Tree-sitter can be used in conjunction with its tree query language as a part of code navigation systems. An example of such a system can be seen in the tree-sitter tags command, which emits a textual dump of the interesting syntactic nodes in its file argument. A notable application of this is GitHub’s support for search-based code navigation. This document exists to describe how to integrate with such systems, and how to extend this functionality to any language with a Tree-sitter grammar.
Tagging is the act of identifying the entities that can be named in a program. We use Tree-sitter queries to find those entities. Having found them, you use a syntax capture to label the entity and its name.
The essence of a given tag lies in two pieces of data: the role of the entity that is matched (i.e. whether it is a definition or a reference) and the kind of that entity, which describes how the entity is used (i.e. whether it’s a class definition, function call, variable reference, and so on). Our convention is to use a syntax capture following the @role.kind capture name format, and another inner capture, always called @name, that pulls out the name of a given identifier.
You may optionally include a capture named @doc to bind a docstring. For convenience purposes, the tagging system provides two built-in functions, #select-adjacent! and #strip! that are convenient for removing comment syntax from a docstring. #strip! takes a capture as its first argument and a regular expression as its second, expressed as a quoted string. Any text patterns matched by the regular expression will be removed from the text associated with the passed capture. #select-adjacent!, when passed two capture names, filters the text associated with the first capture so that only nodes adjacent to the second capture are preserved. This can be useful when writing queries that would otherwise include too much information in matched comments.
This query recognizes Python function definitions and captures their declared name. The function_definition syntax node is defined in the Python Tree-sitter grammar.
name: (identifier) @name) @definition.function
property: (property_identifier) @name)
right: [(arrow_function) (function)]
An even more sophisticated query is in the Ruby Tree-sitter repository, which uses built-in functions to strip the Ruby comment character (#) from the docstrings associated with a class or singleton-class declaration, then selects only the docstrings adjacent to the node matched as @definition.class.
name: (_) @name)
name: (_) @name)
(#strip! @doc "^#\\s*")
(#select-adjacent! @doc @definition.class)
The below table describes a standard vocabulary for kinds and roles during the tagging process. New applications may extend (or only recognize a subset of) these capture names, but it is desirable to standardize on the names below.
You can use the tree-sitter tags command to test out a tags query file, passing as arguments one or more files to tag. We can run this tool from within the Tree-sitter Ruby repository, over code in a file called test.rb:
# won't be included
# is adjacent, will be
Invoking tree-sitter tags test.rb produces the following console output, representing matched entities’ name, role, location, first line, and docstring:
tree-sitter tags test.rb
Foo | module def (0, 7) - (0, 10) `module Foo`
Bar | class def (1, 8) - (1, 11) `class Bar`
baz | method def (2, 8) - (2, 11) `def baz` "is adjacent, will be"
It is expected that tag queries for a given language are located at queries/tags.scm in that language’s repository.
Tags queries may be tested with tree-sitter test. Files under test/tags/ are checked using the same comment system as highlights queries. For example, the above Ruby tags can be tested with these comments:
# ^ definition.module
# ^ definition.class
# ^ definition.method