feat: sandboxed home

[dmadisetti]

📝 Summary

Adds --sandbox to marimo edit --sandbox directory/. Leveraging pep-723 where ever it can.

closes #2598 rebases and supercedes #7640

• removes dangerous sandbox
• spawns ipc process only for marimo edit --sandbox (home mode)
• uses lazy front end

UX changes on kernel launch failure:

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Review	Updated (UTC)
marimo-docs	Ready	Preview, Comment	Jan 9, 2026 5:37pm

[akshayka]

Before releasing this feature, we should make sure the UX issues you alluded to in the PR description, in particular when the frontend handles the case in which sandbox notebooks fail to start (e.g., uv sync fails):

• the error should be communicated to the user;
• we should either fail fast (show the error and stop trying to connect to the kernel), or provide a way for the user to fix the error and try again (edit script metadata, then click a button to try creating the venv again)

If you prefer to work in multiple PRs, sandbox home could be feature flagged until the FE handles the failure case

[akshayka]

Shutting down the server on the frontend should also shutdown running notebooks, but that does not appear to be happening right now

[mscolnick]

self.session.notify( KernelStartupErrorNotification(error=error_message))

[mscolnick]

oh maybe you don't have the session? this bit feels fragile

[dmadisetti]

Good point. Changing this up, an advantage to having a session is that we could potentially stream the errors/ startup before closing. Will keep that for a future PR because this is quiet long, but will do the non-blocking kernel launch

[dmadisetti]

I did try to refactor this such that we:

• async attempted startup kernel
• started session and the web socket loop
• push notifications from kernel startup directly to the queue

This also has the advantage of being able to stream messages during startup

In theory. In practice the installation modal popped up after it was done (not sure what I was missing there), I had a couple race condition issues where kernel closed out before WS was ready.

There's also not really a clear mechanism to turn off the session from inside the kernel, and started going down the route of a process interrupt- I didn't want to rely on the FE for teardown since it was leaving orphaned sessions. Then remembered if we just fail before we set up our scaffolding, then we don't have to do all this- so I reverted and just cleaned up.

Maybe we can revisit, but it was looking to be like a fair amount of more work on an already long PR. Maybe there's a simpler way but seemed marginal for just a slight code smell

[akshayka]

Might not be related to this change, but I created a new notebook and marimo was not added to its script metadata:

# /// script 
# requires-python = ">=3.13"
# dependencies = [
#     "matplotlib==3.10.8",
# ]
# ///

Additionally, I am finding a bad interaction with format on save, see video:

ruff-install.mp4

[dmadisetti]

marimo not being added a a bug particular to this (but we should fix). Opening the motebook again and it works as expected.

For the ruff error, I'm just going to throw an error and revisit if we need to. Ruff is installed in the sandbox, but the wrong sandbox! Installation path goes through the kernel