What

transform_schema (the structured-output schema transformer used by messages.parse / beta.messages.parse) raised

ValueError: Schema must have a 'type', 'anyOf', 'oneOf', or 'allOf' field.

for any sub-schema that pydantic emits as an empty {} — namely an Any-typed field, or the items of an untyped list / List[Any].

Repro

import pydantic
from typing import Any, List
import anthropic

class Result(pydantic.BaseModel):
    summary: str
    tags: List[Any]   # also: a bare `list`, or any `Any`-typed field

client = anthropic.Anthropic(api_key="sk-ant-...")
client.beta.messages.parse(
    model="claude-sonnet-4-5",
    max_tokens=100,
    messages=[{"role": "user", "content": "hi"}],
    output_format=Result,
)
# -> ValueError: Schema must have a 'type', 'anyOf', 'oneOf', or 'allOf' field.

The error is raised at request-build time, before any API call. Notably, an untyped dict / Dict[str, Any] field does not trip it (it keeps type: object), so the failure is specific to untyped lists and Any fields — surprising and easy to hit.

Root cause: in the array branch, items: {} recurses into transform_schema({}), and a schema with no type/anyOf/oneOf/allOf hit the raise. The same happens for an Any object property.

Change

An empty {} is a valid, unconstrained ("any") JSON Schema. Instead of raising when there is no type and no combinator, leave the sub-schema untyped (permissive). One-line behavior change in src/anthropic/lib/_parse/_transform.py.

Tests

Added to tests/lib/_parse/test_transform.py:

• empty {} schema
• untyped array items ({"type": "array", "items": {}})
• an Any object property
• end-to-end pydantic model with List[Any] + Any field (regression for the original crash)

Verification

• uv run pytest tests/lib/_parse → 26 passed
• uv run ruff check → all checks passed
• uv run pyright (strict) on changed files → 0 errors, 0 warnings