Workflow sandbox re-imports annotated_types, so pydantic silently drops constraints

[anthonypalmeira]

Summary

SandboxRestrictions.passthrough_modules_with_temporal passes pydantic and, since 1.33.0 (#1834), pydantic_core through, but not annotated_types. pydantic imports annotated_types inside functions, so each sandbox gets its own copy, and pydantic's lru_cached class → constraint map (_get_at_to_constraint_map) only knows the copy that filled it first. A constraint built from the other copy falls through to "ignore any other unknown metadata", with no error:

• The host builds a constrained model first: a workflow's Annotated[int, annotated_types.Ge(1)] is not enforced.
• A workflow builds one first: host-side Field(ge=0) and conint(ge=0) are not enforced for the rest of the process, activities included.

Each sandbox also pays a fresh import: a median 4.7 ms here, against 0.003 ms passed through. When that import happens during an activation, the sandbox warns Module annotated_types was imported after initial workflow load.

Repro

temporalio 1.33.0, pydantic 2.13.4, annotated-types 0.7.0, Python 3.12.3, macOS 26 arm64. It drives the sandbox's Importer the way tests/worker/workflow_sandbox/test_importer.py does.

"""pydantic drops constraints built from the sandbox's annotated_types copy, or the host's."""

import subprocess
import sys
import tempfile
from pathlib import Path

MODELS = """
from typing import Annotated
import annotated_types
from pydantic import BaseModel

class Positive(BaseModel):
    x: Annotated[int, annotated_types.Ge(1)]
"""


def case(first, passthrough):
    from pydantic import BaseModel, Field, ValidationError
    from temporalio.worker.workflow_sandbox import SandboxRestrictions
    from temporalio.worker.workflow_sandbox._importer import Importer
    from temporalio.worker.workflow_sandbox._restrictions import RestrictionContext

    restrictions = SandboxRestrictions.default
    if passthrough:
        restrictions = restrictions.with_passthrough_modules("annotated_types")

    def in_sandbox():
        with Importer(restrictions, RestrictionContext()).applied():
            import sandboxed_models
        return sandboxed_models.Positive

    def on_host():
        class Host(BaseModel):
            n: int = Field(ge=0)

        return Host

    if first == "host":
        host, sandboxed = on_host(), in_sandbox()
    else:
        sandboxed, host = in_sandbox(), on_host()

    def result(model, **values):
        try:
            model(**values)
        except ValidationError:
            return "enforced"
        return "DROPPED"

    return f"sandbox Ge(1): {result(sandboxed, x=0)}, host ge=0: {result(host, n=-1)}"


if __name__ == "__main__":
    if len(sys.argv) == 4:  # one case per fresh process: pydantic's cache is process-wide
        sys.path.insert(0, sys.argv[3])
        print(case(sys.argv[1], sys.argv[2] == "passthrough"))
    else:
        models = tempfile.mkdtemp()
        Path(models, "sandboxed_models.py").write_text(MODELS)
        for first in ("host", "sandbox"):
            for mode in ("default", "passthrough"):
                out = subprocess.run(
                    [sys.executable, __file__, first, mode, models],
                    capture_output=True,
                    text=True,
                    check=True,
                ).stdout.strip()
                print(f"{first + ' first,':<14} {mode + ':':<13} {out}")

host first,    default:      sandbox Ge(1): DROPPED, host ge=0: enforced
host first,    passthrough:  sandbox Ge(1): enforced, host ge=0: enforced
sandbox first, default:      sandbox Ge(1): enforced, host ge=0: DROPPED
sandbox first, passthrough:  sandbox Ge(1): enforced, host ge=0: enforced

Proposal

Add "annotated_types" next to "pydantic" and "pydantic_core" in passthrough_modules_with_temporal. It is a pure-Python module of constraint dataclasses that imports only the standard library and typing_extensions, with no I/O, time or randomness. temporalio's own StrandsPlugin and GoogleGenAIPlugin already pass it through; other users need a runner override such as SandboxRestrictions.default.with_passthrough_modules("annotated_types").

I can open the PR: the one-line change, a CHANGELOG entry, and a regression test in tests/worker/workflow_sandbox/test_importer.py built from the repro.

[tconley1428]

That seems reasonable, go ahead.