind name, but ast-grep also supports a limited ESQuery-style syntax for some kind selectors.
Separate kind and pattern checks do not change how the pattern is parsed. If parse shape is the problem, switch to one pattern object with context and selector.
regex matches the whole node text. Reach for nthChild when position among named siblings matters and range when the match must be limited to a known source span.
Regex syntax follows Rust regex, not PCRE. Do not assume look-around or backreferences are available, and usually pair regex with kind or pattern so the expensive text check only runs on the right node shapes.
nthChild accepts a number, an An+B string, or an object with position, reverse, and ofRule. Counting is 1-based and only considers named siblings.
range matches by source position with 0-based line and column; start is inclusive and end is exclusive.
Relational rules describe structure around the target node. Use inside, has, follows, and precedes when the match depends on ancestors, descendants, or neighboring nodes.
Read relational rules as: target node relates to surrounding node. The top-level rule still matches the target; the relational subrule matches the surrounding node that filters it.
Config Cheat Sheet
Basic info keys define the rule itself. Use id for the unique rule name, language for the parser target, url for rule documentation, and metadata for custom project data that VT Code should preserve with the rule.
One YAML file can hold multiple rules when you separate documents with ---.
Finding keys define what gets matched. rule is the core matcher, constraints narrows meta-variable captures, and utils holds reusable helper rules that you call through matches.
utils can be purely local to the current file or can supplement global utility-rule files loaded through utilDirs. Keep shared building blocks global only when multiple rule files genuinely need them.
constraints runs after rule matched, only targets single meta variables like $ARG, and is a poor fit inside not.
Patching keys define reusable fixes. Use transform to derive new meta-variables before replacement, fix for either a string replacement or a template object with expandStart / expandEnd, and rewriters when the transformation is too complex for one inline fix.
Linting keys define what scan results report. Use severity, message, note, and labels for diagnostics, then files and ignores to scope where the rule applies.
Severity levels are error, warning, info, hint, and off. hint is the default severity in ast-grep project scans.
error findings make raw ast-grep scan exit non-zero; VT Code normalizes that CLI behavior into structured findings on the public scan path instead of surfacing a tool error.
severity: off disables the rule during scanning. note supports Markdown but cannot interpolate meta variables.
Source suppression uses ast-grep-ignore comments.
Transformation Objects
transform builds new strings from captured meta variables before fix runs.
Each transform entry introduces a new variable name without a leading $. Inside the transform object, source still points at an existing capture or prior transform result using the normal $VAR form.
Later transforms can consume variables created by earlier transforms, so transform order matters when you are stacking multiple string operations.
replace uses a Rust regex over one meta variable. source must be $VAR style, replace is the regex, by is the replacement text, and regex capture groups can be reused in by.
Regex capture groups are only available inside the replace field of the replace transform and can only be referenced from the by field of that same transform. Regular regex rules do not expose those capture groups.
substring slices a meta variable by character index with inclusive startChar and exclusive endChar. Negative indexes count from the end, and slicing is based on Unicode characters rather than raw bytes.
substring behaves like Python string slicing, so omit either bound when the slice should stay open-ended.
convert changes identifier-style casing through toCase. Common outputs are lowerCase, upperCase, capitalize, camelCase, snakeCase, kebabCase, and pascalCase.
Use separatedBy to control how convert splits words before rebuilding the target case. Supported separators include dash, dot, space, slash, underscore, and CaseChange.
CaseChange splits at transitions such as astGrep, ASTGrep, or XMLHttpRequest, which matters when converting mixed acronym identifiers.
Use replace transforms for conditional punctuation or whitespace when a multi-capture may be empty. The common pattern is deriving MAYBE_COMMA or similar from so the extra separator only appears when matches exist.
Rewriters
rewriters is an experimental feature for advanced multi-node rewrites. Prefer ast-grep’s API instead when the YAML starts carrying too much control flow or state.
A rewriter is a smaller rule object with only id, rule, constraints, transform, utils, and fix. id, rule, and fix are required.
Rewriters are only usable through the rewrite transformation. They are not standalone scan/report rules.
Meta variables captured inside one rewriter do not leak to sibling rewriters or the outer rule.
Rewriter-local transform variables and utils are also scoped to that one rewriter.
A rewriter transform can call other rewriters from the same rewriters list when the rewrite pipeline needs multiple internal passes.
Pattern Syntax
Pattern code must be valid code that tree-sitter can parse.
Patterns match syntax trees, so a query can match nested expressions instead of only top-level text.
$VAR matches one named AST node.
$$$ARGS matches zero or more AST nodes in places like arguments, parameters, or statements.
Reusing the same captured name means both occurrences must match the same syntax.
Prefixing a meta variable with _ disables capture, so repeated $_X occurrences do not need to match the same content.
$$VAR captures unnamed nodes when named-node matching is too narrow.
If a short snippet is ambiguous, move to an object-style pattern with more context plus selector instead of guessing.
Pattern Parsing Deep Dive
Pattern creation has four stages: preprocess meta-variable text when the language needs a custom expandoChar, parse the snippet, choose the effective node, then detect meta variables inside that effective node.
Invalid pattern code usually fails because a meta variable is standing in for syntax that the parser treats as an operator or keyword. Patterns like $LEFT $OP $RIGHT or { $KIND foo() {} } should become rule objects using parseable code plus kind, regex, has, or other rule fields.
Incomplete or ambiguous snippets can appear to work only because tree-sitter recovered from an error. Treat that as best-effort behavior, not a stable contract across ast-grep upgrades, and prefer valid context plus selector.
The default effective node is the leaf node or the innermost node with more than one child. Override it with selector when the real match should be a statement instead of the inner expression, especially for follows and precedes.
Meta variables are detected only when the whole AST node text matches meta-variable syntax. Mixed text like obj.on$EVENT, lowercase names like $jq, or string-content fragments do not become meta variables.
$$VAR captures unnamed nodes such as operators when the grammar exposes them only as anonymous tokens. $$$ARGS is lazy: it stops before the next node that satisfies the rest of the pattern.
When pattern behavior is surprising, inspect the parsed tree and effective node first. In VT Code, start with public structural debug_query; in the Playground, use the pattern view for the same questions.
Pattern Core Concepts
ast-grep is structural search, not plain text search. pattern matches syntax-tree shape, while regex is the escape hatch when node text itself matters.
Tree-Sitter gives ast-grep a concrete syntax tree, not a stripped-down abstract syntax tree. That CST detail is why punctuation and modifiers can still matter even when matching stays syntax-aware.
Named nodes carry a kind; unnamed nodes are punctuation or literal tokens. Meta variables match named nodes by default, and $$VAR is the opt-in when unnamed nodes matter.
kind belongs to the node itself. field belongs to the parent-child relationship, so use relational has or inside with field when role matters more than raw node kind.
A node is significant to ast-grep when it is named or has a field. Trivial nodes can still matter for exact matching, so do not assume every important token has its own named node.
Match Algorithm
The default strictness is smart. Every node you spell out in the pattern is respected, but unnamed nodes in the target code can be skipped.
Unnamed nodes written in the pattern are not skipped. A shorter pattern like function $A() {} can match async function, while async function $A() {} requires async to be present.
Use strictness to tune what ast-grep may skip during matching.
cst: skip nothing
smart: default, skip unnamed nodes in code only
ast: skip unnamed nodes on both sides
relaxed: also skip comments
signature: ignore text and compare mostly named-node kinds
This explains why quote differences can disappear under ast, comments can disappear under relaxed, and even different callee text can match under signature.
In VT Code, read-only structural queries already expose strictness. Use the bundled skill when the task is choosing between levels, or when the user needs raw CLI --strictness or YAML pattern-object strictness.
Custom Languages
Use custom language support when the parser exists in tree-sitter form but ast-grep does not ship it as a built-in language.
The basic workflow is:
install tree-sitter CLI and obtain the grammar
compile the parser as a shared library
register it in workspace sgconfig.yml under customLanguages
Prefer tree-sitter build --output <lib> to compile the dynamic library. If the installed tree-sitter is too old for build, use TREE_SITTER_LIBDIR with tree-sitter test as the fallback path.
Reusing a parser library built by Neovim is valid when it already matches the grammar/version you need.
Register libraryPath, extensions, and optional expandoChar in sgconfig.yml. expandoChar matters when $VAR is not valid syntax in the target language and must be rewritten to a parser-friendly prefix.
Use tree-sitter parse <file> to inspect parser output when the custom grammar or file association is unclear.
VT Code’s public structural queries can use a custom language only after the local ast-grep project config is in place. The setup, compilation, and debugging work stays on the bundled ast-grep skill path.
Language Injection
ast-grep can search embedded languages inside a host document. Built-in injection already covers HTML with CSS in <style> and JavaScript in <script>.
Use languageInjections in sgconfig.yml when the embedded language is project-specific, such as CSS inside styled-components or GraphQL inside tagged template literals.
A languageInjections entry needs hostLanguage, a rule, and injected. The rule should capture the embedded subregion with a meta variable such as $CONTENT.
Typical patterns are styled.$TAG\$CONTENT`for CSS-in-JS andgraphql`$CONTENT`` for GraphQL template literals.
ast-grep parses the extracted subregion with the injected language, not the parent document language. That is why CSS patterns can match inside JavaScript once injection is configured.
Use languageGlobs when the whole file should be parsed as a different or superset language. Use languageInjections when only a nested region inside the file changes language.
In VT Code, read-only structural query / scan / test can consume existing injection config. Designing or debugging languageInjections itself stays on the bundled ast-grep skill path.
FAQ Highlights
If a pattern fragment fails, the usual fix is to provide more valid context and then narrow the real target with selector. This is the standard workaround for subnodes like JSON pairs or class fields that are not standalone code.
If a rule behaves strangely, reduce it to the smallest repro, confirm whether it is matching an expression or a statement, and use all to make rule order explicit when later checks depend on earlier meta-variable captures.
CLI and Playground can disagree because parser versions and text encodings differ. In VT Code, prefer the public structural debug_query flow first, then compare the parsed AST or CST before assuming the rule is wrong.
Meta variables must occupy one whole AST node. use$HOOK and similar prefix/suffix patterns will not work; capture the full node and narrow it with constraints.regex instead. Use $$VAR for unnamed nodes, and remember that $$$MULTI is lazy.
Do not combine separate kind and pattern rules to force a different parse shape. Use one pattern object with context and selector so the parser sees the intended node kind.
ast-grep rules are single-language. Share coverage across related languages by parsing both with the superset via languageGlobs, or keep separate rules when the AST differences matter.
ast-grep does not provide scope, type, control-flow, data-flow, taint, or constant-propagation analysis. If the task needs those, switch tools instead of stretching rule syntax.
Find & Patch
ast-grep rewrites are still find first, patch second. rule plus optional constraints finds the target, transform derives replacement strings, and fix patches the final text.
The simple workflow rewrites one matched node at a time. When one node must expand into multiple outputs, use rewriters plus transform.rewrite instead of forcing everything into one inline fix.
transform.rewrite lets you run sub-rules over a matched meta-variable, generate one fix per sub-node, and join the results with joinBy.
transform.rewrite is still experimental. It rewrites descendants of the captured source, prevents overlapping rewriter matches, prefers higher-level AST matches first, and for one node only applies the first matching rewriter in the declared order.
This is the right model for list-style rewrites such as exploding a barrel import into multiple single imports, and it is the canonical example for rewriter usage.
Keep this declarative workflow on the ast-grep skill path. VT Code’s public structural surface stays read-only and does not expose rewrite/apply behavior.
Rewrite Essentials
Use ast-grep run --pattern ... --rewrite ... for one-off rewrites.
Use YAML fix in rule files for reusable rewrites that should live with the rule.
Use --interactive to review rewrite hunks before applying them.
Use --update-all or -U only when the user clearly wants non-interactive apply behavior.
Meta variables captured in pattern can be reused in fix.
String fix is raw replacement text, not a parsed Tree-Sitter pattern. Meta variables can appear anywhere in the replacement string.
fix indentation is preserved relative to the matched source location, so multiline rewrites must be authored with deliberate indentation.
Non-matched meta variables become empty strings in rewritten output.
If appended uppercase text would be parsed as part of a meta variable name, use transforms instead of writing $VARName directly.
Use transform.rewrite when a matched list must be rewritten element-by-element before the outer fix runs.
Use joinBy to control how rewritten list items are stitched together, for example newline-joined imports in a barrel-import rewrite.
Use FixConfig when replacing only the matched node is not enough, especially for deleting list items or key-value pairs that also need a surrounding comma removed.
In FixConfig, template is the replacement text and expandStart / expandEnd widen the rewritten range to consume commas, brackets, or other surrounding trivia outside the target node.
Keep advanced transform and rewriters in the skill-driven CLI workflow.
Run Command Basics
ast-grep -p 'foo()' and ast-grep run -p 'foo()' are equivalent. run is the default subcommand.
ast-grep run defaults to searching . when no path is provided and can search multiple paths in one invocation.
--globs includes or excludes paths and overrides ignore-file behavior. Prefix a glob with ! to exclude, and let later globs win when multiple patterns match.
--no-ignore changes which ignore sources ast-grep respects. The supported categories are hidden, dot, exclude, global, parent, and vcs.
--follow makes ast-grep traverse symlinks. Expect loop or broken-link errors to surface directly from the CLI when the filesystem is invalid.
Scan Command Basics
ast-grep scan defaults to searching . when no path is provided and can search multiple paths in one invocation.
--config <file> points scan at a project sgconfig.yml root. It is the default scan mode in VT Code’s public workflow="scan" surface.
--rule <file> runs one YAML rule file without project setup and conflicts with --config.
--inline-rules '...' runs one or more inline YAML rules without creating a file on disk. Separate multiple rules with YAML ---. It conflicts with --rule.
--filter <regex> narrows project-config scan to matching rule ids and conflicts with --rule.
--include-metadata only affects JSON output and is already enabled on VT Code’s public scan path so normalized findings can carry rule metadata.
Test Command Basics
ast-grep test validates rule tests from the ast-grep project config.
Rule test files are YAML with id, valid, and invalid. valid cases should produce no issue; invalid cases should produce at least one issue.
Ast-grep’s test output distinguishes four outcomes:
reported: invalid code correctly reports
validated: valid code correctly stays quiet
noisy: valid code reported unexpectedly
missing: invalid code was not reported
--config <file> points test execution at a specific ast-grep root config.
--test-dir <dir> narrows where test YAML files are discovered.
--snapshot-dir <dir> changes the snapshot directory name from the default __snapshots__.
--filter <glob> narrows which rule test cases run.
--skip-snapshot-tests checks test validity without snapshot-output assertions. VT Code exposes this one on the public workflow="test" path.
--include-off includes severity: off rules during test runs.
--update-all generates or refreshes snapshot baselines, usually under __snapshots__/.
--interactive is for selective snapshot updates after rule or test changes.
Snapshot tests cover output details such as spans, labels, or message rendering in addition to simple valid/invalid matching, so --skip-snapshot-tests is useful while a rule is still evolving.
Other Commands
ast-grep new [project|rule|test|util] scaffolds a project or individual items. Common flags are --lang, --yes, --base-dir, and an optional item NAME.
ast-grep lsp starts the language server and accepts an optional --config <file>.
ast-grep completions [shell] generates shell completion scripts for bash, elvish, fish, powershell, or zsh.
ast-grep help and ast-grep --help are the authoritative command-discovery entry points when the exact subcommand or flags are in doubt.
CLI Modes
--interactive is for reviewing rewrite hunks one-by-one; ast-grep’s interactive controls are y, n, e, and q.
--json=pretty|stream|compact is for raw ast-grep JSON output when the user needs native ast-grep payloads or shell pipelines. pretty is the default if a style is not specified. Prefer VT Code’s normalized structural results when those are sufficient.
Raw ast-grep JSON match objects include fields such as text, range, file, lines, optional replacement, optional replacementOffsets, and optional metaVariables. Scan-mode rule matches add fields like ruleId, severity, message, and optional note.
ast-grep JSON positions are zero-based for line, column, and byte offsets. Keep that convention in mind when translating payloads into editor-facing or user-facing locations.
--json=stream emits one JSON object per line and is the better fit for large pipelines; pretty and compact emit one JSON array and are easier to inspect but less streaming-friendly.
--json=<STYLE> must use the equals-sign form. --json stream is parsed as plain --json plus an extra positional argument, not as --json=stream.
--stdin is for piping code into ast-grep. It conflicts with --interactive.
ast-grep run --stdin requires an explicit --lang because stdin has no file extension for language inference.
ast-grep scan --stdin only works with one single rule via --rule / -r.
--stdin only activates when the flag is present and ast-grep is not running in a TTY.
--heading=auto|always|never only changes the human-readable text layout. It does not matter when VT Code is already consuming structured JSON.
--color=auto|always|ansi|never only controls terminal coloring. VT Code’s public structural query forces plain output with --color=never.
API Escalation
Do not force complex transformations into rule syntax when the task needs arbitrary AST inspection or computed replacements.
Escalate to ast-grep’s library API when the task needs conditional replacement logic, counting or ordering matched nodes, per-node patch generation, or replacement text computed from matched content and surrounding nodes.
Node.js NAPI is the main experimental API surface today. The common entry points are parse, kind, and pattern, and the main objects are SgRoot and SgNode.
In NAPI, parse(Lang.<X>, source) returns SgRoot, root() returns SgNode, and traversal/search APIs like find, findAll, field, parent, children, matches, inside, has, replace, and commitEdits live on SgNode.
NapiConfig is the programmatic equivalent of rule YAML for find / findAll, and FindConfig is the config shape for file-based searching.
Python bindings expose the same general model with SgRoot(src, language) plus SgNode methods for rule checks, traversal, searching, and edit generation.
JS language-specific objects like js.parse(...) are deprecated; prefer the unified NAPI functions with Lang.JavaScript.
Rust ast_grep_core is the lowest-level and most efficient option, but also the heaviest lift.
Applying ast-grep fix through the JS/Python APIs is still experimental, so prefer generating explicit patches in code when reliability matters.
If the target language has no suitable JS/Python parser path for the desired automation, prefer a Rust implementation or another repo-native AST approach instead of overcomplicating ast-grep rules.
Use unified_exec For
ast-grep --help
ast-grep new
ast-grep new rule
ast-grep scan -r <rule.yml> <path>
ast-grep scan --rule <file> <path>
ast-grep scan --inline-rules '...' <path>
ast-grep run --pattern <pattern> --rewrite <rewrite>
ast-grep run --json
ast-grep run --stdin --lang <lang>
ast-grep run --no-ignore hidden --follow
ast-grep run --inspect summary --threads 4
ast-grep run --context 2
ast-grep scan --format sarif --report-style short
ast-grep scan --error=rule-id
ast-grep scan --stdin --rule <rule.yml>
ast-grep test --config sgconfig.yml --filter 'rust/*'
ast-grep test --test-dir rule-tests --snapshot-dir __snapshots__
ast-grep test --include-off --update-all
ast-grep lsp
ast-grep completions
ast-grep new project --base-dir . --yes
ast-grep new rule my-rule --lang rust
ast-grep help
ast-grep GitHub Action setup
ast-grep programmatic API experiments and library examples
System package-manager install commands when the user explicitly wants them
sg new
Rewrite or apply flows
Interactive ast-grep flags
Raw ast-grep color control such as --color never
transform, rewrite, joinBy, or rewriters
Non-trivial sgconfig.yml authoring or debugging
Rule authoring tasks that need direct ast-grep CLI iteration beyond public scan/test
Read More
Read references/project-workflows.md when you need the boundary between public scan/test support and skill-driven CLI work, or when you need a quick reminder of ast-grep pattern and rule essentials.
Keep sgconfig.yml authoring on the skill path. VT Code’s public structural tool can consume an existing config through config_path, but it does not expose these top-level schema fields directly.
Relational subrules can themselves use pattern, kind, composites, and captures. Those captures can still be referenced later in fix, which is a practical way to extract surrounding syntax while keeping the target node as the match.
Add relational field when the surrounding node matters by semantic role, not just by shape. field only applies to inside and has.
Add stopBy when ancestor or sibling traversal must continue past the nearest boundary instead of stopping early. The default is neighbor, end searches to the boundary, and a rule object stop is inclusive.
inside means the target is somewhere under a matching ancestor, has means the target node contains a matching descendant, follows means the target comes after a matching sibling or prior node, and precedes means it comes before one.
Composite rules combine checks for the same target node. Use all for explicit conjunction, any for alternatives, not for exclusions, and matches to delegate to a utility rule.
all and any still operate on one target node. They combine sub-rules, not multiple matched nodes.
all is the ordered composite. Use it when later checks depend on captures established by earlier pattern matches, because YAML rule-object field order is not guaranteed.
any is for alternatives, not for "collect all matching nodes". If one node cannot satisfy every branch at once, all is the wrong operator even if the surrounding structure feels plural.
Nested composites still evaluate one node at a time. A rule like has: { all: [...] } means "has one child satisfying every listed rule", not "has one child for each listed rule".
When you need "a node that has X child and has Y child", write all: [ { has: ... }, { has: ... } ] on the outer target instead of putting incompatible checks inside one nested all.
Utility rules keep repeated logic out of the main rule body. Use file-local utils for one config file and global utility-rule files when multiple rules in the project need the same building block.
Local utility rules live under the current file's utils map, are only visible in that one config file, inherit the file's language, and cannot define their own separate constraints.
Global utility rules live in separate files discovered through utilDirs. They must declare id and language, and can only use the utility-safe fields: id, language, rule, constraints, and local utils.
Local utility names must be unique inside one file. A local utility can shadow a global utility with the same name, so check the nearest file first when matches seems to resolve unexpectedly.
Utility rules can call other utility rules through matches, including recursive structural tricks like nested-parentheses matching. Avoid cyclic matches dependency graphs, because ast-grep does not allow recursive cycles there.
Self-reference through relational structure such as inside or has is different from cyclic matches reuse and is allowed when the AST traversal still makes progress.
Switch from a single pattern to a rule object when you need positional constraints, role-sensitive matching, reusable sub-rules, or several structural conditions on one node.
Rule-object fields are logically equivalent to an all across those fields, but not to an ordered all. Keep explicit all when capture order matters; use rule-object style when the checks are independent and flatter indentation helps readability.
ast-grep-ignore suppresses all rules for the same line or following line
ast-grep-ignore: rule-id suppresses one rule
comma-separated rule ids suppress multiple specific rules
next-line suppression only works when there is no preceding AST node on that same comment line
File-level suppression requires the suppression comment on the first line plus an empty second line.
unused-suppression is a built-in hint-style rule with autofix for stale ignore directives, but it only appears in full scan runs when ast-grep is not filtering or disabling rules through CLI narrowing flags.
labels keys must come from meta variables already defined by the rule or constraints.
files supports either plain globs or object entries. Use object syntax when you need options like caseInsensitive glob matching.
ignores runs before files. Both are relative to the sgconfig.yml directory, and the glob should not start with ./.
Rule-level ignores is different from CLI --no-ignore: the CLI flag changes global ignore-file behavior, while YAML ignores only filters files for that rule.
JSON output only includes rule metadata when the ast-grep run enabled metadata output, for example via --include-metadata.
Keep config authoring on the ast-grep skill path. VT Code’s public structural tool runs read-only query/scan/test workflows; it does not expose rule-YAML authoring fields directly.
$$$ARGS
String-form transforms such as replace(...), substring(...), convert(...), and rewrite(...) are valid shorthand in newer ast-grep versions, but keep examples explicit when debugging.
String-form transforms require ast-grep 0.38.3 or newer. Prefer object form when compatibility or debugging clarity matters.
--format=github|sarif is for CI/reporting pipelines, not VT Code’s normalized public scan result shape.
--report-style=rich|medium|short only changes ast-grep’s human-readable diagnostics.
--error, --warning, --info, --hint, and --off override rule severities for one scan run. These flags belong on the CLI skill path, not VT Code’s public structural surface.
--inspect entity is the direct CLI way to inspect each rule’s final enabled severity, including overrides and project-config effects.
unused-suppression can also have its severity overridden on the CLI, but that is still CLI-only behavior outside VT Code’s public structural surface.
--inspect=summary|entity emits file and rule discovery diagnostics to stderr without changing the actual match results.
-C/--context shows symmetric surrounding lines. -A/--after and -B/--before are asymmetric alternatives and conflict with --context.
ast-grep run exits 0 when at least one match is found and 1 when no matches are found. VT Code normalizes that no-match case to an empty matches array on the public structural query path.
ast-grep scan exits 1 when at least one error-severity rule matches and 0 when no rules match. VT Code normalizes that error-finding case to structured findings instead of surfacing a tool error.
