VS Code Language Model provider (vscode-lm)

Routes Cline inference through the VS Code Language Model API (vscode.lm), letting Cline use models contributed by any extension that registers a language model chat provider (a vendor) with VS Code. GitHub Copilot is the most common such vendor, but the implementation is vendor-agnostic — it selects models via vscode.lm.selectChatModels and has no Copilot-specific logic. VS Code only — the vscode.lm API does not exist on other hosts (e.g. JetBrains), so the provider is gated on the API being present.

Files

• vscode-lm-handler.ts — VsCodeLmHandler, a Cline SDK ApiHandler (@cline/llms) backed by vscode.lm. Selects a chat model, streams the response, forwards tool definitions, and surfaces tool calls and usage.
• vscode-lm-format.ts — converts SDK Messages to vscode.LanguageModelChatMessages and back-converts tool results.
• register-vscode-lm.ts — registers the handler with the SDK handler registry and exposes the vscode.lm availability check used for gating.

How it plugs into the SDK

The SDK's @cline/llms handler registry (registerHandler) exists for providers that need host-only dependencies — here, vscode.lm — which cannot live in the host-agnostic SDK package. registerVsCodeLmHandler() (called during extension activation) registers the factory for the vscode-lm provider id when the API is available. The SDK then resolves vscode-lm to this handler for both the main task loop and standalone utility calls.

Model selection

ProviderConfig has no field for a VS Code LM selector, so the selected model travels as a vendor/family[/version/id] string in ProviderConfig.modelId — the model-id channel the rest of the SDK adapter uses. The webview stores the structured LanguageModelChatSelector in plan/actModeVsCodeLmModelSelector; the session factory stringifies it into modelId, and parseVsCodeLmSelector reads it back into a selector here.

Tool calling

Tool definitions are passed to sendRequest via LanguageModelChatRequestOptions.tools (with toolMode: Auto). Tool invocations arrive as LanguageModelToolCallPart and are forwarded as tool-call chunks the SDK runtime executes.

Tool results round-trip as LanguageModelToolResultPart. Two details the VS Code LM API requires:

• Result content is serialized to text. SDK tool executors return rich, untyped result objects (e.g. ToolOperationResult { query, result, success }), not typed text blocks; extractToolOutputText pulls their text out so the model receives the actual output.
• The prompt must not end on a tool result. Copilot models require the final message to be a non-tool-result user message; when a turn ends on tool results, a short trailing user message is appended.