I don't think this is a good idea; what is really needed is some sort of compound mappable artist. It doesn't have to be a new kind of collection, but it needs to be a single ScalarMappable so that when sci() is used and the colormap is changed, the lines and arrows change together.
Also, when a cmap kwarg is allowed, a norm kwarg should also be allowed.

Hmm,... I was hoping it wouldn't come to that :)

I guess I'll take a look at the ContourSet object (as you suggested in the original streamplot PR) and emulate it's behavior.

BTW, after looking back at the original PR, I realized never flagged the streamplot function as experimental, as you suggested. I'm not really sure where that should be marked. Maybe in the return value itself (assuming that's the only major part that's likely change)?

The ContourSet is a model in the sense that it produces and contains more than a single primitive Artist, and in that it requires extensive computation to generate those Artists, but it may not be an ideal model. It evolved a long time ago. It is possible that the Container base class would provide a good starting point. This is just speculation, though; I haven't looked closely.

In the newest PR,streamplot returns aContainer object with lines and arrows. This needs careful review because I had to override the__new__ constructor which didn't play well withCollections. I don't really understand the container object (e.g. why does it subclasstuple?), so I could be breaking something with this hack.

The original purpose of the Container object was to support legend for complex plot commands. For example, errorbar command used to return a tuple of artists, thus I ended up with a simple Container class inherited from tuple. And it never meant to be a full implementation of artist.

I'm not sure what the streamplot need to return. Maybe a single object of a class inherited from a ScallarMappable (and Artist). And I think it might not be a good idea to use the Container class for this purpose.

Thanks for your comments,@leejjoon. After rereading@efiring's comments, I realize that he suggested thatContainer would be a good starting point---not necessarily the base class for a streamplot container.

In my most recent commit, I change the base class to a simple object and give it attributeslines andarrows. As mentioned in the commit message, this is just a temporary fix since this object doesn't allow the user to set the properties of the lines and arrows from the object itself. Nevertheless, it does at least give users access to the arrows, and (I think) future changes (e.g. subclassingScalarMappable andArtist) could be made backward-compatible with this simple object.

@tonysu, sorry for the extra work, but would you rebase this against master, please? Then I will merge it. I should have done so a long time ago. It definitely needs to be done before the RC.

@tonysyu, also, while you are at it, you might want to note, in docstrings, and/or comments, something about the limitations and likely evolution.

@efiring Rebased and added note about future changes tostreamplot return value.

Hmm, Github seems to be screwing up the order of the commits. The discussion page for this PR shows commits in order: 1, 3, 2, 4. The commit page shows them in sequential order, and the diff page looks fine so it's probably just a bug in the discussion page.

This pull requestfails (mergeda7248d9 intocadd152).