Roadmap

What is done, and what comes next, in order. Each step lands as its own pull request with a GitHub issue, and records what it learned in the development log.

Done

• Core crates. Protocol, kernel host, nbformat, executor, contents, API, and server, with the ganymede CLI and the embeddable ganymede-engine.
• Parity with Jupyter. An end-to-end suite runs cdp-kernel and ipykernel through Ganymede and compares the results with jupyter execute and jupyter server.
• Existing frontends. The VS Code Jupyter extension's remote-server flow passes. JupyterLab's own bundle is served at /lab with its /lab/api/* endpoints and runs notebooks in a browser, with cookie login as jupyter_server does it.
• Portability. The engine builds for iOS, Android, and wasm, and runs a notebook inside wasm under WASI.
• Kernel conformance. cdp-kernel (Node 24) and ipykernel run the same scenarios in CI: launch, execute, interrupt, restart, comms with binary buffers, stdin, aborted queues, history, display updates, the runtime exiting mid-cell, and parallel starts, with kernelspec metadata and kernel_info_reply extras passed through untouched.
• Kernel lifetime. Kernels survive websocket drops, and a client that reconnects with the same session_id gets the output it missed, as jupyter_server buffers it. Idle culling is configurable. A kernel that dies is reported dead to its clients and removed; an attached one is watched by heartbeat. A start or restart whose kernel dies or doesn't answer is relaunched on fresh ports, up to three tries. Shutdown sends shutdown_request, then SIGTERM, then SIGKILL, and kernels die with the host.
• JupyterLab without Python. ganymede lab install fetches JupyterLab's prebuilt bundle from PyPI, ganymede lab extension install the prebuilt extensions (the widget manager and anywidget by default), and ganymede lab serves and opens it.
• Releases and install. A vX.Y.Z tag builds and publishes binaries for Linux, macOS, and Windows (the Windows build is not tested beyond compiling). install.sh installs a release and cdp-kernel. CI's workstation job runs the documented install on a machine with no Python and drives JupyterLab with Playwright.
• Notebook trust. Notebooks are signed on save and marked trusted on read as jupyter_server does, with Jupyter's key and signature database, and ganymede trust signs one by hand.

Next

1. Document model. The file is the source of truth and is reloaded when it changes on disk (iCloud). A yrs CRDT layer is optional and later.
2. A first UI of our own, outside the core: a Swift shell with a web view for iPadOS and macOS, running a TypeScript notebook app, with the Rust core embedded in the same process. The TypeScript app is ganymede-ui, a prototype that runs in a browser and installs as a PWA.
3. CurveZMQ for the ZeroMQ transport. ipykernel 7 already advertises it.

Open

• Restarting a kernel automatically when it crashes, as jupyter_server's restarter does. Today a dead kernel is reported and removed.
• Conformance gaps: debugger (debug_request) messages, malformed stdin frames, and kernel death and parallel starts driven through serve.
• A Linux build for musl (Alpine).
• A runtime seam for the browser (timers and spawn on wasm-bindgen-futures) and a websocket or postMessage transport, so the engine runs in a page and not only under WASI.
• Windows interrupts, which ipykernel does through a named event rather than a signal.

Not planned

JupyterLab server extensions and jupyter_server extensions (prebuilt JupyterLab frontend extensions do work), nbconvert export, and the JupyterHub single-user contract. See Replacing Python for why.