marimo-pair

AI pair programming in a running marimo notebook works by programmatically executing and mutating cells through an HTTP kernel, allowing the assistant to safely read, edit, and run code so reactive graph changes appear live in your browser.

To create and mutate notebook cells programmatically without breaking the reactive graph, you must follow guard rails for marimo's code API, using async context and ctx.* for mutations to avoid silent UI inconsistencies and clobbered edits.

You do not need marimo installed locally to execute code through an HTTP kernel, because the workflow uses provided Bash scripts to discover active marimo sessions and execute code deterministically without requiring a local installation.

The best way to debug and refactor Python notebooks with reactive programming is to let an AI work directly inside the running kernel, targeting exploration and refactoring workflows while ensuring reactive graph changes reflect live in the browser.

Editing marimo cells programmatically causes silent UI inconsistencies when you bypass the async context and code_mode guard rails, which are required to ensure kernel-safe cell mutations and prevent clobbered edits in the reactive graph.