Investigate debug visualization extension points

[connor4312]

VS proper has some data visualization like this:

...and personally I've never been a fan of how we represent things like XML elements or binary data in debug hovers and the REPL. In large part this is because we have to have a 'generic' view that gives all possible information, so we can't be more concise when showing things like DOM tree.

So it would be nice to have a way for visualizations to be contributed, with similar affordances to what VS provides.

• These should be contributable by extensions
• These should not be custom webviews. At least not immediately, I have no desired to re-implement the layering of notebooks' webviews in the debug REPL. (Though perhaps we could allow webviews in the hover and as popouts in the console...)
• This does not affect DAP in any way. This is a UI extension, not a protocol extension.
• This should allow the hex editor to move to using the generic APIs and away from the custom handling we currently have.

My initial thought at an API is something like this. Note that we don't bring in DAP types in vscode.d.ts, so they would be typed as unknown and cast by the extension author.

/** Contribution point in package.json `debugVisualizers` */
type DebugVisualizerContributionPoint = Array<{
  // Corresponding ID in the `registerDebugVisualizationProvider`
  id: string;
  // `when` clause that determines when the visualizer may be shown
  when: string;
}>

declare module 'vscode' {
    export namespace debug {
        /**
         * Registers a custom data visualization for variables when debugging.
         *
         * @param id The corresponding ID in the package.json `debugVisualizers` contribution point.
         * @param provider The {@link DebugVisualizationProvider} to register
         */
        export function registerDebugVisualizationProvider<T extends DebugVisualization>(
            id: string,
            provider: DebugVisualizationProvider<T>
        ): Disposable;
    }

    export class DebugVisualization {
        /**
         * The name of the visualization to show to the user.
         */
        name: string;

        /**
         * An icon for the view when it's show in inline actions.
         */
        iconPath?: Uri | { light: Uri; dark: Uri } | ThemeIcon;

        /**
         * Visualization to use for the variable. This may be either:
         * - A command to run when the visualization is selected for a variable.
         * - A {@link TreeDataProvider} which is used to display the data in-line
         *   where the variable is shown. If a single root item is returned from
         *   the data provider, it will replace the variable in its tree.
         *   Otherwise, the items will be shown as children of the variable.
         */
        visualization?: Command | TreeDataProvider<unknown>;

        /**
         * Creates a new debug visualization object.
         * @param name Name of the visualization to show to the user.
         */
        constructor(name: string);
    }

    export interface DebugVisualizationProvider<T extends DebugVisualization = DebugVisualization> {
        /**
         * Called for each variable when the debug session stops. It should return
         * any visualizations the extension wishes to show to the user.
         *
         * Note that this is only called when its `when` clause defined under the
         * `debugVisualizers` contribution point in the `package.json` evaluates
         * to true.
         */
        provideDebugVisualization(context: DebugVisualizationContext, token: CancellationToken): ProviderResult<T[]>;

        /**
         * Invoked for a variable when a user picks the visualizer.
         *
         * It may return a {@link TreeView} that's shown in the Debug Console or
         * inline in a hover. A visualizer may choose to return `undefined` from
         * this function and instead trigger other actions in the UI, such as opening
         * a custom {@link WebviewView}.
         */
        resolveDebugVisualization(visualization: T, token: CancellationToken): ProviderResult<T>;
    }

    export interface DebugVisualizationContext {
        /**
         * The Debug Adapter Protocol Variable to be visualized.
         * @see https://microsoft.github.io/debug-adapter-protocol/specification#Types_Variable
         */
        variable: any;
        /**
         * The Debug Adapter Protocol variable reference the type (such as a scope
         * or another variable) that contained this one. Empty for variables
         * that came from user evaluations in the Debug Console.
         * @see https://microsoft.github.io/debug-adapter-protocol/specification#Types_Variable
         */
        containerId?: number;
        /**
         * The ID of the Debug Adapter Protocol StackFrame in which the variable was found,
         * for variables that came from scopes in a stack frame.
         * @see https://microsoft.github.io/debug-adapter-protocol/specification#Types_StackFrame
         */
        frameId?: string;
        /**
         * The debug session the variable belongs to.
         */
        session: DebugSession;
    }
}

cc @hediet @roblourens @joj

[connor4312]

@thegecko and I had some discussion on the linked issue, continuing here https://github.com/microsoft/vscode/pull/197458#issuecomment-1793666100

[hediet]

I'd love to have better support for visualizations!

I think a fundamental question is if these visualizations should be a modal one-time experience (as in VS https://youtu.be/1Nq4E4aN1WA?t=85, maybe you can see it as copilots "inline chat") or a more interactive non-modal experience (as in my debug visualizer extension, which would correspond to copilots ghost text once you setup the visualization).

The modal experience is much easier to implement, but much more inconvenient to use when you want to inspect how value visualizations change as you step through your code, as the visualization disappears when you step. In particular, the non-modal experience has to remember which visualization you selected the last time for a given variable, so you don't have to select it again the next time you want to visualize the same variable.

I would really like to see the non-modal experience, as it is a much more useful feature.

My initial thought at an API is something like this.

This basically replaces the "generic" tree in the hover with a specialized, extension-contributed one, right?

I think extensions might want to be able to return ProviderResult<TreeView | undefined> to indicate they cannot visualize the given data (even though it matched the selector). E.g. a base64 hex viewer is only applicable to strings that match [A-Fa-f0-9=]+, which you cannot easily express statically through a selector. It would also be nice if one VisualizationProvider could return multiple possible visualizations (ideally without having to compute the tree yet). Then visualization extensions can delegate listing the visualizations to the program to debug.

Maybe the non-modal approach could work like this (similar to how SuggestProviders work):

namespace debug {
  /**
   * Registers a custom data visualization for variables when debugging.
   * 
   * @param selector Identifies data this visualizer applies to. It will only
   * be shown for DAP variables whose properties are a superset of the
   * `variable` property, and whose session type is `debugType`.
   * 
   * @param provideView Invoked for a variable when a user picks the visualizer.
   * It may return a {@link TreeView} that's shown in the Debug Console or
   * inline in a hover.
   */
  export function registerDebugVisualizationProvider<T extends IDebugVisualization>(
    selector: { debugType?: string; variable?: Dap.Variable },

    // Serves as dynamic discoverability mechanism (selector applies first though).
    provide: (context: IDebugVisualizationContext) => ProviderResult<T[] | undefined>,

    // Allows to delay computing `treeView`
    resolve: (visualization: T) => Promise<void>;
  ): Disposable;

  export interface IDebugVisualization {
     name: string; // A human readable (localized) name of this visualization
     id: string; // The id to remember which visualization was picked the last time
     treeView?: TreeView; // If this visualization is selected, this tree view replaces the "generic" tree view of the value
     visualizeCommand?: ICommand; // This command is shown next to the value when an inline visualization cannot be done.
  }

  export interface IDebugVisualizationContext {
    /** The variable the user wants to visualize */
    variable: Dap.Variable;
    /** The container the variable was set in. */
    // authors note: this and the name is required for use of `setVariable` / `setExpression`
    container: Dap.Variable;
    /** The name of the variable in its container. */
    name: string;
    /** The debug session the variable belongs to.  */
    session: DebugSession;
    /** The currently selected debug visualization for this value. This gives extensions a chance to use a single DAP request for discovery and visualization computation. In that case, a resolve call wouldn't have to send a second DAP request to actually compute the visualization */
    preferredDebugVisualizationId?: string;
  }
}

These should not be custom webviews.

I think the super power from visualizations really comes from custom UI. But I understand its a hassle to implement (also, because some visualization libraries are large and take time to load, you want to reuse the webview). ~A nice escape hatch for this could be https://github.com/microsoft/vscode/issues/125763 where extensions could then open the webview on their own when the user clicks on some "visualize" command.~ You already seem to have thought of this.

[thegecko]

It may be useful to outline some real-world use cases which this proposal may (or may not!) be designed to support.

1. As mentioned in #155597 / #197458 we want to be able to change the extension/view a user can use to interact with variable data
2. Allow addition of functionality to change the format of a variable display (e.g. hex, binary, etc.)
3. Add the ability to search/filter variables. Perhaps this should be a different implementation of the variables view contributed by an extension using a webview?

Some further thoughts:

1. Personally I'd prefer to avoid modal dialogs and instead guide the user to opening specific views (webview or otherwise) which can remain open and support a multi-document interface. In our experience the number one feature requested by developers during debug is the ability to put different windows on different monitors in order to view lots of information at once. We should align with the Floating Windows Editor Exploration work ;).
2. Can the existing HexEdit hover button thing be made an extension point or removed?
3. Should we allow existing panels to be entirely replaced by a contribution (e.g. for the enhanced variable view outlined above). Panels can be added, but can default ones be hidden to avoid confusion?

[connor4312]

@hediet

Persistence: it would be possible for an extension who opens its own webview to also have a DebugAdapterTracker to update a webview as the state changes.

Provider/resolver: I agree, that is a good idea. Updated the API in the main issue. "This basically replaces the "generic" tree in the hover with a specialized, extension-contributed one, right?" -- yes

Webviews: yes, webviews are more of a nightmare when it comes to showing them inside tree views. This is a path trod by notebooks in years past. For performance, showing many individual webviews is a poor experience, so notebooks ended up with the editor layer and webview layer. But getting this right has taken years and is still presenting challenges, so I have no appetite at all to do that here. But, as mentioned, inline 'popouts' might be a way to make this happen, and could be added in the future by adding some extra | WebviewPopout to the resolve method's return type.

---

@thegecko

I'll update the main issue with scenarios to support. Searching/filter: I don't think this is in scope for this work.

1. If/when multi-window supports webviews (I haven't tried it personally) then I would expect such webview-visualizing extensions to work automatically with this API :)
2. Yes, it should also use this API
3. Technically one could already replace the existing variable views if they wanted to by registering custom views, a DebugAdapterTracker, and optionally automatically running commands to hide the built-in views. But I don't think this is necessarily in-scope for this work (and is not something we'd probably build a special path to support, in general)

---

Other thought. I realized the "variable" selector as I have defined it is not very good, since e.g. the common case thegecko is looking would want any defined memoryReference.

If we continue to go with an object selector, I would type it as variable: Record<string, boolean | string | RegExp>with the rules:

• The "variable" selector is a mapping of property names on the Variable to selection logic. All defined properties must match for the visualizer to be available for a variable.
  • A value of "true" will match any non-null value for the property
  • A string will match if the corresponding string in the variable is the same
  • A regular expression will match if the corresponding property in the variable, when cast to a string, matches the expression
• For example, { memoryReference: true, type: /string|buffer/ } would match any variable that has a defined memoryReference and set their type to string or buffer

I don't want to have a more tailored selector with properties like hasMemoryReference: boolean to avoid leaking DAP knowledge into vscode.d.ts.

[thegecko]

Other thought. I realized the "variable" selector as I have defined it is not very good

We currently use when clauses in the manifest file to ensure things appear at the right point, I wonder if that system can be leveraged?

It just so happens there exists a canViewMemory clause in VS Code which is true when the current variable has a memoryReference.

[vogelsgesang]

This should allow the hex editor to move to using the generic APIs and away from the custom handling we currently have.

I would find it particularly nice, if the "visualization extension points" could also be reused not only for the hex editor, but also for "Go to source location" proposed in https://github.com/microsoft/debug-adapter-protocol/issues/372. Afaict, the current proposal would already support adding such a "Go to source location" button, or am I missing something?

[vogelsgesang]

also, CC @vadimcn since the proposed extension points here might also be very useful for https://github.com/vadimcn/codelldb and its data visualizations

[hediet]

A visualizer may choose to return undefined from

• this function and instead trigger other actions in the UI, such as opening
• a custom {@link WebviewView}. */ resolveDebugVisualization(visualization: T): ProviderResult<TreeView | undefined>;

I think this should align with the resolve idea/signature of completion providers (I'm very sure Joh will insist on this anyway). Also, I don't think resolve should cause side effects, such as opening a webview. This will make it impossible to automatically update the tree view (e.g. in the watch window) when you step through the code.

Rather it should return the command that does this.

WebviewPopout

That would be nice!

[connor4312]

We currently use when clauses in the manifest file to ensure things appear at the right point, I wonder if that system can be leveraged?

Good point, we should do this for extension activation anyway.

I think this should align with the resolve idea/signature of completion providers (I'm very sure Joh will insist on this anyway)... Also, I don't think resolve should cause side effects, such as opening a webview.

Maybe, we do have precedent for resolve* causing side-effects, such as resolveWebviewView. Will discuss it.

[connor4312]

I've done some initial work on this, as well as initial refinement in the API, which is in the main branch of VS Code. It currently supports command-generating visualizers. You can see an example of a basic 'base64 visualizer' in https://github.com/microsoft/vscode-extension-samples/tree/connor4312/debug-viz-demo/proposed-api-sample, and try it in VS Code Insiders.

• The package.json contribution tells VS Code to try to resolve variables that look like base64 with the 'base64Decoder' visualizer. VS Code activates the extension on-demand.
• This is registered in the extension.ts with a provider that just returns a command for the editor to dispatch
• Selecting the visualizer opens a new untitled editor with the decoded value

[thegecko]

@connor4312 nice work! I think this is an excellent addition and looks very flexible.

(obviously not base64 but modified to get the viz to show)

Are there any specific things you want feedback or testing around?

[connor4312]

Any feedback you want to provide on the initial version is great -- though the current exposure is fairly simple and "works for me" is also good feedback 🙂

The more gnarly part will be letting it support subtrees, which I'll look at this coming iteration.

[starball5]

This sounds interesting. I wonder how it'll affect or help with C++ things (Natvis and GDB pretty printers, which I think the MS C/C++ extension currently supports).

[thegecko]

Any feedback you want to provide on the initial version is great -- though the current exposure is fairly simple and "works for me" is also good feedback 🙂

Well, it "works for me"! As long as the hex edit recommendation is removed when this is merged :)

Playing some more, I notice multiple visualizations are put into a menu:

Would it be worth supporting group names (including inline) to control this and group accordingly if multiple appear?

The more gnarly part will be letting it support subtrees, which I'll look at this coming iteration.

Is this supporting branches further down the graph?

[aiday-mar]

Hi @connor4312 , since this is a big work item, I suppose this will not be finalized this milestone, hence I am changing the milestone to March 2024. Feel free to change it back if you'd like to make this a candidate issue.

[hediet]

Just tried out the API again.

What when context should I use when I can visualize everything? "true" did not work.

Also, it seems like it only works in the variable view, not in the watch view or with hovers. Would be very nice if this would be supported there as well.

[thegecko]

@connor4312 can we have an update on this being merged, please? We are keen to promote our new memory inspector in the VS Code UI: https://marketplace.visualstudio.com/items?itemName=eclipse-cdt.memory-inspector