Prior art

What to read before working on a part of Ganymede.

Reading list

• jupyter_client messaging spec (v5.5) and JEP 91 (subshells), JEP 92 (optional features / supported_features).
• jupyter_server websocket protocol doc and services/kernels/connection/base.py for the v1 binary layout.
• runtimed workspace: jupyter-protocol, runtimelib, nbformat, jupyter-websocket-client. Zed's repl crate is a consumer.
• Deno's built-in kernel (deno jupyter): a kernel in Rust on the zeromq crate, which is the other side of the same wire.
• evcxr_jupyter: Rust kernel, useful as a test target.
• nbclient for execution semantics, jupyter_ydoc for the CRDT shape.
• xeus (C++) for a non-Python view of the kernel protocol.
• cdp-kernel docs: architecture, runtimes, known gaps, and the widgets and comms guide. The generated kernel.json is the first thing to read.
• Jupyverse (David Brochart): a from-scratch jupyter_server replacement in Python. Useful for the exact endpoint list JupyterLab's frontend needs, including /lab/api/*.
• VS Code Jupyter extension: the remote-server client, for what a non-JupyterLab frontend actually calls.
• anywidget: the widget protocol subset to support instead of ipywidgets.
• Datalayer's jupyter-react: a React frontend that reuses @jupyterlab/rendermime and @jupyterlab/outputarea without the rest of JupyterLab. Relevant to a TypeScript web UI.
• JupyterLite: JupyterLab in the browser with a TypeScript reimplementation of the REST API in a service worker and in-browser kernels. The proof that the frontend can be pointed at a non-HTTP backend through ServerConnection, and the closest precedent for the iPad profile.
• jupyter_server's gateway client and Jupyter Kernel Gateway: the remote provisioner pattern.

The reference stack by size

Jupyter's components, with rough Python line counts, and the Rust equivalents that existed when Ganymede started:

Layer	Reference impl	Approx. size	Rust equivalent
Wire protocol (ZMQ, HMAC, message schema)	jupyter_client messaging spec v5.5	spec only	jupyter-protocol crate exists
Kernel lifecycle (spec discovery, launch, ports, restart, interrupt, shutdown)	jupyter_client KernelManager/Provisioner	~10k	runtimelib covers part
Notebook file format	nbformat v4.5 + JSON schema	~3k	nbformat crate exists
Server (REST: kernels, sessions, contents, kernelspecs; websocket bridge; auth)	jupyter_server	~15k	nothing complete
Headless execution	nbclient	~2k	nothing
Realtime collaboration	jupyter_collaboration (Yjs / pycrdt)	~5k + JS	yrs crate is the Rust Yjs
Frontend	JupyterLab (@jupyterlab/*, Lumino)	~300k TS	nothing; Zed REPL is a partial native UI
Widgets	ipywidgets + @jupyter-widgets/*	~50k TS	nothing

The last two rows are the bulk of the work and the least Rust-shaped.