The interactive window component enables extensions to offer REPL like experience to its users. VS Code provides the user interface and extensions provide the execution environment, code completions, execution results rendering and so on.
The interactive window consists of notebook editor at the top and regular monaco editor at the bottom of the viewport. Extensions can extend the interactive window by leveraging the notebook editor API and text editor/document APIs:
- Extensions register notebook controllers for the notebook document in the intearctive window through
vscode.notebooks.createNotebookController. The notebook document has a special notebook view typeinteractive, which is contributed by the core instead of extensions. The registered notebook controller is responsible for execution. - Extensions register auto complete provider for the bottom text editor through
vscode.languages.registerCompletionItemProvider. The resource scheme for the text editor isinteractive-inputand the language used in the editor is determined by the notebook controller contributed by extensions.
Users can type in code in the text editor and after users pressing Shift+Enter, we will insert a new code cell into the notebook document with the content from the text editor. Then we will request execution for the newly inserted cell. The notebook controller will handle the execution just like it;s in a normal notebook editor.
Registering a new editor type in the workbench consists of two steps
- Register an editor input factory which is responsible for resolving resources with given
globpatterns. Here we register anInteractiveWindowInputfor all resources withvscode-interactive-inputscheme:vscode-interactive-input:/**. - Register an editor pane factory for the given editor input type. Here we register
InteractiveEditorfor our own editor inputInteractiveWindowInput.
The workbench editor service is not aware of how models are resolved in EditorInput, neither how EditorPanes are rendered. It only cares about the common states and events on EditorInput or EditorPane, i.e., display name, capabilities (editable), content change, dirty state change. It's EditorInput/EditorPane's responsibility to provide the right info and updates to the editor service. One major difference between Interactive Editor and other editor panes is Interactive Window is never dirty so users never see a dot on the editor title bar.
The Interactive.open command will manually create an EditorInput specific for the Interactive Window and resolving that Input will go through the following workflow:
The NotebookEditorModelResolverService will create a NotebookFileWorkingCopyModelFactory and use that to create a WorkingCopyManager which is then used to create a SimpleNotebookEditorModel.
When the SimpleNotebookEditorModel is requested to load, it will ask the WorkingCopyManager to create a new StoredWorkingCopy which reads content from a resource URI with the fileService. That content is passed to the the ModelFactory which retreives a NotebookSerializer from the notebookService and constructs a NotebookTextModel.
The FileSystem provider that is registered for vscode-interactive schema will always return an empty buffer for any read, and will drop all write requests as nothing is stored on disk for Interactive Window resources. The notebookSerializer that is registered for the interactive viewtype knows to return an empty notebook data model when it deserializes an empty buffer when the model is being resolved.
Restoring the interactive window happens through the EditorModelCache, where the full notebook data is stored, and can be used to repopulate the EditorInput without needing to go through the editor model resolution flow, effectively skipping any filesystem reads.
EditorInput is responsible for resolving models for the given resources but in Interactive Window it's much simpler as we are not resolving models ourselves but delegating to Notebook and TextEditor. InteractiveEditorInput does the coordination job.
