Thanks for the PR!

Checking pflag is not enough IMHO. The same issue will cause Set and SetDefault not working. For that reason I think we will have to revert that PR and implement key delimiter instead.

I'm not seeing the same behaviour for Set and SetDefault because they create an interface for "foo" which is then found by Get or IsSet. The problem before was that while these were returning true, the pflags, and as I just suspected env, were returning nil for "foo" instead, hence the new modified version of IsSet.

I can add the tests for Set and SetDefault to make sure?

Indeed, you are right. :)

Here is a breaking test though:

func TestUnmarshalOnlyPFlagSet2(t *testing.T) {
	flags := pflag.NewFlagSet("test", pflag.ContinueOnError)
	flags.String("foo.bar", "cobra_flag", "")

	v := New()
	assert.NoError(t, v.BindPFlags(flags))

	v.SetConfigType("yaml")
	v.ReadConfig(bytes.NewBuffer([]byte(`foo.bar: file_content`)))

	config := &struct {
		Foo struct {
			Bar string
		}
	}{}

	assert.NoError(t, v.Unmarshal(config))
	assert.Equal(t, "cobra_flag", config.Foo.Bar)
	t.Log(v.AllSettings())
}

Flags and env vars generally override config entries. This is a tough scenario, because keys are not identical in this case.

Yes, I'd noticed that. This change is at least consistent with pre #673 behviour but I guess if we also want a solution to the config override then reverting #673 and implementing a key delimter setter is the only viable way forward.

I'm afraid that any breaking change will cause trouble since this library is widely used. I hope we can clear things up a bit in a v2, but for now, I think reverting and allowing to set the key delimiter is a better option.