This is a loose collection of guidelines that you should be following when proposing API. The process for adding API is the following
- read and understand this document
- add API into the
vscode.proposed.d.tsfile - trigger a discussion about the proposed API
We DO NOT want to break API. Therefore be careful and conservative when proposing new API. It needs to hold up in the long term. Expose only the minimum but still try to anticipate potential future requests.
The API is structured into different namespaces, like commands, window, workspace etc. Namespaces contain functions, constants, and events. All types (classes, enum, interfaces) are defined in the global, vscode, namespace.
The API should have a JavaScript’ish feel. While that is harder to put in rules, it means we use namespaces, properties, functions, and globals instead of object-factories and services. Also take inspiration from popular existing JS API, for instance window.createStatusBarItem is like document.createElement, the members of DiagnosticsCollection are similar to ES6 maps etc.
Events aren’t defined on the types they occur on but in the best matching namespace. For instance, document changes aren’t send by a document but via the workspace.onDidChangeTextDocument event. The event will contain the document in question. This global event pattern makes it easier to manage event subscriptions because changes happen less frequently.
Events follow the on[Did|Will]VerbSubject patterns, like onDidChangeActiveEditor or onWillSaveTextDocument. It doesn’t hurt to use explicit names.
Objects that live in the main thread but can be controlled/instantiated by extensions are declared as interfaces, e.g. TextDocument or StatusBarItem. When you allow to create such objects your API must follow the createXYZ(args): XYZ pattern. Because this is a constructor-replacement, the call must return synchronously.
Objects the API hands out to extensions should not contain more than what the API defines. Don’t expect everyone to read vscode.d.ts but also expect folks to use debugging-aided-intellisense, meaning whatever the debugger shows developers will program against. We don’t want to appear as making false promises. Prefix your private members with _ as that is a common rule or even better use function-scopes to hide information.
Reading data, like an editor selection, a configuration value, etc is synchronous. Setting state that reflects on the main side is asynchronous. Despite updates being async your ‘extension host object’ should reflect the new state synchronously. This happens when setting an editor selection
editor.selection = newSelection
|
|
V
1. On the API object set the value as given
2. Make an async-call to the main side ala `trySetSelection`
3. The async-call returns with the actual selection (it might have changed in the meantime)
4. On the API object set the value again
We usually don’t expose the fact that setting state is asynchronous. We try to have API that feels sync -editor.selection is a getter/setter and not a method.
Whenever possible you should define a data model and define provider-interfaces. This puts VS Code into control as we can decide when to ask those providers, how to deal with multiple providers etc. The ReferenceProvider interface is a good sample for this.
Sometimes it is expensive for a provider to compute parts of its data. For instance, creating a full CompletionItem (with all the documentation and symbols resolved) conflicts with being able to compute a large list of them quickly. In those cases, providers should return a lightweight version and offer a resolve method that allows extensions to enrich data. The CodeLensProvider and CompletionItemProvider interfaces are good samples for this.
Calls into a provider should always include a CancellationToken as the last parameter. With that the main thread can signal to the provider that its result won’t be needed anymore. When adding new parameters to provider-functions it is OK to have the token not at the end anymore.
Objects that should be returned by a provider are usally represented by a class that extensions can instantiate, e.g. CodeLens. We do that to provide convience constructors and to be able to populate default values.
Data that we accept in methods calls, i.e. parameter types, like in registerRenameProvider or showQuickPick, are declared as interfaces. That makes it easy to fullfill the API contract using class-instances or plain object literals.
Data the API returns is strict, e.g activeTextEditor is an editor or undefined, but not null. On the other side, providers can return relaxed data. We usually accept 4 types: The actual type, like Hover, a Thenable of that type, undefined or null. With that we want to make it easy to implement a provider, e.g. if you can compute results synchronous you don’t need to wrap things into a promise or if a certain condition isn’t met simple return etc.
Although providers can return ‘relaxed’ data you need to verify it. The same is true for arguments etc. Throw validation errors when possible, drop data object when invalid.
Don’t send the data that a provider returned over the wire. Often it containa more information than we need and often there are cyclic dependencies. Use the provider data to create objects that your protocol speaks.
When API-work started only numeric-enums were supported, today TypeScript supports string-or-types and string-enums. Because fewer concepts is better, we stick to numeric-enums.
We define the API with strictNull-checks in mind. That means we use the optional annotation foo?: number and null or undefined in type annotations. For instance, its activeTextEditor: TextEditor | undefined. Again, be strict for types we define and relaxed when accepting data.
We add JSDoc for all parts of the API. The doc is supported by markdown syntax. When document string-datatypes that end up in the UI, use the phrase ‘Human-readable string…’