I'm wondering if it's better to add user_agent agent directly to storage::RequestOptions instead of to CommonOptions. WDYT?

You are probably right. Putting it in CommonOptions now forces us to think about how a per-client setting works.

When I was thinking about the feature request I was thinking about how in (say) secret manager I can say:

let client = SecretManagerService::builder().build().await?;
let get = client.get_secret().with_user_agent("foo/1.0").send().await?;

And the goal was to do the analogous thing for Storage:

let client = Storage::builder().build().await?;
let reader = client.read_object("projects/_/buckets/my-bucket", "my-object")
    .with_user_agent("foo/1.0") // ERROR, LET'S FIX THIS
    .send().await?;

This PR gives applications a different feature, they can do this now:

let client = Storage::builder()
    .set_user_agent("foo/1.0") // NEW, COOL, BUT NOT WHAT ASKED.
    .build().await?;
let reader = client.read_object("projects/_/buckets/my-bucket", "my-object")
    .with_user_agent("bar/1.0") // STILL AN ERROR...
    .send().await?;

To be clear: we will want the change in this PR too, someday, but it is not (I believe) what #3615 is asking.

In passing: the other changes in #3615 (host header, using reqwest::Client, library instrumentation for observability) are already done. You can mark that issue as fixed once the support for setting the user header is done.

Thanks. There are a couple of extra changes that look out of place. I trust you to revert them and then merge the code.