[Feature]: Run Python notebook cells with a shared user/agent kernel

[themeaningW]

Problem

When working with an agent on a Python notebook, I want to run a cell, inspect its output or plot, and ask the agent to continue using the same variables without switching to a separate Jupyter application or starting another kernel.

This is a proposal for Python notebook execution in T3, not embedding JupyterLab or its extension system.

Related work

• [Feature]: Read-only `.ipynb` notebook preview #11033 proposes read-only notebook previews.
• [Feature]: Edit Jupyter notebooks (.ipynb) cell by cell #17622 proposes cell-level source editing and explicitly leaves execution and kernel management out of scope.
• feat(web): preview opendocument, epub, rich text, and notebooks #21469 currently adds notebook previews, including saved outputs.

This proposal is a separate execution follow-up to that work. It would build on the agreed notebook viewer rather than introduce a competing preview or expand the source-editing proposal.

Proposed initial scope

• Web and desktop execution of Python code cells in writable workspace .ipynb files.
• One persistent Jupyter kernel per notebook, with variables shared between cells and between user-triggered and agent-triggered execution.
• Run one cell or all code cells in order; stop Run all at the first error.
• Display stream output, errors, text results, and static PNG/JPEG plots alongside the cells.
• Interrupt and restart controls. If interruption fails and the kernel must be shut down, explicitly report that its variables were cleared.
• Save execution counts and outputs while preserving notebook metadata and unknown fields. If the notebook changes during execution, preserve those edits and report that the new results were not saved.
• Agent tools use the same execution service and kernel as the UI. Code runs on the environment's host with permissions equivalent to terminal execution; agent execution requires full access.
• Use an installed Python/Jupyter environment and retain bounded file/output handling.

Closing the preview would preserve kernel state; restarting the host server would clear it.

Out of scope

JupyterLab's UI or plugins, interactive widgets, JavaScript kernels, live synchronization with an external JupyterLab session, cell-aware diffs, and a native mobile execution UI. Cell-level source editing remains the scope of #17622.

Local prototype

I have a personal Windows prototype based on v0.1.1-nightly.20261016.3203. It is an exploration of this workflow, not an approved implementation or a branch validated against current main.

Focused checks passed:

• Three real-kernel integration tests covering shared variables, matplotlib PNG output and saved metadata, preservation of concurrent file edits, and interruption of a native sleep followed by another execution.
• One HTTP MCP tool-registration test and 25 file-preview regression tests.
• Web and server typechecks.
• A browser pass: assign answer = 40 in one cell, obtain 42 in another, render a matplotlib plot, run all cells, and verify that restarting clears variables.

The prototype still edits source through the raw JSON view. Packaged Electron startup, macOS/Linux, relay/tunnel access, and native mobile behavior have not been verified. Host Python setup and kernel selection need an upstream design decision.

Questions for maintainers

1. Is shared Python-kernel execution a direction you would consider for T3?
2. Should it wait for feat(web): preview opendocument, epub, rich text, and notebooks #21469 to land, or use another integration point?
3. What host Python setup and kernel-selection workflow would fit T3 without bundling Python?
4. Is web/desktop first, with agent access to the same kernel, an acceptable initial scope?

I am asking for direction and scope approval under CONTRIBUTING.md before submitting an upstream PR. If approved, I would port the agreed scope onto the agreed base, rerun focused checks, and provide before/after UI evidence and an execution recording.

Prepared with gpt-6.1-sol through Codex in T3 Code.