Hmm, I'm uncertain how to make this patch look nice in the github review UI. It looks okay in my local git history viewer, but afaict github really wants to display them squashed?

After sleeping on it, there are definitely a few things this patch could do better.

1. Hard coding of "DataFusion CLI {DATAFUSION_CLI_VERSION}"
2. Calling Args::parse which will exit instead of Args::try_parse or Args::try_parse_from which will return a Result.
3. Probably should just bite the bullet and make Args public, extending the existing example to use the command line arguments for configuring it's CliSessionContext, instead of calling SessionContext::new().
4. That'll require taking a CliSessionContext parameter, and probably taking ownership of PrintOptions from a CliSession, perhaps the fields of CliSession should just be pub?.

Mostly fixed in 249072e, except the first point which I'm not sure needs to be done and seems minor to me.

So the only thing I'm a bit uncertain/hesitant about (I'm really not all that familiar with clap), is whether this will work
for anyone who wants to extend the arguments using #[command(flatten)], adding their own command line options, but reuse the datafusion-cli ones. I kind of assume that use case will need additional changes with something like CliSession::from_clap_args(clap_args: entry_point::Args) -> Result<CliSession, CliError>, but uncertain and hesitate to implement it without anything trying to actually do that.

Edit: After thinking about, this seemed like I should just do this now as it seemed inevitable, and it's better if we don't have to break semver later to handle it. Fixed in e3d941a

Alright I think this has ended up in a reasonable state (modulo bike shedding), and I can't imagine anything else (crosses fingers), so I believe it should be ready for review

Alas, I think there is one more thing to do here:

I just tried to use this to set up a custom SessionStateBuilder ... with_type_planner(...) call,
and It feels like this doesn't allow you to call SessionContext::new_with_state for the context returned by
from_args. (from there you can still use the exec_from_repl whose CliContext::session_state can return a custom state. But it just feels off, because it looks like SessionContext, and CliSessionContext have two different SessionState values.

It feels to me that this calls for a builder API...

CliSession::builder().session_context(SessionContext::new_with_state(...)).args(args).build()

Then if you don't suppy the builder with one via session_context It can just build a default one.
anyhow, I'm going to try that.

Well, I pushed what I think is right, but as a noob I've struggled to come up with a good example that actually utilizes the ability to customize session_state. There are examples in datafusion-examples, but they're pretty involved. What I've struggled to find is a simple one.

So, I feel that the major thing left is to somehow make this reviewable in the github UI.
Somewhere I've read it has been suggested to split the code movement and the changes into separate PRs rather than patches to work around how the github UI displays commits as though they were squashed.

I think it is worth trying, so I'm going to close this and try separating into two PRs and clean up the history a bit.