GitHub - janbjorge/gyml: gyml

5 min read Original article ↗

YAML syntax. JSON semantics. Zero drama.


The problem

YAML's spec contains 211 grammar productions across ten chapters. It has five scalar styles, five ways to write null, anchors, aliases, tags, and implicit type coercion rules that vary by implementation. The Norway Problem is a well-known example: NO is a valid YAML 1.1 boolean (False), which surprises anyone using it as a country code or similar string value.

GYML is a strict subset of YAML that trades that flexibility for predictability. It keeps YAML's block indentation syntax and applies JSON's type semantics: one spelling per type, no implicit coercion surprises, no anchors or tags. What you write is what you get.


What GYML is

A strict subset of YAML. Everything valid GYML is also valid YAML. The reverse is absolutely not true, which is entirely the point.

The rules fit on a sticky note:

  • One way to write each type. true/false for booleans. null for null. Decimal integers and floats. Double-quoted strings. Done.
  • Block style only. No flow mappings {a: 1}, no flow sequences [a, b]. Empty {} and [] are fine as explicit empty literals.
  • No dark arts. Anchors (&), aliases (*), and tags (!!) are rejected at the lexer. Your config file is not a program.
  • Duplicate keys are a hard error. Silent overwrites have caused too many bugs in too many codebases.
  • Strict indentation. Multiples of 2 spaces. Tabs rejected. No exceptions.

Install

# pip
pip install gyml

# uv (recommended)
uv add gyml

Usage

Parse a string

from gyml import loads

config = loads("""
server:
  host: 0.0.0.0
  port: 8080
  debug: false
  workers: 4
""")

print(config["server"]["port"])   # 8080  (int, not string)
print(config["server"]["debug"])  # False (bool, not string)

Parse a file

from gyml import load

config = load("config.gyml")

Both functions return plain Python objects: dict, list, str, int, float, bool, or None. No wrapper types. No schema required. What you see is what you get.

Error handling

Errors are precise. When something is wrong you get the exact line, column, and a message that tells you how to fix it, not a stack trace pointing at a C extension.

from gyml import loads, ParseError

try:
    loads("port: 0xFF")
except ParseError as e:
    print(e)
    # line 1, col 7: "0xFF" -- hex/octal/binary literals are not allowed

ParseError exposes three attributes:

Attribute Type Description
message str Human-readable description of the violation
line int 1-based line number in the source
col int 1-based column number in the source

CLI

Convert any .gyml file to pretty-printed JSON:

gyml config.gyml                  # pretty-print to stdout
gyml config.gyml | jq '.server'   # pipe into jq
gyml config.gyml > out.json       # write to file

What valid GYML looks like

# All scalar types
name: Alice
age: 30
score: 9.5
active: true
notes: null

# Double-quoted strings preserve reserved words as strings
status: "true"
id: "404"

# Nested mapping
database:
  host: localhost
  port: 5432
  ssl: true
  name: mydb

# Sequence of scalars
allowed_hosts:
  - localhost
  - 127.0.0.1
  - 10.0.0.0

# Sequence of mappings (dash alone on its line, mapping indented below)
users:
  -
    name: Alice
    role: admin
  -
    name: Bob
    role: viewer

# Empty collection literals
cache: {}
tags: []

# Deeply nested is fine
services:
  web:
    image: nginx
    ports:
      - 80
      - 443
  db:
    image: postgres
    env:
      POSTGRES_DB: mydb
      POSTGRES_USER: admin

What GYML rejects (and why)

Every one of these is valid YAML. Every one of them has caused a real bug in someone's production system.

Input Error Why it's banned
country: NO not a valid boolean Norway Problem; use "NO"
enabled: yes not a valid boolean use true
level: on not a valid boolean use true
value: ~ not a valid null use null
value: NULL not a valid null use null
port: 0xFF hex/octal/binary not allowed no implicit base conversion
ratio: .inf .inf/.nan not allowed not a JSON value
count: 1_000 underscore separators not allowed ambiguous in plain scalars
key: 'val' single-quoted strings not allowed one quoting style, double
true: val boolean literals not allowed as keys use "true"
a: 1\na: 2 duplicate key silent overwrites are bugs
key:\n bare empty value write null explicitly
data: &anchor val anchors not allowed configs aren't programs
ref: *anchor aliases not allowed configs aren't programs
obj: !!python/object tags not allowed absolutely not


Get going

# clone
git clone https://github.com/janbjorge/gyml.git
cd gyml

# install deps (uv required)
uv sync

# run the tests
uv run pytest

# lint + format check
uv run ruff check gyml/ tests/
uv run ruff format --check gyml/ tests/

# type-check
uv run ty check gyml/

All four must pass clean before any change is considered done.


Contributing

Contributions are welcome. A few ground rules:

  • No runtime dependencies. The whole value proposition is zero deps. Don't add any third-party libraries.
  • No Any. The codebase is fully typed. Keep it that way.
  • Tests for everything. New behaviour gets a test. Bug fixes get a regression test. Use ok(), ok_eq(), and fail() from conftest.py.
  • Read AGENTS.md before touching anything. It documents the architecture, naming conventions, and what the lexer vs. parser is responsible for.
  • Ruff and ty must be clean. Run uv run ruff check gyml/ tests/ and uv run ty check gyml/ before opening a PR. CI will reject anything that isn't.

Open an issue first if you're planning something bigger than a bug fix. It's worth a quick conversation before spending time on an approach that might not fit the design.