msgspec-toon

TOON (Token-Oriented Object Notation) is a line-oriented text format that reduces repeated keys in uniform JSON-shaped data.

msgspec is a high-performance serialization and validation library whose typed Struct objects are a faster, lighter alternative to Pydantic models.

msgspec-toon is a native TOON 4.1 codec for Python. It decodes TOON text directly into msgspec.Struct objects. It does not build an intermediate dict and list tree.

Use it when tabular data must fit in a language model context window, but the application still needs typed Python objects and fast in-process conversion.

This is a beta release. The project passes the pinned TOON 4.1.1 corpus, but it does not yet support every type that msgspec.json supports.

Install

Install the public beta from PyPI:

uv add msgspec-toon

The package requires Python 3.13 or newer. Its only runtime dependency is the exact pin msgspec==0.21.1. Benchmark codecs and tokenizers are optional development dependencies. They are not installed with the library.

Decode TOON into a Struct

import msgspec
import msgspec_toon as toon


class Metadata(msgspec.Struct, frozen=True):
    alias: str
    region: str


class Worker(msgspec.Struct, frozen=True):
    pid: int
    provider: str
    metadata: Metadata


class Document(msgspec.Struct, frozen=True):
    workers: list[Worker]


wire = b"""workers[2]{pid,provider,metadata{alias,region}}:
  9007199254740993,claude,worker-a,west
  80916,claude,worker-b,east"""

decoder = toon.Decoder(Document)
document = decoder.decode(wire)

assert isinstance(document, Document)
assert document.workers[0].pid == 9007199254740993
assert toon.encode(document) == wire

The parser uses the target type while it reads the input. It constructs the final Struct objects directly. The G2 allocation proof records zero temporary built-in dictionaries and lists for this path.

Read and write TOON files

The codec accepts bytes, bytearray, memoryview, or str. Encoding returns bytes, so normal Python file APIs work without an adapter.

from pathlib import Path

import msgspec_toon as toon


source = Path("workers.toon")
target = Path("workers-copy.toon")

decoder = toon.Decoder(Document)
encoder = toon.Encoder()

value = decoder.decode(source.read_bytes())
target.write_bytes(encoder.encode(value))

The conversion itself does not open files, sockets, or subprocesses. Your application controls all I/O.

Use untyped values

Omit type when you need normal Python dictionaries and lists:

value = toon.decode(b"name: ada\nactive: true")
wire = toon.encode(value)

The public surface follows msgspec.json where support exists:

• encode and decode
• reusable Encoder and Decoder
• enc_hook and dec_hook
• strict decoding by default
• msgspec-compatible encode, decode, and validation errors

The encoder can use one through 16 spaces per indentation level. The decoder default is two spaces. If you select another width, the producer and consumer must use the same setting:

wire = toon.encode(value, delimiter="\t", indent=1)
toon.decode(wire, indent_size=1)

If they do not match, the decode error reports the observed leading-space count and tells the caller to pass the matching indent_size. Using indent=1 can reduce tokens in nested documents, but it makes the width an out-of-band pipeline setting. Payload shape still decides whether TOON uses fewer tokens than compact JSON.

Decode accepts a leading UTF-8 byte order mark, Windows CRLF line endings, and a final newline. These inputs normalize only while reading; encoding does not emit a byte order mark.

The functional and reusable decoders both accept float_hook:

from decimal import Decimal

value = toon.decode(b"1.25", float_hook=Decimal)
decoder = toon.Decoder(float_hook=Decimal)

assert value == decoder.decode(b"1.25") == Decimal("1.25")

Typed encode and decode support these msgspec-native values:

• datetime, date, time, and timedelta
• UUID
• Decimal
• string Enum and integer Enum

The encoder normalizes these values before it calls enc_hook. Decimal values keep all their digits and trailing zeroes. Typed decode constructs each value without a built-in container tree.

from decimal import Decimal
from uuid import UUID

toon.encode(Decimal("1.2300"))
# b'"1.2300"'

toon.encode(Decimal("1.2300"), decimal_format="number")
# b'1.2300'

toon.encode(UUID("12345678-1234-5678-1234-567812345678"), uuid_format="hex")
# b'"12345678123456781234567812345678"'

TOON 4.1 uses one canonical spelling for whole numbers. Thus, the encoder writes 1.0 as 1. It writes -0.0 as 0, as required by the official fixtures.

An untyped decode returns an int for these values. A typed float decode returns a float, but it cannot recover the sign of -0.0 because the sign is not on the wire.

The codec does not implement sorted or deterministic output. Both encoder entry points raise NotImplementedError for these order values. They never ignore an accepted option.

Typed decode supports recursive Structs, array-like Structs, tagged Struct unions in both object and positional forms, object as an open value, unions of bool, int, float, and str, native scalar values, and permissive scalar conversion. object uses the same requested open-value path as Any; it does not add an intermediate tree around a typed value. Scalar unions select an exact wire category before a widening conversion, matching msgspec when types overlap.

For a tagged array_like Struct, position zero is the discriminator and the declared fields follow it. A Struct union must use one shape throughout: mixing object-form and array-like Struct variants fails during plan construction. Set strict=False to accept the same bool, integer, and float string conversions as msgspec 0.21.1. Strict mode stays the default.

Other multi-member unions remain explicit plan errors. Use a tagged Struct union for object variants. Use object or Any when the value shape is intentionally open.

Non-string mapping keys are intentionally rejected. See the mapping-key policy.

Why not wrap another TOON codec?

A wrapper must first convert a Struct into built-in containers. Typed decode must parse a built-in tree and then call msgspec.convert. Those extra trees can cost more than the codec work.

TOON 4 nested field groups are important. They let a uniform nested record use one tabular header:

workers[2]{pid,provider,metadata{alias,region}}:
  20324,claude,worker-a,west
  80916,claude,worker-b,east

Older encoders can fall back to a larger entry form for the same data.

Tokens and speed

The figure shows absolute end-to-end time against absolute o200k_base token count. Pareto status is calculated separately for each payload shape and record count. Lines connect the same implementation across record counts. They show workload scaling, not an unmeasured continuous Pareto curve.

The generated benchmark report also publishes the detailed axes:

• Direct encode, decode, and total time for each measured codec.
• Absolute token counts, including compact JSON, under tiktoken o200k_base.

The report crosses four payload shapes with four record counts. Canonical TOON uses more tokens than compact JSON for the measured irregular shapes.

All timing rows come from one session and one release build. The estimator is the mean across ten independent worker processes. It never reports the minimum. Against v0.2.0b5, the 0.3.0b1 release guard found entry decode 9-24% faster, keyed decode 9-11% faster, and entry encode 11-15% faster. Its 2-8% untyped decode gain covered only uniform tabular records; it missed slower nested and irregular records. 0.3.0b2 adds both shapes to the permanent gate. Against 0.3.0b1, nested records are neutral at -1.1% with a 2.1% minimum detectable effect, and irregular records are 9.0% faster.

0.3.0b3 also varies distinct ordinary-record key count while record count and encoded size stay fixed. It replaces the 0.3.0b2 cache scan with hash lookup. Against 0.3.0b2, decode is 35.8% faster at 32 keys and 87.4% faster at 512 keys. The complete 100-point guard reports no reproduced slowdown. These are direct time comparisons, not ratios to an arbitrary reference row. The raw evidence is in the conformance/report.json file.

Conformance and safety

• All 538 pinned TOON 4.1.1 fixtures pass in both directions.
• Typed decode creates no intermediate built-in container tree.
• Integers keep Python precision and do not route through float.
• Errors contain coordinates and static messages, never input payload text.
• Malformed input must return an error. It must not panic or terminate Python.
• Canonical output is byte-locked by tests.

The generated support matrix in the conformance/report.json file lists supported, rejected, and not-yet-supported msgspec features. Unsupported behavior fails clearly. It does not silently return a different value.

Optional msgspec Struct fast path

The stock package reads Struct fields through msgspec's public Python attributes. This is the compatible path for msgspec==0.21.1.

The repository also contains a versioned Struct-access capsule proposal for msgspec. You can build the same codec against that patch in an isolated environment:

make fastpath-build
make fastpath-check
make fastpath-bench
.venv-fastpath/bin/python

This workflow fetches a hash-pinned msgspec commit, applies the preserved patch, and builds both release wheels. It does not modify the normal .venv. The build fails unless the capsule path is active. Published wheels do not depend on the unreleased API.

Develop and reproduce

Use uv for all Python environment work:

uv sync --locked                         # library and developer tools
make build                              # release extension in .venv
make check                              # Rust and Python checks
make qualify                            # canonical release gate and evidence
uv run python conformance/run.py        # pinned 538-fixture corpus
make g2                                 # allocation proof in a separate build

uv sync --group bench --locked           # opt in to benchmark packages
make bench                              # same-run codec and typed ladders
make public-report                      # raw JSON, R charts, and BENCHMARKS.md

make public-report uses the host Rscript, ggplot2, jsonlite, and scales. It does not install R or add R packages to the Python environment. The release guide documents installed-artifact verification and Trusted Publishing.

License

MIT