Backtestspydanticrun-20260930-144322#12978
mypy [arg-type]-error when passing in type that can be converted to RootModel
Our patch was judged on par with the merged fix.
In this judgment, Patch A = our agent, Patch B = the merged fix.
Both patches make the same two code changes in pydantic/mypy.py. First, they register `_RootModelMetaclass` in `get_metaclass_hook`, so mypy's dataclass_transform spec is reset and the plugin handles RootModel classes. Second, they bypass the 'don't generate __init__ if one exists' guard when the class is `pydantic.root_model.RootModel` itself. That lets the plugin-generated `__init__` (untyped unless `init_typed` is set) replace RootModel's hand-written typed `__init__`. This fixes the root cause: `RootModel[float]('1.0')` and its subclasses now follow the same plugin settings as BaseModel fields. The only code difference is `self._cls.info.fullname` (A) versus `self._cls.fullname` (B), which are equivalent for a ClassDef. On tests, both add cases for a parametrized alias (`RootModel[float]`) and a subclass, with expected outputs for all three configs. Under plugin_ini (init_typed unset) there are no errors. Under the default configs the errors remain, so the tests would fail before the change and pass after. A also includes a BaseModel control case (`FloatModel(f='1.0')`) that shows parity with the issue's example. B adds an explanatory comment in the test and a clearer inline comment on why RootModel is excluded. A updates the docstrings. These are cosmetic, offsetting differences; correctness, coverage, and scope are the same.
The judge saw only the two patches, labelled A and B in random order, with issue links and version numbers removed. The reasoning above is shown verbatim.
Solve
43m 38s24 turns
Wall-clock time in the container
Full test suite
No new failures
12,253 passed · 501 skipped · 12,754 total
Contamination audit
Clean
No sign the agent saw the real fix
Patch size
87vs 63 merged
Changed lines, ours vs merged
The issue
View on GitHubInitial Checks
- I confirm that I'm using Pydantic V2
Description
The below example code (named foo.py) passes when running python foo.py as the input "1.0"-string gets converted to float. However, running mypy foo.py (with or without --strict) gives the following error:
foo.py:11: error: Argument 1 to "RootModel" has incompatible type "str"; expected "float" [arg-type]
Found 1 error in 1 file (checked 1 source file)
I would expect the BaseModel-fields and the RootModel-field to behave the same in this case. That is, if the value can be converted – either via a custom made validator or through the conversion table – then mypy should not raise an error.
Example Code
import pydantic
class Foo(pydantic.BaseModel):
f: float
RFoo = pydantic.RootModel[float]
f = Foo(f="1.0")
rf = RFoo("1.0")
assert isinstance(f.f, float)
assert isinstance(rf.root, float)
assert f.f == rf.root
Python, Pydantic & OS Version
pydantic version: 2.12.5
pydantic-core version: 2.41.5
pydantic-core build: profile=release pgo=false
python version: 3.11.14 (main, Feb 3 2026, 22:51:56) [Clang 21.1.4 ]
platform: Linux-5.15.0-173-generic-x86_64-with-glibc2.35
related packages: typing_extensions-4.15.0 mypy-1.19.1
commit: unknownThe two patches
pydantic/mypy.pyindex cd28e700..1026ecf6 100644@@ -88,6 +88,7 @@ CREATE_MODEL_FULLNAME = 'pydantic.main.create_model' BASESETTINGS_FULLNAME = 'pydantic_settings.main.BaseSettings' ROOT_MODEL_FULLNAME = 'pydantic.root_model.RootModel' MODEL_METACLASS_FULLNAME = 'pydantic._internal._model_construction.ModelMetaclass'+ROOT_MODEL_METACLASS_FULLNAME = 'pydantic.root_model._RootModelMetaclass' FIELD_FULLNAME = 'pydantic.fields.Field' DATACLASS_FULLNAME = 'pydantic.dataclasses.dataclass' MODEL_VALIDATOR_FULLNAME = 'pydantic.functional_validators.model_validator'@@ -142,8 +143,8 @@ class PydanticPlugin(Plugin): return None def get_metaclass_hook(self, fullname: str) -> Callable[[ClassDefContext], None] | None:- """Update Pydantic `ModelMetaclass` definition."""- if fullname == MODEL_METACLASS_FULLNAME:+ """Update Pydantic `ModelMetaclass` and `_RootModelMetaclass` definitions."""+ if fullname in (MODEL_METACLASS_FULLNAME, ROOT_MODEL_METACLASS_FULLNAME): return self._pydantic_model_metaclass_marker_callback return None @@ -171,7 +172,7 @@ class PydanticPlugin(Plugin): transformer.transform() def _pydantic_model_metaclass_marker_callback(self, ctx: ClassDefContext) -> None:- """Reset dataclass_transform_spec attribute of ModelMetaclass.+ """Reset dataclass_transform_spec attribute of ModelMetaclass (and _RootModelMetaclass). Let the plugin handle it. This behavior can be disabled if 'debug_dataclass_transform' is set to True', for testing purposes.@@ -916,8 +917,14 @@ class PydanticModelTransformer: The added `__init__` will be annotated with types vs. all `Any` depending on the plugin settings. """- if '__init__' in self._cls.info.names and not self._cls.info.names['__init__'].plugin_generated:- return # Don't generate an __init__ if one already exists+ if (+ '__init__' in self._cls.info.names+ and not self._cls.info.names['__init__'].plugin_generated+ and self._cls.info.fullname != ROOT_MODEL_FULLNAME+ ):+ # Don't generate an __init__ if one already exists (except for `RootModel` itself,+ # so that the `root` argument follows the plugin settings, as with other model fields)+ return typed = self.plugin_config.init_typed model_strict = bool(config.strict)tests/mypy/modules/root_models.pyindex 804d1d43..b6c5b532 100644@@ -46,3 +46,19 @@ m = Model[str](m1=Maybe(None), m2=Maybe('dog'), m3=Maybe([])) Model(m1=None, m2={}, m3=[]) assert_type(m.m1, Maybe[int])+++class FloatModel(BaseModel):+ f: float+++FloatRoot = RootModel[float]+++class SubFloat(RootModel[float]):+ pass+++FloatModel(f='1.0')+FloatRoot('1.0')+SubFloat('1.0')tests/mypy/outputs/mypy-default_ini/root_models.pyindex 01cf5bbb..cc6e3c6d 100644@@ -54,3 +54,22 @@ Model(m1=None, m2={}, m3=[]) # MYPY: error: Argument "m3" to "Model" has incompatible type "list[Never]"; expected "Maybe[Any]" [arg-type] assert_type(m.m1, Maybe[int])+++class FloatModel(BaseModel):+ f: float+++FloatRoot = RootModel[float]+++class SubFloat(RootModel[float]):+ pass+++FloatModel(f='1.0')+# MYPY: error: Argument "f" to "FloatModel" has incompatible type "str"; expected "float" [arg-type]+FloatRoot('1.0')+# MYPY: error: Argument 1 to "RootModel" has incompatible type "str"; expected "float" [arg-type]+SubFloat('1.0')+# MYPY: error: Argument 1 to "SubFloat" has incompatible type "str"; expected "float" [arg-type]tests/mypy/outputs/mypy-plugin_ini/root_models.pyindex 5bb30a04..6835c712 100644@@ -49,3 +49,19 @@ m = Model[str](m1=Maybe(None), m2=Maybe('dog'), m3=Maybe([])) Model(m1=None, m2={}, m3=[]) assert_type(m.m1, Maybe[int])+++class FloatModel(BaseModel):+ f: float+++FloatRoot = RootModel[float]+++class SubFloat(RootModel[float]):+ pass+++FloatModel(f='1.0')+FloatRoot('1.0')+SubFloat('1.0')tests/mypy/outputs/pyproject-default_toml/root_models.pyindex 01cf5bbb..cc6e3c6d 100644@@ -54,3 +54,22 @@ Model(m1=None, m2={}, m3=[]) # MYPY: error: Argument "m3" to "Model" has incompatible type "list[Never]"; expected "Maybe[Any]" [arg-type] assert_type(m.m1, Maybe[int])+++class FloatModel(BaseModel):+ f: float+++FloatRoot = RootModel[float]+++class SubFloat(RootModel[float]):+ pass+++FloatModel(f='1.0')+# MYPY: error: Argument "f" to "FloatModel" has incompatible type "str"; expected "float" [arg-type]+FloatRoot('1.0')+# MYPY: error: Argument 1 to "RootModel" has incompatible type "str"; expected "float" [arg-type]+SubFloat('1.0')+# MYPY: error: Argument 1 to "SubFloat" has incompatible type "str"; expected "float" [arg-type]
pydantic/mypy.pyindex cd28e7004..40aee0627 100644@@ -88,6 +88,7 @@ CREATE_MODEL_FULLNAME = 'pydantic.main.create_model' BASESETTINGS_FULLNAME = 'pydantic_settings.main.BaseSettings' ROOT_MODEL_FULLNAME = 'pydantic.root_model.RootModel' MODEL_METACLASS_FULLNAME = 'pydantic._internal._model_construction.ModelMetaclass'+ROOT_MODEL_METACLASS_FULLNAME = 'pydantic.root_model._RootModelMetaclass' FIELD_FULLNAME = 'pydantic.fields.Field' DATACLASS_FULLNAME = 'pydantic.dataclasses.dataclass' MODEL_VALIDATOR_FULLNAME = 'pydantic.functional_validators.model_validator'@@ -143,7 +144,7 @@ class PydanticPlugin(Plugin): def get_metaclass_hook(self, fullname: str) -> Callable[[ClassDefContext], None] | None: """Update Pydantic `ModelMetaclass` definition."""- if fullname == MODEL_METACLASS_FULLNAME:+ if fullname in (MODEL_METACLASS_FULLNAME, ROOT_MODEL_METACLASS_FULLNAME): return self._pydantic_model_metaclass_marker_callback return None @@ -916,7 +917,13 @@ class PydanticModelTransformer: The added `__init__` will be annotated with types vs. all `Any` depending on the plugin settings. """- if '__init__' in self._cls.info.names and not self._cls.info.names['__init__'].plugin_generated:+ if (+ '__init__' in self._cls.info.names+ and not self._cls.info.names['__init__'].plugin_generated+ # `RootModel` declares a typed `__init__` for the benefit of type checkers when the plugin+ # isn't used, but the plugin-generated one should take priority over it:+ and self._cls.fullname != ROOT_MODEL_FULLNAME+ ): return # Don't generate an __init__ if one already exists typed = self.plugin_config.init_typedtests/mypy/modules/root_models.pyindex 804d1d43e..f95bd7fe1 100644@@ -22,6 +22,18 @@ pets1 = Pets1(['dog', 'cat']) pets2 = Pets2(['dog', 'cat']) pets3 = Pets3(['dog', 'cat']) +RFloat = RootModel[float]+++class RFloatSub(RootModel[float]):+ pass+++# When the plugin is used with `init_typed` unset, arbitrary input+# should be accepted as it may be coerced to the `root` type:+rfloat = RFloat('1.0')+rfloat_sub = RFloatSub('1.0')+ class Pets4(RootModel[list[str]]): pets: list[str]tests/mypy/outputs/mypy-default_ini/root_models.pyindex 01cf5bbb3..df3354187 100644@@ -23,6 +23,20 @@ pets1 = Pets1(['dog', 'cat']) pets2 = Pets2(['dog', 'cat']) pets3 = Pets3(['dog', 'cat']) +RFloat = RootModel[float]+++class RFloatSub(RootModel[float]):+ pass+++# When the plugin is used with `init_typed` unset, arbitrary input+# should be accepted as it may be coerced to the `root` type:+rfloat = RFloat('1.0')+# MYPY: error: Argument 1 to "RootModel" has incompatible type "str"; expected "float" [arg-type]+rfloat_sub = RFloatSub('1.0')+# MYPY: error: Argument 1 to "RFloatSub" has incompatible type "str"; expected "float" [arg-type]+ class Pets4(RootModel[list[str]]): pets: list[str]tests/mypy/outputs/mypy-plugin_ini/root_models.pyindex 5bb30a044..12b4c84bd 100644@@ -23,6 +23,18 @@ pets1 = Pets1(['dog', 'cat']) pets2 = Pets2(['dog', 'cat']) pets3 = Pets3(['dog', 'cat']) +RFloat = RootModel[float]+++class RFloatSub(RootModel[float]):+ pass+++# When the plugin is used with `init_typed` unset, arbitrary input+# should be accepted as it may be coerced to the `root` type:+rfloat = RFloat('1.0')+rfloat_sub = RFloatSub('1.0')+ class Pets4(RootModel[list[str]]): pets: list[str]tests/mypy/outputs/pyproject-default_toml/root_models.pyindex 01cf5bbb3..df3354187 100644@@ -23,6 +23,20 @@ pets1 = Pets1(['dog', 'cat']) pets2 = Pets2(['dog', 'cat']) pets3 = Pets3(['dog', 'cat']) +RFloat = RootModel[float]+++class RFloatSub(RootModel[float]):+ pass+++# When the plugin is used with `init_typed` unset, arbitrary input+# should be accepted as it may be coerced to the `root` type:+rfloat = RFloat('1.0')+# MYPY: error: Argument 1 to "RootModel" has incompatible type "str"; expected "float" [arg-type]+rfloat_sub = RFloatSub('1.0')+# MYPY: error: Argument 1 to "RFloatSub" has incompatible type "str"; expected "float" [arg-type]+ class Pets4(RootModel[list[str]]): pets: list[str]
What the agent reported
I fixed this in the mypy plugin (pydantic/mypy.py). With the plugin on, RootModel[float]("1.0") now type-checks the same way Foo(f="1.0") does. A new regression test fails without the fix and passes with it. None of this is committed.
What was wrong. With the plugin's default settings, it gives model constructors Any-typed arguments, which is why Foo(f="1.0") is accepted. Two things kept RootModel from getting the same treatment:
- Direct use,
RootModel[float](...):RootModelhas its own__init__(self, /, root: RootModelRootType = ...). The plugin never replaces an existing__init__, so mypy always checkedrootagainstfloat. - Subclasses,
class Sub(RootModel[float]): these hit the same error.RootModeluses a special type-checking-only metaclass that carries its own@dataclass_transform. The plugin switches that off on the normal model metaclass but not on this one. So mypy's own dataclass handling built a strictly typed__init__(self, root: float)for subclasses.
Without the plugin, mypy rejects all three calls, including Foo(f="1.0"). That is already consistent, so I left that mode alone.
What I changed.
- Added
ROOT_MODEL_METACLASS_FULLNAME = 'pydantic.root_model._RootModelMetaclass'.get_metaclass_hooknow switches offdataclass_transformfor it too, just as it already did for the normal metaclass. add_initializernow also generates__init__forRootModelitself. SorootisAnyby default, and typed wheninit_typed = Trueor the model is strict, like any other field.- Tests: I added the report's scenario (a
BaseModelwith a float field,RootModel[float], and aRootModel[float]subclass, each called with'1.0') totests/mypy/modules/root_models.py. I updated the expected outputs: mypy-plugin_ini: no errors on those lines.mypy-default_iniandpyproject-default_toml(no plugin): all three lines error, which is what happens today.
Test evidence.
- mypy isn't installed in the project environment and installing packages was blocked, so
./bt-testskips all 53 mypy tests. I ran them with the locked version (mypy 1.18.2) taken from the local uv cache throughPYTHONPATH, usingpytest tests/mypy --test-mypy. - Without the fix: the new
mypy-plugin.ini/root_models.pytest fails, and the only difference is the twoRootModel/ subclass arg-type errors. The two no-plugin configurations pass. - With the fix: 48 passed, 5 failed. Those same 5 also fail on the untouched checkout:
pydantic_settings.pyunder four configurations andstrict_equality.py. - By hand, I confirmed that
init_typed = Trueandstrict=Truestill report the error, and that results without the plugin are unchanged. ./bt-test tests/test_root_model.py tests/mypy: 78 passed, 53 skipped.
Side effects. Only mypy's view of RootModel changes; runtime and pyright are unaffected. One edge case gets stricter: with init_forbid_extra = True, calling RootModel[dict[str, int]](a=1) directly is now rejected, even though it works at runtime. Subclasses already behaved this way. No test covers that case, or the init_typed = True case; I checked the latter by hand but didn't add a test for it.
Requirements:
- Passing a convertible value to
RootModel[float]should not raise[arg-type]under mypy: done when the plugin is on. Covered by the regression test above. BaseModelfields and theRootModelroot should behave the same: done for plugin mode, and it now includesRootModelsubclasses, which had the same bug. Without the plugin they already matched, since both error; the test outputs for that mode show this.- The same should hold with or without
--strict: this flag didn't affect the result either way. With the plugin, theAnytyping followsinit_typedand each model's ownstrictsetting, the same as forBaseModelfields.