Add depenecies from within a running notebook

[ianhi]

Problem

When exploring a new project I don't always know what tools I will need installed from the start. With conda environments as I've learned more about what I will need I would install from within the notebook using !pip install .... which is very convenient because you do not need to go through the loop of closing the notebook, installing, then re-opening.

I'd like to be able to achieve the same with juv

SOlution?

Where I do !juv add <something> and then can seamlessly import something

Potential Issue

Running this command will modify the notebook, which then creates a conflict between the in memory notebook and the on disk version and eventually prompts a dialog asking if you want to revert or overwrite, neither of which is an ideal option. (It would be very cool if that jupyterlab dialog also had a merge option)

[manzt]

I totally understand the use case and have actually explored a runtime "client" before (as a special cell magic):

• Add experimental cellmagic #17

The main challenge was handling writes to the notebook file from the CLI at the same time as the front end (as you mentioned), which led to tricky synchronization issues I wasn’t comfortable shipping. I'm open to revisiting it, but I suspect a front-end-specific extension (e.g., for JupyterLab) might be a more robust solution for writing the metadata. What we really want is some way to install the dependency on the kernel side, but write the metadata in the front-end UI for the user to save so it's a part of the same document.

What we want to avoid is a state where the metadata and the running environment differ. FWIW, our VS Code extension supports adding dependencies while editing because it manages the virtual environment alongside the notebook and is aware of the front end (VS Code). This API was possible because the VS Code extension API offers mechanism to handle the synchronization issues mentioned.

Aside: One can always !uv pip install within the running notebook, but sadly the PEP 723 metadata sadly won't be written.