extension/packages/jupyter

Hanzo AI for JupyterLab

@hanzo/jupyter — an in-notebook AI assistant for JupyterLab 4. AI where data scientists live.

• Assistant side panel — a model picker, a prompt, and streaming output. It can attach the active cell, the selected cells, or the whole notebook as context.
• Cell actions (command palette, cell context menu, and cell toolbar):
  • Explain this cell — plain-language walkthrough of a code cell.
  • Fix / debug — reads the cell and its error traceback, inserts a corrected cell below.
  • Add docstring / comments — documents the code in place.
  • Optimize — vectorizes / cleans up without changing behavior.
  • Generate a cell from a prompt — describe what you want, get a new cell.
• Reads cell outputs (especially tracebacks) for the fix action, and inserts AI-generated code as a new cell below the active one.

Model calls route through the Hanzo AI gateway (api.hanzo.ai, /v1/...) via the published headless client @hanzo/ai — no bespoke HTTP. Your Hanzo hk- key is set in JupyterLab settings, never in source.

Architecture

The JupyterLab widget layer is kept thin; the model-facing logic is pure and unit-tested.

File	Responsibility
src/index.ts	The JupyterLab plugin — the imperative shell. Registers the panel, the five commands, the cell toolbar button + context menu, tracks the active notebook, and adapts live cells into the pure data shapes.
src/panel.tsx	The side-panel ReactWidget (model picker + prompt + streaming output over @hanzo/ai).
src/notebook.ts	Pure. Cell / notebook → context text, error-output & traceback extraction, ANSI stripping, truncation, fence stripping.
src/hanzo.ts	Thin over @hanzo/ai + the pure action catalog: system frame, per-action prompt builders, request shaping, and the two async model edges (runAction, streamChat).
src/config.ts	Settings resolution + command/plugin ids.
schema/plugin.json	The settings-registry schema (API key, default model, gateway base URL).

Requirements

• JupyterLab >= 4.0
• Node >= 18 and jlpm (the JupyterLab-pinned Yarn, shipped with JupyterLab)
• A Hanzo hk- API key (from https://hanzo.ai) for authenticated models; the gateway also serves public/limited models anonymously.

Build

# from the repo root, install workspace deps
pnpm install

cd packages/jupyter

# compile TypeScript → lib/ and build the federated labextension bundle
jlpm build          # dev build
jlpm build:prod     # minified production build

jlpm build runs build:lib (tsc) then build:labextension (jupyter labextension build .), emitting the federated bundle under hanzo_jupyter/labextension.

Install (development)

# from packages/jupyter — installs the Python package and links the source
# extension so a rebuild is picked up on JupyterLab reload
pip install -e .
jupyter labextension develop . --overwrite
jlpm build

jupyter labextension list   # should show @hanzo/jupyter … enabled OK
jupyter lab

For a normal install from a built wheel:

pip install .
jupyter lab

Settings — your Hanzo key

Open Settings → Settings Editor → Hanzo AI (or edit Advanced Settings JSON):

{
  "apiKey": "hk-…",
  "model": "zen5",
  "baseUrl": "https://api.hanzo.ai"
}

• apiKey — your Hanzo hk- key. Sent as a Bearer token to api.hanzo.ai. Leave blank to use the gateway's public models. Stored by the JupyterLab settings registry, never committed to source.
• model — the default model id for cell actions and the panel. Any id from /v1/models (the panel's picker lists them live). Defaults to a Zen model.
• baseUrl — override only for a self-hosted gateway.

Usage

• Panel — click the Hanzo tab in the right sidebar. Pick a model and a context mode, type a prompt, Send (or ⌘/Ctrl+Enter). Output streams in.
• Cell actions — right-click a code cell, use the cell toolbar, or the command palette ("Hanzo AI"). Fix/document/optimize/generate insert a new cell below; explain inserts a markdown note.

Test / typecheck

jlpm test        # vitest — pure logic (notebook.ts, hanzo.ts, config.ts)
jlpm typecheck   # tsc --noEmit

The JupyterLab widget APIs are integration-tested by running JupyterLab; the unit tests cover the pure core: context assembly & truncation, traceback extraction, the action-catalog prompt builders, and @hanzo/ai request shaping (with a mocked client).

Publish

• npm (the source package): npm publish (publishConfig.access is public).

• PyPI (the prebuilt extension end users pip install):

  jlpm build:prod
  pip install build twine
  python -m build              # sdist + wheel with the bundled labextension
  twine upload dist/*
  

License

MIT