feat(condense): allow provider profile override

Author: twinlunarstarz-dev

packages/types/src/global-settings.ts

@@ -116,6 +116,8 @@ export const globalSettingsSchema = z.object({
 	openRouterImageApiKey: z.string().optional(),
 	openRouterImageGenerationSelectedModel: z.string().optional(),
 
+	condensingApiConfigOverride: z.boolean().optional(),
+	condensingApiConfigId: z.string().optional(),
 	customCondensingPrompt: z.string().optional(),
 
 	autoApprovalEnabled: z.boolean().optional(),

packages/types/src/vscode-extension-host.ts

@@ -303,6 +303,8 @@ export type ExtensionState = Pick<
 	| "customModePrompts"
 	| "customSupportPrompts"
 	| "enhancementApiConfigId"
+	| "condensingApiConfigOverride"
+	| "condensingApiConfigId"
 	| "customCondensingPrompt"
 	| "codebaseIndexConfig"
 	| "codebaseIndexModels"

src/core/condense/__tests__/index.spec.ts

@@ -1250,6 +1250,60 @@ describe("summarizeConversation with custom settings", () => {
 		expect(createMessageCalls[0][0]).toContain("CRITICAL: This is a summarization-only request")
 	})
 
+	it("should use the selected condensing handler while counting the resulting context with the current handler", async () => {
+		const condensingApiHandler = {
+			createMessage: vi.fn().mockImplementation(() => {
+				return (async function* () {
+					yield { type: "text" as const, text: "Summary from selected profile" }
+					yield { type: "usage" as const, totalCost: 0.01, outputTokens: 40 }
+				})()
+			}),
+			getModel: vi.fn().mockReturnValue({
+				id: "condensing-model",
+				info: {
+					contextWindow: 4000,
+					supportsImages: false,
+					supportsVision: false,
+					maxTokens: 2000,
+				},
+			}),
+		} as unknown as ApiHandler
+
+		const result = await summarizeConversation({
+			messages: sampleMessages,
+			apiHandler: mockMainApiHandler,
+			condensingApiHandler,
+			systemPrompt: defaultSystemPrompt,
+			taskId: localTaskId,
+		})
+
+		expect(condensingApiHandler.createMessage).toHaveBeenCalledOnce()
+		expect(mockMainApiHandler.createMessage).not.toHaveBeenCalled()
+		expect(mockMainApiHandler.countTokens).toHaveBeenCalled()
+		expect(result.summary).toBe("Summary from selected profile")
+	})
+
+	it("should fall back to the current handler when the selected condensing handler is invalid", async () => {
+		const warnSpy = vi.spyOn(console, "warn").mockImplementation(() => {})
+		const invalidHandler = {
+			getModel: vi.fn(),
+		} as unknown as ApiHandler
+
+		try {
+			await summarizeConversation({
+				messages: sampleMessages,
+				apiHandler: mockMainApiHandler,
+				condensingApiHandler: invalidHandler,
+				systemPrompt: defaultSystemPrompt,
+				taskId: localTaskId,
+			})
+		} finally {
+			warnSpy.mockRestore()
+		}
+
+		expect(mockMainApiHandler.createMessage).toHaveBeenCalledOnce()
+	})
+
 	/**
 	 * Test that telemetry is called for custom prompt usage
 	 */

src/core/condense/index.ts

@@ -224,6 +224,8 @@ export type SummarizeResponse = {
 export type SummarizeConversationOptions = {
 	messages: ApiMessage[]
 	apiHandler: ApiHandler
+	/** Optional API handler used only for the summarization request. Token counting still uses apiHandler. */
+	condensingApiHandler?: ApiHandler
 	systemPrompt: string
 	taskId: string
 	isAutomaticTrigger?: boolean
@@ -257,6 +259,7 @@ export async function summarizeConversation(options: SummarizeConversationOption
 	const {
 		messages,
 		apiHandler,
+		condensingApiHandler,
 		systemPrompt,
 		taskId,
 		isAutomaticTrigger,
@@ -311,8 +314,17 @@ export async function summarizeConversation(options: SummarizeConversationOption
 	// This is necessary because some providers (like Bedrock via LiteLLM) require the `tools` parameter
 	// when tool blocks are present. By converting them to text, we can send the conversation for
 	// summarization without needing to pass the tools parameter.
+	const handlerToUse =
+		condensingApiHandler && typeof condensingApiHandler.createMessage === "function"
+			? condensingApiHandler
+			: apiHandler
+
+	if (condensingApiHandler && handlerToUse !== condensingApiHandler) {
+		console.warn("Selected API handler for condensing is invalid; using the current mode's API handler.")
+	}
+
 	const messagesWithTextToolBlocks = transformMessagesForCondensing(
-		maybeRemoveImageBlocks([...messagesWithToolResults, finalRequestMessage], apiHandler),
+		maybeRemoveImageBlocks([...messagesWithToolResults, finalRequestMessage], handlerToUse),
 	)
 
 	const requestMessages = messagesWithTextToolBlocks.map(({ role, content }) => ({ role, content }))
@@ -321,7 +333,7 @@ export async function summarizeConversation(options: SummarizeConversationOption
 	const promptToUse = SUMMARY_PROMPT
 
 	// Validate that the API handler supports message creation
-	if (!apiHandler || typeof apiHandler.createMessage !== "function") {
+	if (!handlerToUse || typeof handlerToUse.createMessage !== "function") {
 		console.error("API handler is invalid for condensing. Cannot proceed.")
 		const error = t("common:errors.condense_handler_invalid")
 		return { ...response, error }
@@ -332,7 +344,7 @@ export async function summarizeConversation(options: SummarizeConversationOption
 	let outputTokens = 0
 
 	try {
-		const stream = apiHandler.createMessage(promptToUse, requestMessages, metadata)
+		const stream = handlerToUse.createMessage(promptToUse, requestMessages, metadata)
 
 		for await (const chunk of stream) {
 			if (chunk.type === "text") {

src/core/context-management/index.ts

@@ -251,6 +251,7 @@ export type ContextManagementOptions = {
 	contextWindow: number
 	maxTokens?: number | null
 	apiHandler: ApiHandler
+	condensingApiHandler?: ApiHandler
 	autoCondenseContext: boolean
 	autoCondenseContextPercent: number
 	systemPrompt: string
@@ -294,6 +295,7 @@ export async function manageContext({
 	contextWindow,
 	maxTokens,
 	apiHandler,
+	condensingApiHandler,
 	autoCondenseContext,
 	autoCondenseContextPercent,
 	systemPrompt,
@@ -361,6 +363,7 @@ export async function manageContext({
 			const result = await summarizeConversation({
 				messages,
 				apiHandler,
+				condensingApiHandler,
 				systemPrompt,
 				taskId,
 				isAutomaticTrigger: true,

src/core/task/Task.ts

@@ -1572,6 +1572,37 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 		}
 	}
 
+	private async getCondensingApiHandler(
+		state:
+			| {
+					condensingApiConfigOverride?: boolean
+					condensingApiConfigId?: string
+					listApiConfigMeta?: Array<{ id: string }>
+			  }
+			| undefined,
+	): Promise<ApiHandler | undefined> {
+		const configId = state?.condensingApiConfigId
+		if (!state?.condensingApiConfigOverride || !configId) {
+			return undefined
+		}
+
+		if (!state.listApiConfigMeta?.some((config) => config.id === configId)) {
+			console.warn(`[Task] Context condensing profile "${configId}" no longer exists; using the current mode.`)
+			return undefined
+		}
+
+		try {
+			const profile = await this.providerRef.deref()?.providerSettingsManager.getProfile({ id: configId })
+			return profile?.apiProvider ? buildApiHandler(profile) : undefined
+		} catch (error) {
+			console.warn(
+				`[Task] Failed to load context condensing profile "${configId}"; using the current mode.`,
+				error,
+			)
+			return undefined
+		}
+	}
+
 	public async condenseContext(): Promise<void> {
 		// CRITICAL: Flush any pending tool results before condensing
 		// to ensure tool_use/tool_result pairs are complete in history
@@ -1583,6 +1614,7 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 		const state = await this.providerRef.deref()?.getState()
 		const customCondensingPrompt = state?.customSupportPrompts?.CONDENSE
 		const { mode, apiConfiguration } = state ?? {}
+		const condensingApiHandler = await this.getCondensingApiHandler(state)
 
 		const { contextTokens: prevContextTokens } = this.getTokenUsage()
 
@@ -1638,6 +1670,7 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 		} = await summarizeConversation({
 			messages: this.apiConversationHistory,
 			apiHandler: this.api,
+			condensingApiHandler,
 			systemPrompt,
 			taskId: this.taskId,
 			isAutomaticTrigger: false,
@@ -3802,6 +3835,8 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	private async handleContextWindowExceededError(): Promise<void> {
 		const state = await this.providerRef.deref()?.getState()
 		const { profileThresholds = {}, mode, apiConfiguration } = state ?? {}
+		const customCondensingPrompt = state?.customSupportPrompts?.CONDENSE
+		const condensingApiHandler = await this.getCondensingApiHandler(state)
 
 		const { contextTokens } = this.getTokenUsage()
 		const modelInfo = this.api.getModel().info
@@ -3876,10 +3911,12 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 				maxTokens,
 				contextWindow,
 				apiHandler: this.api,
+				condensingApiHandler,
 				autoCondenseContext: true,
 				autoCondenseContextPercent: FORCED_CONTEXT_REDUCTION_PERCENT,
 				systemPrompt: await this.getSystemPrompt(),
 				taskId: this.taskId,
+				customCondensingPrompt,
 				profileThresholds,
 				currentProfileId,
 				metadata,
@@ -4042,11 +4079,13 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 				lastMessageTokens,
 				useAvailableInputForContextPercent,
 			})
+			let condensingApiHandler: ApiHandler | undefined
 
 			// Send condenseTaskContextStarted BEFORE manageContext to show in-progress indicator
 			// This notification must be sent here (not earlier) because the early check uses stale token count
 			// (before user message is added to history), which could incorrectly skip showing the indicator
 			if (contextManagementWillRun && autoCondenseContext) {
+				condensingApiHandler = await this.getCondensingApiHandler(state)
 				await this.providerRef
 					.deref()
 					?.postMessageToWebview({ type: "condenseTaskContextStarted", text: this.taskId })
@@ -4111,6 +4150,7 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 					maxTokens,
 					contextWindow,
 					apiHandler: this.api,
+					condensingApiHandler,
 					autoCondenseContext,
 					autoCondenseContextPercent,
 					systemPrompt,