applyPostTransforms should not be applied for all builder schemas

packages/angular_devkit/core/src/json/schema/registry.ts から始め、特に applyPostTransforms オプションを確認してから、2 番目の引数なしで validation が呼び出されている packages/angular_devkit/architect/src/architect.ts を調べます。issue で説明されている builder schema の動作を再現し、他の schema の既存の defaults を変更せずに、特定の builder で addUndefinedDefaults を無効化する方法を判断します。

🚀 Feature request

Command (mark with an x)

- [ ] new
- [ ] build
- [ ] serve
- [ ] test
- [ ] e2e
- [ ] generate
- [ ] add
- [ ] update
- [ ] lint
- [ ] xi18n
- [x] run
- [ ] config
- [ ] help
- [ ] version
- [ ] doc

Description

I am currently writing a builder and the appropriate JSON schema to validate its options.
The builder will be used in different targets, but all of these targets usually share some base options which is why I added some target reference support, similar to the browserTarget
option of the @angular-devkit/build-angular:dev-server builder.
An example config could look like this:

"target1": {
    "builder": "mybuilder:f":
    "options": {
        "a": "foo",
        "b": {
            "asd": "wer"
        }
    }
},
"target2": {
    "builder": "mybuilder:f":
    "options": {
        "targetRef": "myproject:target1", // References the options of target1 above
        "a": "bar"
    }
},

So, at the end the options of target2 should be evaluated to the following:

{
    "a": "bar",
    "b": {
        "asd": "wer"
    }
}

This should be as simple as using some Object.assign code:

const target2Options = {  "a": "bar"  };
const target1Options = {  "a": "foo", "b": { "asd": "wer" }  };
const result = Object.assign({}, target1Options, target2Options);

But the problem is, that Angular modifies the incoming options-object. It passes in an empty object for the b property of target2's options: { "a": "bar", "b": {} }. Thus, result.b will be {} instead of { "asd": "wer" }.
This is done by the Angular's postTransform addUndefinedDefaults which is always added to the CoreSchemaRegistry. In theory postTransforms can be disabled as seen here using applyPostTransforms: false. However, it is not possible to override this value when the builder's schema is compiled and the validator is used, see here (no second argument is passed to validation).

Of course, I could add some code to my builder that checks for that {} and ignores it. But actually, it should be possible to explicitly set b: {} to override the inherited value, e.g.:

"target3": {
    "builder": "mybuilder:f":
    "options": {
        "targetRef": "myproject:target1",
        "b": {}
    }
},

So I really need to distinguish between the {} set by the user and the {} set by the postTransform.

NOTE: All of this affects options that are set to type object or array (in which case [] is used as default) in the schema.

Describe the solution you'd like

It would be nice to be able to disable postTransforms for a specific builder, for example in the builder code itself.

Actually I do not know why the addUndefinedDefaults transform is used at all, but there is probably a reason for that. Is it? 😛

Describe alternatives you've considered

I cannot imagine any workarounds. Angular modifies the options passed to a builder where it should not, IMHO. In the above case, the user does not specified a b property, but Angular passes in an empty object for b. And currently there is no way to change that behavior, as far as I can see.

EDIT:
I did find a workaround, but it is an ugly hack as it relies on the internals of addUndefinedDefault and thus could easily break when Angular is updated.

Instead of this schema:

"b": {
    "type": "object",
    "additionalProperties": {
        "type": "string"
    }
}

One can use this schema:

"b": {
    "oneOf": [
        {
            "type": "null"
        },
        {
            "type": "object",
            "additionalProperties": {
                "type": "string"
            }
        }
    ]
}

This works, because Angular detects that two types are possible for b: objectand null. This causes Angular to not override the value.
Of course, null is now a valid value for b, too, but that may not be a problem in some cases.