msgspec Skill

msgspec is a high-performance Python library for serialization, deserialization, and typed validation. This guidance targets the 0.22.0 release.

Code Style Rules

• Use PEP 604 for unions: T | None (not Optional[T])
• from __future__ import annotations rule — Library/shared modules that define runtime-introspected msgspec.Struct subclasses should avoid postponed annotations unless the consuming tool resolves them. Consumer modules that only use Structs MAY use future annotations.
• Annotate every serialized field; only annotated attributes become Struct fields
• Use kw_only=True for Structs with more than 2 fields
• Put wire-name configuration on msgspec.field(name=...) or the Struct's rename= option; msgspec.Meta defines constraints and JSON Schema metadata, not field aliases

Quick Reference

Struct Definition and Options

import msgspec


# Basic struct
class User(msgspec.Struct):
    id: int
    name: str
    email: str | None = None


# Performance options
class Event(msgspec.Struct, frozen=True, gc=False, cache_hash=True):
    """frozen=True: immutable + hashable. gc=False: skip GC tracking."""

    event_type: str
    payload: dict[str, object]


# Keyword-only + omit defaults on wire (recommended for >2 fields / compact payloads)
class Config(msgspec.Struct, kw_only=True, omit_defaults=True, repr_omit_defaults=True):
    host: str
    port: int = 5432
    ssl: bool = False


# Array-like encoding (tuple encoding, more compact)
class Point(msgspec.Struct, array_like=True):
    x: float
    y: float


# Rename fields for serialization ("lower", "upper", "camel", "pascal", "kebab", callable, or mapping)
class ApiResponse(msgspec.Struct, rename="camel"):
    user_id: int  # serialized as "userId"
    created_at: str  # serialized as "createdAt"


# Per-field wire alias and default_factory via msgspec.field()
class Resource(msgspec.Struct):
    resource_id: int = msgspec.field(name="id")
    tags: list[str] = msgspec.field(default_factory=list)


# Reject unknown fields at API boundaries
class StrictInput(msgspec.Struct, forbid_unknown_fields=True):
    name: str
    value: int

Struct / defstruct Option	Default	Purpose
kw_only	False	Make all fields keyword-only in __init__
frozen	False	Immutable instances; adds __hash__
cache_hash	False	Precompute and cache hash on frozen=True Structs
gc	True	Set False to omit Cyclic GC header on short-lived, non-circular Structs
omit_defaults	False	Omit fields equal to their default (or empty built-in list/dict/set/bytearray default_factory; custom default_factory callables are never omitted) when encoding or calling to_builtins()
repr_omit_defaults	False	Omit fields equal to their default in repr()
forbid_unknown_fields	False	Raise ValidationError on unknown keys during decoding / convert()
rename	None	"lower", "upper", "camel", "pascal", "kebab", Callable[[str], str | None], or Mapping[str, str] (leading _ is preserved for "camel"/"pascal")
array_like	False	Encode/decode Struct as a positional array ([x, y]) instead of an object
tag / tag_field	None	Discriminated union tag (True, False, str, int, or Callable[[str], str | int]) and discriminator key (defaults to "type" when tag is set)
eq / order	True / False	Generate ==/!= and </<=/>/>= comparison methods
weakref / dict	False / False	Add __weakref__ slot or __dict__ attribute storage (neither dict=True nor weakref=True/inherited __weakref__ can be combined with gc=False in 0.22.0+)

Partial Updates with UNSET and UnsetType

Use msgspec.UNSET (msgspec.UnsetType) to distinguish omitted fields from explicit null (None). bool(msgspec.UNSET) is False. Fields set to UNSET are automatically omitted when encoding or calling msgspec.to_builtins().

import msgspec


class UserPatch(msgspec.Struct, kw_only=True, forbid_unknown_fields=True):
    name: str | msgspec.UnsetType = msgspec.UNSET
    email: str | None | msgspec.UnsetType = msgspec.UNSET


patch = msgspec.json.decode(b'{"email": null}', type=UserPatch)
assert patch.name is msgspec.UNSET  # omitted by client
assert patch.email is None  # explicitly cleared by client
assert msgspec.json.encode(patch) == b'{"email":null}'

Validation Constraints

from datetime import datetime
from typing import Annotated

import msgspec
from msgspec import Meta


class Product(msgspec.Struct):
    name: Annotated[str, Meta(min_length=1, max_length=100)]
    price: Annotated[float, Meta(gt=0)]
    quantity: Annotated[int, Meta(ge=0, le=10_000)]
    sku: Annotated[str, Meta(pattern=r"^[A-Z]{2}-\d{4}$")]
    batch_size: Annotated[int, Meta(multiple_of=5)]
    expires_at: Annotated[datetime, Meta(tz=True)]


# Reusable constraint aliases
PositiveInt = Annotated[int, Meta(gt=0)]
NonEmptyStr = Annotated[str, Meta(min_length=1)]
Percentage = Annotated[float, Meta(ge=0.0, le=100.0)]


class Order(msgspec.Struct):
    id: PositiveInt
    label: NonEmptyStr
    discount: Percentage = 0.0

Serialization (json, msgpack, yaml, toml) and Raw

import msgspec

# JSON -- singleton encoder/decoder (cache these!)
# Encoder kwargs: enc_hook=None, decimal_format="string"|"number"|Callable[[Decimal], Any], uuid_format="canonical"|"hex", order=None|"deterministic"|"sorted"
# Decoder kwargs: type=Any, *, strict=True, dec_hook=None, float_hook=None
encoder = msgspec.json.Encoder()
decoder = msgspec.json.Decoder(User)

data = encoder.encode(user)  # bytes
user = decoder.decode(b'{"id":1,"name":"Alice"}')

# Newline-delimited JSON (NDJSON) and zero-allocation buffer writes
ndjson = encoder.encode_lines([user])
users = decoder.decode_lines(ndjson)
buf = bytearray()
encoder.encode_into(user, buf, 0)
pretty = msgspec.json.format(data, indent=2)

# Functional API (convenience, slightly slower)
data = msgspec.json.encode(user, order="deterministic")
user = msgspec.json.decode(data, type=User)

# MessagePack (binary, more compact; uuid_format also accepts "bytes"; Decoder accepts ext_hook and msgspec.msgpack.Ext)
mp_data = msgspec.msgpack.encode(user)
user = msgspec.msgpack.decode(mp_data, type=User)

# YAML (requires msgspec[yaml] / PyYAML) and TOML (requires msgspec[toml] / tomli_w + tomllib/tomli)
yaml_bytes = msgspec.yaml.encode(user)
user_from_yaml = msgspec.yaml.decode(yaml_bytes, type=User)
toml_bytes = msgspec.toml.encode(user)
user_from_toml = msgspec.toml.decode(toml_bytes, type=User)


# Deferred decoding / zero-copy payload slicing with msgspec.Raw
class Envelope(msgspec.Struct):
    kind: str
    payload: msgspec.Raw


env = msgspec.json.decode(b'{"kind":"user","payload":{"id":1,"name":"Alice"}}', type=Envelope)
detached_raw = env.payload.copy()  # detach from input buffer if retaining long-term
inner_user = msgspec.json.decode(detached_raw, type=User)


# Hooks are only for unsupported custom types. datetime, date, time, timedelta,
# UUID, Decimal, and Enum are already supported natively.
def enc_hook(obj: object) -> object:
    if isinstance(obj, complex):
        return (obj.real, obj.imag)
    raise NotImplementedError(f"Unsupported type: {type(obj)}")


def dec_hook(target_type: type, obj: object) -> object:
    if target_type is complex:
        real, imag = obj
        return complex(real, imag)
    raise NotImplementedError(f"Unsupported type: {target_type}")


custom_encoder = msgspec.json.Encoder(enc_hook=enc_hook)
custom_decoder = msgspec.json.Decoder(MyStruct, dec_hook=dec_hook)

dec_hook runs only for unsupported custom annotations. TypeError and ValueError raised by the hook become path-aware ValidationErrors. ValidationError and DecodeError raised by the hook propagate directly without being wrapped in another ValidationError (0.21.1+).

Exception hierarchy: MsgspecError(Exception) is the base class for EncodeError and DecodeError. ValidationError is a subclass of DecodeError (ValidationError -> DecodeError -> MsgspecError). Always catch ValidationError before DecodeError when distinguishing schema validation errors from malformed wire syntax.

Canonical Litestar serializers (match-your-stack)

Litestar apps typically need to_json(value, as_bytes=True) that handles UUID / datetime / Enum / Decimal for Channels broadcasts, log contexts, and JSONB writes. Pick the branch that matches your project.

Branch A — sqlspec is in-stack. Re-export sqlspec's serializer; it already installs an enc_hook covering UUID, datetime, Enum, Decimal, Pydantic, dataclasses, attrs, and msgspec.Struct.

# myapp/utils/serialization.py
from sqlspec.utils.serializers import from_json, to_json

__all__ = ("from_json", "to_json")

Usage:

from myapp.utils.serialization import to_json

payload = to_json(order, as_bytes=True)
await backend.publish(payload, channels=[f"orders:{order.id}:events"])

Branch B — sqlspec is not in-stack. Use a plain msgspec Encoder; the package natively handles UUID, datetime, date, time, timedelta, Decimal, Enum, dataclasses, attrs classes, and Structs.

# myapp/utils/serialization.py
from typing import Any

import msgspec


_encoder = msgspec.json.Encoder()


def to_json(value: Any) -> bytes:
    if isinstance(value, bytes):
        return value
    return _encoder.encode(value)

Type Coercion with convert() and to_builtins()

import msgspec

raw = {"id": "42", "name": "Alice"}  # id is a string

# Strict mode (default): raises on type mismatch
user = msgspec.convert(raw, User)  # ValidationError: id must be int

# Lax mode: coerces compatible types (e.g. str -> int, 0/1 -> bool, epoch int/float -> datetime)
user = msgspec.convert(raw, User, strict=False)  # id coerced to 42

# str_keys: dict keys are strings (useful for JSON/TOML-loaded dicts)
data = {"1": "Alice", "2": "Bob"}
result = msgspec.convert(data, dict[int, str], str_keys=True)

# Convert with dec_hook or custom builtin_types passthrough
measurement = msgspec.convert(raw_measurement, Measurement, dec_hook=dec_hook, builtin_types=(bytes,))

# Convert a dataclass, attrs instance, or ORM model to a Struct by reading attributes
from dataclasses import dataclass


@dataclass
class LegacyUser:
    id: int
    name: str


legacy = LegacyUser(id=1, name="Alice")
user = msgspec.convert(legacy, User, from_attributes=True)

# Convert a Struct or supported type graph back to Python builtin types
builtins_dict = msgspec.to_builtins(user, str_keys=True, order="deterministic")

from_attributes=False is the default. Plain mappings convert to object-like output types without this option; dataclass, attrs, ORM, and other objects require from_attributes=True. msgspec.Raw instances pass through convert() when the target field/type is msgspec.Raw; to parse a Raw buffer into a Struct, use msgspec.json.decode(raw, type=Struct) (or msgpack.decode).

Struct Helpers (msgspec.structs) and Inspection (msgspec.inspect)

import copy
import msgspec

# Shallow conversions (keyed by Python attribute names; preserves nested Structs and UNSET values)
field_map = msgspec.structs.asdict(user)
field_tuple = msgspec.structs.astuple(user)

# Copy-and-replace (both call __post_init__ as of msgspec 0.21.0)
updated = msgspec.structs.replace(user, name="Bob")
updated_std = copy.replace(user, name="Bob")

# Mutate a frozen=True Struct (e.g. inside __post_init__ to normalize a field)
msgspec.structs.force_setattr(event, "event_type", event.event_type.lower())

# Runtime field introspection (FieldInfo: name, encode_name, type, default, default_factory, required)
for f in msgspec.structs.fields(User):
    assert f.required is (f.default is msgspec.NODEFAULT and f.default_factory is msgspec.NODEFAULT)

# TypeGuards for Struct instances and classes (added in msgspec 0.20.0; checks StructMeta)
assert msgspec.inspect.is_struct(user)
assert msgspec.inspect.is_struct_type(User)
struct_info = msgspec.inspect.type_info(User)

Dynamic Struct Creation

import msgspec

# Runtime struct from field definitions (accepts all Struct keyword options plus bases, module, namespace)
fields = [
    ("id", int),
    ("name", str),
    ("score", Annotated[float, Meta(ge=0.0)]),
]
DynamicModel = msgspec.defstruct("DynamicModel", fields, kw_only=True)

# With defaults
fields_with_defaults = [
    ("id", int),
    ("active", bool, True),  # (name, type, default)
]
FlexModel = msgspec.defstruct("FlexModel", fields_with_defaults)

Tagged Unions (Discriminated Unions)

import msgspec


# Default tag field is "type", tag value is the qualified class name
class Dog(msgspec.Struct, tag=True):
    name: str
    breed: str


class Cat(msgspec.Struct, tag=True):
    name: str
    indoor: bool


Animal = Dog | Cat

# Deserialize: inspects "type" field to pick correct class
animal = msgspec.json.decode(b'{"type":"Dog","name":"Rex","breed":"Lab"}', type=Animal)


# Custom tag values
class CreateEvent(msgspec.Struct, tag="create"):
    resource: str


class DeleteEvent(msgspec.Struct, tag="delete"):
    resource: str
    soft: bool = True


Event = CreateEvent | DeleteEvent


# Custom tag field name
class V1Request(msgspec.Struct, tag="v1", tag_field="version"):
    payload: str


class V2Request(msgspec.Struct, tag="v2", tag_field="version"):
    payload: str
    metadata: dict[str, str] = {}


Request = V1Request | V2Request

All Struct variants in a multi-Struct union must be tagged, use the same tag_field, use unique tag values, and use one tag type (str or int) consistently. A union may contain non-Struct types, but it may contain at most one untagged Struct.

Validation and 0.20–0.22 Behavior

• Direct Struct construction trusts the caller and does not enforce field annotations. Typed decode() and convert() perform runtime type and Meta constraint validation.
• msgspec.StructMeta is publicly exposed, and msgspec.inspect.is_struct() / msgspec.inspect.is_struct_type() provide TypeGuard checks for Struct instances and classes (0.20.0+).
• msgspec.structs.replace() and Python's copy.replace() call __post_init__ (0.21.0+).
• ValidationError and DecodeError raised inside dec_hook propagate directly without being double-wrapped in another ValidationError (0.21.1+).
• In 0.22.0, setting gc=False on a Struct with weakref=True (or an inherited __weakref__ slot) raises ValueError (matching gc=False + dict=True).
• In 0.22.0, msgspec.json.Encoder and msgspec.msgpack.Encoder accept a callable for decimal_format (Callable[[Decimal], Any]) in addition to "string" and "number".
• In 0.22.0, Literal[True]/Literal[False], mixed-type Literals (such as Literal[1, None]), PEP 695 generic classes (class Box[T](msgspec.Struct): ...), dict keys using plain str-valued Enums in JSON, overriding an inherited field alias back to its own name, and Python 3.15+ frozendict (msgspec.inspect.FrozenDictType) are supported.
• msgspec.json.schema() and schema_components() accept ref_template="#/$defs/{name}", mark set and frozenset fields with uniqueItems: True, emit JSON-compatible []/"" defaults for default_factory=set/bytearray, place {"type": "null"} last in optional union anyOf, and preserve Meta constraints/metadata in propertyNames for annotated dict keys (0.21–0.22).

<workflow>

Workflow

Step 1: Define Structs

Create msgspec Structs for all data shapes. Use kw_only=True for Structs with more than 2 fields. Use frozen=True for immutable value objects. Use forbid_unknown_fields=True for API-boundary input validation.

Step 2: Add Constraints

Annotate fields with Annotated[Type, Meta(...)] for numeric ranges, string lengths, and regex patterns. Define reusable constraint aliases at module level to avoid repetition.

Step 3: Choose Serialization Strategy

Use msgspec.json for JSON APIs and msgspec.msgpack for binary protocols or internal messaging. Instantiate reusable Encoder/Decoder objects once at module level. Add enc_hook/dec_hook only for unsupported custom types; msgspec natively supports datetime, UUID, Decimal, and Enum.

Step 4: Handle Polymorphism

Use tagged unions (tag=True or tag="value") for discriminated unions. Define a union type alias (Event = CreateEvent | DeleteEvent) and decode against it. Use tag_field to customize the discriminator field name.

Step 5: Validate

Test round-trip encode/decode. Confirm ValidationError is raised for constraint violations. Verify tag dispatch selects the correct Struct type for all union variants.

</workflow> <guardrails>

Guardrails

• Annotate every serialized field -- only annotated attributes become Struct fields.
• Reuse Encoder/Decoder instances -- configured codec objects are designed for repeated calls.
• Use kw_only=True for Structs with >2 fields -- prevents positional argument confusion and makes instantiation self-documenting.
• Use forbid_unknown_fields=True at API boundaries -- rejects payloads with unexpected keys, preventing silent data loss.
• Use Meta for supported field constraints -- typed decode and convert() check these constraints and report the failing path; direct Struct construction does not.
• Never pass rename to Meta -- alias one field with msgspec.field(name=...) or configure the Struct with rename=.
• Avoid non-integral float multiple_of constraints -- binary floating-point precision may reject mathematically valid values; use an integer unit when possible.
• Use gc=False for short-lived, non-circular objects -- eliminates GC overhead for hot-path objects like request/response shapes.
• Tagged unions for polymorphism -- faster than manual dispatch and eliminates isinstance chains.
• from __future__ import annotations rule — Library/shared modules that define runtime-introspected types (advanced-alchemy models, sqlspec configs, msgspec Structs, dishka providers) avoid postponed annotations unless their consumers resolve them. Consumer applications MAY use it. The restriction applies only to modules that define introspected types, not handler/service/test modules that use them.
• Use strict=False only at trust boundaries -- lax coercion is useful for converting legacy dicts but can mask type errors in internal code.
• Prefer sqlspec.utils.serializers.to_json when sqlspec is in-stack — its built-in enc_hook covers UUID, datetime, Enum, Decimal, Pydantic, msgspec.Struct, dataclasses, and attrs in one import. Hand-rolling is only needed when sqlspec is not a dependency.

</guardrails> <validation>

Validation Checkpoint

Before delivering msgspec code, verify:

• All Struct fields have explicit type annotations
• If this library/shared module defines runtime-introspected types, avoid from __future__ import annotations unless all consumers resolve postponed annotations. Consumer modules may use it.
• Encoder/Decoder instances are module-level singletons (not created per-request)
• API-boundary Structs use forbid_unknown_fields=True
• Numeric/string constraints use Meta (not manual if checks)
• Field aliases use msgspec.field(name=...) or Struct rename=; Meta does not accept rename.
• Hooks are used only for unsupported custom types; native datetime/UUID/Decimal/Enum paths do not duplicate built-in handling
• Object-to-Struct conversion uses from_attributes=True
• Tagged union tag values are unique across all variants in a union
• Tagged union variants share one tag_field and one tag value type
• kw_only=True on Structs with more than 2 fields
• If sqlspec is in-stack, to_json is imported from sqlspec.utils.serializers (not hand-rolled)

</validation> <example>

Example

Task: Define an event system with tagged unions, constraints, and JSON serialization.

# Library/shared modules that define runtime-introspected Structs usually avoid postponed annotations.

# events.py
from datetime import UTC, datetime
from typing import Annotated
import uuid

import msgspec
from msgspec import Meta

# --- Constraint aliases ---
NonEmptyStr = Annotated[str, Meta(min_length=1, max_length=255)]
PositiveInt = Annotated[int, Meta(gt=0)]


# --- Event variants (tagged union) ---
class UserCreatedEvent(msgspec.Struct, tag="user.created", tag_field="event_type", kw_only=True, gc=False):
    event_id: uuid.UUID
    user_id: PositiveInt
    email: NonEmptyStr
    occurred_at: datetime


class UserDeletedEvent(msgspec.Struct, tag="user.deleted", tag_field="event_type", kw_only=True, gc=False):
    event_id: uuid.UUID
    user_id: PositiveInt
    occurred_at: datetime
    reason: str | None = None


UserEvent = UserCreatedEvent | UserDeletedEvent

# --- Reusable codec; datetime and UUID are supported natively ---
_encoder = msgspec.json.Encoder()
_decoder = msgspec.json.Decoder(UserEvent)


def encode_event(event: UserEvent) -> bytes:
    return _encoder.encode(event)


def decode_event(data: bytes) -> UserEvent:
    return _decoder.decode(data)


# --- Usage ---
event = UserCreatedEvent(
    event_id=uuid.uuid4(),
    user_id=42,
    email="alice@example.com",
    occurred_at=datetime.now(UTC),
)
payload = encode_event(event)
# b'{"event_type":"user.created","event_id":"...","user_id":42,"email":"alice@example.com","occurred_at":"..."}'

recovered = decode_event(payload)
assert isinstance(recovered, UserCreatedEvent)

</example>

---

References Index

For detailed guides and reference tables, refer to the following documents in references/:

• Meta Constraints Reference -- Full table of all Meta constraint parameters with examples for numeric, string, bytes, and OpenAPI metadata.
• Tagged Union Patterns -- Discriminated union patterns: default tags, custom tag fields/values, nested unions, API versioning, and event systems.
• Litestar Patterns — CamelizedBaseStruct, sqlspec-vs-manual to_json branches, hybrid msgspec + Pydantic schema pattern, __post_init__ validation for Litestar apps.

---

Official References

• https://msgspec.dev
• https://pypi.org/project/msgspec/0.22.0/
• https://github.com/msgspec/msgspec/tree/0.22.0
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/structs.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/constraints.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/supported-types.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/jsonschema.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/converters.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/extending.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/api.rst
• https://github.com/msgspec/msgspec/blob/0.22.0/docs/changelog.md

Shared Styleguide Baseline

• Use shared styleguides for generic language/framework rules to reduce duplication in this skill.
• General Principles
• Python
• Keep this skill focused on tool-specific workflows, edge cases, and integration details.