This isn't currently working because the Node.js script runner isn't able to find pyodide where it gets installed: https://github.com/agriyakhetarpal/scikit-image/actions/runs/9471953221/job/26096398007

I think I have two potential fixes; one is to do npm install pyodide --global (this didn't work), and the other is to install to the {project} directory, i.e., the root of the repository, as a prefix because that is where I have instructed cibuildwheel to run the JavaScript file from. I'll test this in a moment. Edit: started a run here: https://github.com/agriyakhetarpal/scikit-image/actions/runs/9472581755/job/26098187649, let's see if this works. Edit 2: okay, this works, but now I can't do import skimage because it will pick up skimage from the {project} directory... argh. I'll have to think about this so that other libraries that use the Node.js runner such as scikit-learn can adopt this too.

Summoning @hoodmane – would you consider this a bug or rather an intricacy of how npm works (I am not proficient with JavaScript enough)? In any case, I would like to improve the Pyodide docs around all of this – not just for highlighting how to configure/use cibuildwheel with it, but also for out-of-tree builds in general and recommendations around it, in the coming time. I still have some other documentation work from our previous discussions in my backlog :)

Maybe add a note to emscripten to explain the difference in build methodology?

Yes, sure; where should I add this? Do you mean adding one above the workflow?

Maybe add a note to emscripten to explain the difference in build methodology?

Yes, sure; where should I add this? Do you mean adding one above the workflow?

Sure, anywhere to document that there is a discrepancy in the way we build emscripten, and why. Probably should be mentioned in both builds!

Got it! I won't say there's a difference in how we build for Emscripten. The only changes from a regular pip install build && python -m build are:

• we use pyodide build which monkeypatches pypa/build and uses a pywasmcross.py script to cross-compile from Linux to wasm32-unknown-emscripten, instead of compiling natively
• cibuildwheel uses pyodide build as well and sets up Emscripten, Node.js, etc. automatically
• the testing differs by a bit as pytest's cacheprovider plugin is somewhat broken – which I noted using a comment.

This translates to just one extra option to pass to cibuildwheel, which can only be configured using an environment variable at the moment and is done using CIBW_PLATFORM: pyodide.

Or, were you referring to how this job and the emscripten.yaml workflow differ from each other?

Ah, OK, I just wasn't sure of the reasoning of why there was any difference at all, and wanted to make sure others who look into this in the future can follow. However you choose to document that is fine.

Okay, I added a note about this in 49ac338. The git blame should also point to this PR so developers can examine why the difference exists. Thanks for the review!

I addressed the merge conflicts here. Thanks!

I notice that zizmor did not warn me about missing SHAs for the cibuildwheel action, as it's not running with the --pedantic argument. I assume it was removed at some later point, but I've still added the hash.

I'm having a bit of trouble running the tests on my fork to see if this is working: https://github.com/agriyakhetarpal/scikit-image/actions/runs/14802982173

I'll debug this, and we should be good to proceed once we have a passing wheel build.

I created a copy of this branch, the job is passing: https://github.com/agriyakhetarpal/scikit-image/actions/runs/14803174454/job/41566395477 :)

Thanks, @agriyakhetarpal!

We're having a chat with the Anaconda team soon and will mention the CORS header issue so that, hopefully, we can start using the dev wheels from the nightly repo directly.

Thanks for this, @stefanv! Actually, the CORS headers issue from the Anaconda.org side has been resolved earlier this year (see pyodide/pyodide#4898 (comment)), so these wheels are now usable in the interactive docs for the "dev" versions. Here's an example from scikit-learn: scikit-learn/scikit-learn#31085

@matthewfeickert ☝️