feat: Support inline code completion (FIM)

packages/types/src/global-settings.ts

@@ -232,6 +232,18 @@ export const globalSettingsSchema = z.object({
 	 * Tools in this list will be excluded from prompt generation and rejected at execution time.
 	 */
 	disabledTools: z.array(toolNamesSchema).optional(),
+
+	/**
+	 * FIM (Fill-in-the-Middle) inline code completion settings.
+	 * These are configured separately from the chat model to allow using
+	 * a cheap/fast FIM-specialized model (e.g., DeepSeek Coder, Codestral).
+	 */
+	fimEnabled: z.boolean().optional(),
+	fimProvider: z.enum(["openai-compatible", "deepseek", "mistral", "ollama"]).optional(),
+	fimModelId: z.string().optional(),
+	fimBaseUrl: z.string().optional(),
+	fimDebounceMs: z.number().min(0).optional(),
+	fimMaxTokens: z.number().min(1).optional(),
 })
 
 export type GlobalSettings = z.infer<typeof globalSettingsSchema>
@@ -285,6 +297,7 @@ export const SECRET_STATE_KEYS = [
 // Global secrets that are part of GlobalSettings (not ProviderSettings)
 export const GLOBAL_SECRET_KEYS = [
 	"openRouterImageApiKey", // For image generation
+	"fimApiKey", // For FIM inline code completion
 ] as const
 
 // Type for the actual secret storage keys

src/__mocks__/vscode.js

@@ -96,6 +96,7 @@ export const languages = {
 		clear: () => {},
 		dispose: () => {},
 	}),
+	registerInlineCompletionItemProvider: () => mockDisposable,
 }
 
 export const extensions = {
@@ -152,6 +153,19 @@ export const CodeActionKind = {
 
 export const EventEmitter = mockEventEmitter
 
+export const InlineCompletionTriggerKind = {
+	Invoke: 0,
+	Automatic: 1,
+}
+
+export const InlineCompletionItem = class {
+	constructor(insertText, range, command) {
+		this.insertText = insertText
+		this.range = range
+		this.command = command
+	}
+}
+
 export default {
 	workspace,
 	window,
@@ -171,4 +185,6 @@ export default {
 	EventEmitter,
 	CodeAction,
 	CodeActionKind,
+	InlineCompletionTriggerKind,
+	InlineCompletionItem,
 }

src/extension.ts

@@ -36,6 +36,7 @@ import { openAiCodexOAuthManager } from "./integrations/openai-codex/oauth"
 import { McpServerManager } from "./services/mcp/McpServerManager"
 import { CodeIndexManager } from "./services/code-index/manager"
 import { MdmService } from "./services/mdm/MdmService"
+import { FimService } from "./services/fim"
 import { migrateSettings } from "./utils/migrateSettings"
 import { autoImportSettings } from "./utils/autoImportSettings"
 import { API } from "./extension/api"
@@ -168,6 +169,28 @@ export async function activate(context: vscode.ExtensionContext) {
 
 	const contextProxy = await ContextProxy.getInstance(context)
 
+	// Initialize FIM (Fill-in-the-Middle) inline completion service.
+	const fimService = new FimService(outputChannel)
+	context.subscriptions.push(fimService)
+
+	// Initialize FIM with current settings
+	const initFimSettings = async () => {
+		const globalSettings = contextProxy.getGlobalSettings()
+		const fimApiKey = await context.secrets.get("fimApiKey")
+		fimService.updateSettings(globalSettings, fimApiKey)
+	}
+
+	void initFimSettings()
+
+	// Listen for secret storage changes to update FIM API key
+	context.secrets.onDidChange(async (e: vscode.SecretStorageChangeEvent) => {
+		if (e.key === "fimApiKey") {
+			const globalSettings = contextProxy.getGlobalSettings()
+			const fimApiKey = await context.secrets.get("fimApiKey")
+			fimService.updateSettings(globalSettings, fimApiKey)
+		}
+	})
+
 	// Initialize code index managers for all workspace folders.
 	const codeIndexManagers: CodeIndexManager[] = []
 

src/services/fim/FimApiClient.ts

@@ -0,0 +1,153 @@
+/**
+ * Lightweight API client for FIM (Fill-in-the-Middle) completion requests.
+ *
+ * Supports the `/v1/completions` endpoint used by OpenAI-compatible servers,
+ * DeepSeek, Ollama, and similar providers. This is the legacy completions
+ * endpoint (not chat completions), which is better suited for raw FIM prompts.
+ */
+
+import { formatFimPrompt } from "./FimTokenFormatter"
+
+export interface FimRequestOptions {
+	/** The FIM provider type */
+	provider: "openai-compatible" | "deepseek" | "mistral" | "ollama"
+	/** Base URL for the API endpoint */
+	baseUrl: string
+	/** API key for authentication */
+	apiKey?: string
+	/** Model ID to use */
+	modelId: string
+	/** Text before the cursor */
+	prefix: string
+	/** Text after the cursor */
+	suffix: string
+	/** Maximum tokens to generate */
+	maxTokens: number
+	/** Abort signal for cancellation */
+	signal?: AbortSignal
+}
+
+export interface FimResponse {
+	/** The generated completion text */
+	completion: string
+}
+
+/**
+ * Normalize a base URL by removing trailing slashes.
+ */
+function normalizeBaseUrl(url: string): string {
+	return url.replace(/\/+$/, "")
+}
+
+/**
+ * Build the API endpoint URL based on the provider type.
+ */
+function buildEndpointUrl(provider: string, baseUrl: string): string {
+	const normalized = normalizeBaseUrl(baseUrl)
+
+	switch (provider) {
+		case "ollama":
+			return `${normalized}/api/generate`
+		case "mistral":
+			return `${normalized}/v1/fim/completions`
+		default:
+			// openai-compatible and deepseek use /v1/completions
+			return `${normalized}/v1/completions`
+	}
+}
+
+/**
+ * Build the request body based on the provider type.
+ */
+function buildRequestBody(options: FimRequestOptions): Record<string, unknown> {
+	const { provider, modelId, prefix, suffix, maxTokens } = options
+
+	switch (provider) {
+		case "ollama":
+			return {
+				model: modelId,
+				prompt: prefix,
+				suffix: suffix,
+				stream: false,
+				options: {
+					num_predict: maxTokens,
+					temperature: 0.2,
+					top_p: 0.9,
+				},
+			}
+		case "mistral":
+			return {
+				model: modelId,
+				prompt: prefix,
+				suffix: suffix,
+				max_tokens: maxTokens,
+				temperature: 0.2,
+				top_p: 0.9,
+				stop: ["\n\n"],
+			}
+		default: {
+			// openai-compatible and deepseek: format the FIM prompt with special tokens
+			const prompt = formatFimPrompt(modelId, prefix, suffix)
+			return {
+				model: modelId,
+				prompt,
+				max_tokens: maxTokens,
+				temperature: 0.2,
+				top_p: 0.9,
+				stop: ["\n\n", "<|fim", "<fim_", "[/MIDDLE]"],
+			}
+		}
+	}
+}
+
+/**
+ * Extract the completion text from the provider response.
+ */
+function extractCompletion(provider: string, data: Record<string, unknown>): string {
+	switch (provider) {
+		case "ollama": {
+			return (data.response as string) ?? ""
+		}
+		default: {
+			// OpenAI-compatible response format
+			const choices = data.choices as Array<{ text?: string; message?: { content?: string } }> | undefined
+			if (!choices || choices.length === 0) {
+				return ""
+			}
+			return choices[0].text ?? choices[0].message?.content ?? ""
+		}
+	}
+}
+
+/**
+ * Send a FIM completion request to the configured provider.
+ */
+export async function requestFimCompletion(options: FimRequestOptions): Promise<FimResponse> {
+	const url = buildEndpointUrl(options.provider, options.baseUrl)
+	const body = buildRequestBody(options)
+
+	const headers: Record<string, string> = {
+		"Content-Type": "application/json",
+	}
+
+	if (options.apiKey) {
+		headers["Authorization"] = `Bearer ${options.apiKey}`
+	}
+
+	const response = await fetch(url, {
+		method: "POST",
+		headers,
+		body: JSON.stringify(body),
+		signal: options.signal,
+	})
+
+	if (!response.ok) {
+		const errorText = await response.text().catch(() => "Unknown error")
+		throw new Error(`FIM API request failed (${response.status}): ${errorText}`)
+	}
+
+	const data = (await response.json()) as Record<string, unknown>
+	const completion = extractCompletion(options.provider, data)
+
+	return { completion }
+}