Turtles

Turtles is a small Python DSL for writing parsers that feel like dataclasses: you define a grammar as a collection of Rule classes, parse some input, and get back a hydrated object you can inspect, transform, or serialize.

This is especially useful when you have a custom format (or “mostly structured” text) and want a real parser instead of a giant regex, or handrolled parser.

NOTE: Implementation is still evolving. Please open an issue if you hit unexpected behavior.

Install

pip install turtles

Requires Python 3.12+. If you’re on Python <3.14, it’s recommended to add from __future__ import annotations at the top of your grammar modules so you can forward-reference rules (and define rules in any order).

Quickstart

Define a grammar (in a .py file), parse input, and use the structured result.

from turtles import Rule, char, repeat, at_least, separator

# define rules for our grammar
class Int(Rule, int):
    value: repeat[char["0-9"], at_least[1]]

class Float(Rule, float):
    whole: Int
    "."
    frac: Int

Number = Float | Int

class KV(Rule):
    key: repeat[char["a-zA-Z_"], at_least[1]]
    "="
    value: Number

class Row(Rule):
    items: repeat[KV, separator[" "], at_least[1]]


# parse some input with the grammar
src = "temp=21.5 humidity=45 retries=0"
row = Row(src)

# Work with hydrated objects
assert row.items[0].key == "temp"
assert row.items[0].value == 21.5

# Convert the whole parse result to plain Python containers
data = row.as_dict()
assert data == {
    "items": [
        {"key": "temp", "value": 21.5},
        {"key": "humidity", "value": 45},
        {"key": "retries", "value": 0},
    ]
}

# # Helpful while iterating on a grammar
print(repr(row))
# Row
# └── items: [3 items]
#     ├── [0]: KV
#     │   ├── key: temp
#     │   └── value: Float(float)
#     │       ├── whole: Int(int)
#     │       │   └── value: 21
#     │       └── frac: Int(int)
#     │           └── value: 5
#     ├── [1]: KV
#     │   ├── key: humidity
#     │   └── value: Int(int)
#     │       └── value: 45
#     └── [2]: KV
#         ├── key: retries
#         └── value: Int(int)
#             └── value: 0

Important notes

• Rules must be defined in a real source file (not a REPL / exec) because Turtles inspects source to build the grammar.
• Named fields (e.g. key: ...) become attributes on the hydrated result.
• Unnamed fields are anonymous and only used to guide parsing, but are omitted from the result.
• optional[...] and Rule | None captures are omitted from .as_dict() when absent.
• Repeats of terminals become strings; repeats of Rules become lists of hydrated Rule instances.

Type Mixins

Rules can inherit from Python's built-in types (int, str, float, bool) to make parsed values behave like native types:

from turtles import Rule, char, repeat, at_least, sequence

class Integer(Rule, int):
    value: '0' | sequence[char['1-9'], repeat[char['0-9']]]

result = Integer("42")
assert result == 42              # Compares as int
assert isinstance(result, int)   # Type checks pass
assert result + 8 == 50          # Arithmetic works
assert result.as_dict() == 42    # as_dict() returns the int value

# Fields are still accessible
assert result.value == "42"

This is useful when you want parsed numeric or string values to integrate seamlessly with Python code.

Custom Converters

For more complex transformations, define a __convert__ method to transform parsed results into any Python type:

from turtles import Rule, char, repeat, at_least

class Point(Rule):
    x: repeat[char['0-9'], at_least[1]]
    ','
    y: repeat[char['0-9'], at_least[1]]
    
    def __convert__(self):
        return (int(self.x), int(self.y))

result = Point("10,20")
assert result == (10, 20)         # Compares as tuple
assert result.__class__ is tuple  # Type is tuple
assert result.as_dict() == (10, 20)

# Original fields still accessible
assert result.x == "10"
assert result.y == "20"

The converter runs after hydration, so all fields are populated before __convert__ is called. You can convert to any type: tuples, dataclasses, named tuples, custom classes, etc.

from dataclasses import dataclass
from turtles import Rule, char, repeat, at_least

@dataclass
class Coordinate:
    x: int
    y: int

class CoordRule(Rule):
    x: repeat[char['0-9'], at_least[1]]
    ','
    y: repeat[char['0-9'], at_least[1]]
    
    def __convert__(self):
        return Coordinate(int(self.x), int(self.y))

result = CoordRule("5,10")
assert result == Coordinate(5, 10)

Parse errors

Turtles automatically outputs user-friendly modern-style error messages whenever an input fails to parse

from turtles import ParseError

try:
    Row("not_a_kv_pair")
except ParseError as e:
    print(e)

Example output:

Error: incomplete KV: missing key

    ╭─[test.py:1:1]
  1 | not_a_kv_pair
    · ┬           ╱╲
    · │           ╰─ expected "="
    · ╰─ Row started here
    ╰───
  help: The input appears incomplete. Try adding "=".

Example grammars

The turtles/examples/ directory contains complete grammar examples:

The test suite is also a great source of patterns:

Contributions welcome! Open a PR with new example grammars.

More Examples

Semantic Versions

from turtles import Rule, repeat, char, separator, sequence, at_least

class NumId(Rule):
    id: '0' | sequence[char['1-9'], repeat[char['0-9']]]

class Id(Rule):
    id: repeat[char['a-zA-Z0-9-'], at_least[1]]

class Prerelease(Rule):
    "-"
    ids: repeat[Id, separator['.'], at_least[1]]

class Build(Rule):
    "+"
    ids: repeat[Id, separator['.'], at_least[1]]

class SemVer(Rule):
    major: NumId
    "."
    minor: NumId
    "."
    patch: NumId
    prerelease: Prerelease | None
    build: Build | None


result = SemVer('1.2.3-alpha+build.5')

assert result.major.id == '1'
assert result.minor.id == '2'
assert result.patch.id == '3'
assert result.prerelease.ids[0].id == 'alpha'
assert result.build.ids[0].id == 'build'
assert result.build.ids[1].id == '5'

Toy JSON Parser

from turtles import Rule, char, repeat, at_least, separator

class Whitespace(Rule):
    repeat[char[' \t\n\r']]

class Comma(Rule):
    Whitespace
    ','
    Whitespace

class JNull(Rule):
    "null"

class JBool(Rule):
    value: "true" | "false"

class JNumber(Rule):
    value: repeat[char['0-9'], at_least[1]]

class JString(Rule):
    '"'
    value: repeat[char["A-Za-z0-9 !#$%&'()*+,-./:;<=>?@^_`{|}~"]]
    '"'

class JArray(Rule):
    '['
    Whitespace
    items: repeat[JSONValue, separator[Comma]]
    Whitespace
    ']'

class Pair(Rule):
    key: JString
    Whitespace
    ':'
    Whitespace
    value: JSONValue

class JObject(Rule):
    '{'
    Whitespace
    pairs: repeat[Pair, separator[Comma]]
    Whitespace
    '}'

# Rule union - JSONValue can be any of these types
JSONValue = JNull | JBool | JNumber | JString | JArray | JObject


src = '{ "A": { "a": null }, "B": [ true, false, 1, 2, 3 ] }'
result = JSONValue(src)

assert isinstance(result, JObject)
assert len(result.pairs) == 2
assert result.pairs[0].key.value == "A"
assert isinstance(result.pairs[1].value, JArray)

# Tree visualization
print(repr(result))

# Convert to plain Python containers
result.as_dict()

Note: This is a simplified grammar. See turtles/examples/json.py for a complete RFC 8259 implementation with floats, escapes, and full unicode support.

DSL Reference

Backend

Turtles uses a GLL (Generalized LL) parser backend. GLL is a general parsing algorithm for arbitrary context-free grammars, including grammars with ambiguity and left recursion, while still keeping the implementation reasonably small.

At a high level, parsing works like this:

1. The Rule class body is inspected and compiled into a context-free grammar.
2. The GLL parser runs against the input and produces a compact shared parse forest.
3. Turtles extracts a parse tree (with optional disambiguation rules like precedence/associativity) and hydrates it back into instances of your Rule classes.

Read more: https://dotat.at/tmp/gll.pdf

Looking for the old Turtles?

⚠️ The turtles project has been rebooted. v2.0.0 and onward will not be compatible with the original v1.0.0 release. If you are looking for the original project, see Roguelazer/turtles.