zx.parseFormSafe type inference does not handle ZodEffects

[mdoury]

I noticed the zx.parseFormSafe type inference does not handle well ZodEffects schemas resulting from the use of refine, superRefine or transform. I will create a PR with a failing test and go from there. Do you have any idea how to fix this?

[mdoury]

Just for context, I'm trying to parse a ZodEffects schema using zx.parseFormSafe with a FormData object because the form contains an image.

Here is a simplified reproduction of the error:

const schema = z
  .object({
    password: z.string().min(8),
    confirmPassword: z.string().min(8),
  })
  .refine(({ password, confirmPassword }) => password === confirmPassword);
const formData = new FormData();
formData.set('password', 'foo');
formData.set('confirmPassword', 'bar');
const result = zx.parseFormSafe(formData, schema);
if (result.success) {
  return json(result);
} 
const { formErrors, fieldErrors } = result.error.formErrors;
return json({ success: false, formErrors, fieldErrors });

The type of result.data seems correctly inferred, however, the type of result.error is not.
As a consequence, when I try to access fieldErrors.password or fieldErrors.confirmPassword in my template TypeScript is angry:

Property 'password' does not exist on type 'SerializeObject<UndefinedToOptional<{ _output?: string[] | undefined; _input?: string[] | undefined; refinement?: string[] | undefined; innerType?: string[] | undefined; _parse?: string[] | undefined; _type?: string[] | undefined; ... 27 more ...; isNullable?: string[] | undefined; }>>'. ts(2339)

I tried parsing directly with safeParse and safeParseAsync from zod and it gets it right.

It seems to be coming from the definition of SafeParsedData:

type SafeParsedData<T extends ZodRawShape | ZodTypeAny> = T extends ZodTypeAny
  ? SafeParseReturnType<T, ParsedData<T>>
  : T extends ZodRawShape
  ? SafeParseReturnType<ZodObject<T>, ParsedData<T>>
  : never;

The ZodEffects case is not handled, and SafeParseReturnType<Input, Output> gets the right Output which explain the correct inference of result.data, but the Input is visibly not appropriate.