dts-tree-sitter

dts-tree-sitter generates TypeScript .d.ts files for interacting the AST from a given tree-sitter grammar.

Usage

npm i @asgerf/dts-tree-sitter

npx @asgerf/dts-tree-sitter INPUT > OUTPUT.d.ts

where INPUT is used to locate a node-types.json file in one of the following locations:

• ${INPUT}
• ${INPUT}/node-types.json
• ${INPUT}/src/node-types.json
• node_modules/${INPUT}/src/node-types.json

Example

The tree-sitter-javascript grammar can be compiled like this:

npm i tree-sitter-javascript
npx @asgerf/dts-tree-sitter tree-sitter-javascript > generated.d.ts

In the resulting grammar, two of the node types look like this:

export interface ClassDeclarationNode extends SyntaxNodeBase {
  type: SyntaxType.ClassDeclaration;
  bodyNode: ClassBodyNode;
  decoratorNodes?: DecoratorNode[];
  nameNode: IdentifierNode;
}

export interface ClassBodyNode extends SyntaxNodeBase {
  type: SyntaxType.ClassBody;
  memberNodes?: (MethodDefinitionNode | PublicFieldDefinitionNode)[];
}

This can be used like this (see full example):

import * as g from "./generated";

function getMemberNames(node: g.ClassDeclarationNode) {
    let result = [];
    for (let member of node.bodyNode.memberNodes) {
        if (member.type === g.SyntaxType.MethodDefinition) {
            result.push(member.nameNode.text);
        } else {
            result.push(member.propertyNode.text);
        }
    }
    return result;
}

Observe TypeScript do its magic: the type check in the if promotes the type of member to a MethodDefinitionNode in the 'then' branch, and to PublicFieldDefinitionNode in the 'else' branch.

Typed Tree Cursors

Tree sitter's TreeCursor allows fast traversal of an AST, and has two properties with correlated types: nodeType, and currentNode. Once you've checked nodeType, it's annoying to have to cast currentNode to the correponding type right afterwards:

if (cursor.nodeType === g.SyntaxType.Function) {
  let node = cursor.currentNode as g.Function; // annoying cast
}

There's another way, which is handy in large switches: Cast the cursor itself to a TypedTreeCursor before switching on nodeType. Then the guarded use of currentNode has the expected type. For example:

function printDeclaredNames() {
    let cursor = tree.walk();
    do {
        const c = cursor as g.TypedTreeCursor;
        switch (c.nodeType) {
            case g.SyntaxType.ClassDeclaration:
            case g.SyntaxType.FunctionDeclaration:
            case g.SyntaxType.VariableDeclarator: {
                let node = c.currentNode;
                console.log(node.nameNode.text);
                break;
            }
        }
    } while(gotoPreorderSucc(cursor));
}

• node gets the type ClassDeclarationNode | FunctionDeclarationNode | VariableDeclaratorNode.
• This allows safe access to node.nameNode, since each of those types have a name field.
• We don't pay the cost of invoking currentNode for other types of nodes.

Trouble-shooting

I get an error about "excessive stack depth" during compilation

This happens if you compare types from the general tree-sitter.d.ts file with those from the generated .d.ts file. Every type from tree-sitter.d.ts has a stronger version in the generated file; make sure you don't mix and match.

I get UnnamedNode types in places where I don't expect them

This can happen if the grammar contains rules and literals with the same name. For example this grammar rule,

  func: $ => seq('func', $.name, $.body)

will produce a named node with type func, while the 'func' literal will produce an unnamed node with type func as well.

This means a check like node.type === 'func' is not an exact type check, and the type of node will only be restricted to FuncNode | UnnamedNode<'func'>. This is not a bug in the generated .d.ts file: there really are two kinds of nodes you need to handle after that check.

Some possible solutions are:

• Change the grammar to avoid rules with the same name as a keyword.
• Write the check as node.isNamed && node.type === 'func'.
• Change the declared type of node from SyntaxNode to NamedNode.