perf(pydantic): reuse TypeAdapters instead of building one per payload

[PranavMishra28]

Closes #1695.

@tconley1428 you said on the issue that a small LRU cache by default seemed reasonable and asked whether there's any drawback to enabling it by default. This implements that, and the two drawbacks I could find are handled below rather than left implicit.

Problem

PydanticJSONPlainPayloadConverter.from_payload called TypeAdapter(hint) on every payload. Pydantic caches the core schema on the class for BaseModel subclasses, so that's cheap for plain models — but activity and workflow hints frequently aren't classes (Annotated discriminated unions, list[Union[...]], generics), and for those TypeAdapter.__init__ rebuilds the schema every time.

Reproduced @ZacharyHampton's shape before touching anything:

build+validate per payload: 0.1375 ms
validate only:              0.0015 ms
ratio:                      92x

(Issue reported 57× on their box and 521× for pydantic-ai's CallToolResult; same effect, machine-dependent multiple.)

Measured through the converter after the change: 0.0058 ms per call, 199 cache hits to 1 miss.

The two drawbacks, and what this does about them

1. The cache holds type hints strongly. So it's bounded — maxsize=256. A worker sees a small fixed set of activity and workflow signatures, so that covers real usage while stopping a pathological caller from retaining unbounded types. Happy to change the number or expose it if you'd rather it be configurable.

2. Not every hint is hashable, so not every hint can be a cache key — Annotated[list[X], SomeUnhashableMetadata()] raises TypeError from lru_cache. That falls back to per-call construction instead of failing the conversion, with a test covering it. This is the case I'd most want a second opinion on, since it's the one that would turn a working conversion into an error if handled naively.

Caching is keyed on the hint alone, which is safe here because to_json_options only affects serialization, not from_payload.

How tested

Five cases in tests/contrib/pydantic/test_pydantic.py:

• test_repeated_conversion_builds_one_type_adapter — black-box: counts TypeAdapter constructions across two conversions of one hint. Against the unfixed converter it reports built 2 TypeAdapters for one hint. This is the one that pins behavior; the others below assert on cache internals.
• reuse across payloads (1 miss, 4 hits), distinct hints not sharing an adapter, the unhashable-hint fallback, and the type_hint is None → Any path.

tests/contrib/pydantic/ 21 passed. ruff check, ruff format --check, and mypy clean on the changed files.

One note on scope: I did not touch to_payload. SchemaSerializer construction there is already hoisted into __init__, so it doesn't have this problem.

Developed with Claude Code; reviewed and tested by Pranav before marking ready for review.

[CLAassistant]

All committers have signed the CLA.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 69582a3948

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[tconley1428]

There's already an open PR with ongoing discussion. #1703