add deferred tools support for kimi in openai-completions api

packages/ai/src/api/openai-completions.ts

@@ -77,6 +77,26 @@ function hasToolHistory(messages: Message[]): boolean {
 	return false;
 }
 
+function getDeferredToolNames(messages: Message[]): Set<string> {
+	const names = new Set<string>();
+	for (const message of messages) {
+		if (message.role === "toolResult") {
+			for (const name of message.addedToolNames ?? []) {
+				names.add(name);
+			}
+		}
+	}
+	return names;
+}
+
+function getToolsByName(tools: Tool[] | undefined, names: Iterable<string>): Tool[] {
+	if (!tools) return [];
+	const toolsByName = new Map(tools.map((tool) => [tool.name, tool]));
+	return Array.from(names)
+		.map((name) => toolsByName.get(name))
+		.filter((tool): tool is Tool => tool !== undefined);
+}
+
 function isTextContentBlock(block: { type: string }): block is TextContent {
 	return block.type === "text";
 }
@@ -117,14 +137,23 @@ interface OpenAICompatCacheControl {
 	ttl?: string;
 }
 
-type ResolvedOpenAICompletionsCompat = Omit<Required<OpenAICompletionsCompat>, "cacheControlFormat"> & {
+type ResolvedOpenAICompletionsCompat = Omit<
+	Required<OpenAICompletionsCompat>,
+	"cacheControlFormat" | "deferredToolsMode"
+> & {
 	cacheControlFormat?: OpenAICompletionsCompat["cacheControlFormat"];
+	deferredToolsMode?: OpenAICompletionsCompat["deferredToolsMode"];
 };
 
 type ResolvedChatTemplateKwargValue = string | number | boolean | null;
 
 type ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;
 
+type KimiToolSystemMessageParam = {
+	role: "system";
+	tools: OpenAI.Chat.Completions.ChatCompletionTool[];
+};
+
 type OpenAIEncryptedReasoningDetail = {
 	type: "reasoning.encrypted";
 	id: string;
@@ -585,8 +614,11 @@ function buildParams(
 		params.temperature = options.temperature;
 	}
 
-	if (context.tools && context.tools.length > 0) {
-		params.tools = convertTools(context.tools, compat);
+	const deferredToolNames =
+		compat.deferredToolsMode === "kimi" ? getDeferredToolNames(context.messages) : new Set<string>();
+	const activeTools = context.tools?.filter((tool) => !deferredToolNames.has(tool.name));
+	if (activeTools && activeTools.length > 0) {
+		params.tools = convertTools(activeTools, compat);
 		if (compat.zaiToolStream) {
 			(params as any).tool_stream = true;
 		}
@@ -1025,6 +1057,7 @@ export function convertMessages(
 			params.push(assistantMsg);
 		} else if (msg.role === "toolResult") {
 			const imageBlocks: Array<{ type: "image_url"; image_url: { url: string } }> = [];
+			const deferredToolNames = new Set<string>();
 			let j = i;
 
 			for (; j < transformedMessages.length && transformedMessages[j].role === "toolResult"; j++) {
@@ -1051,6 +1084,12 @@ export function convertMessages(
 				}
 				params.push(toolResultMsg);
 
+				if (compat.deferredToolsMode === "kimi") {
+					for (const name of toolMsg.addedToolNames ?? []) {
+						deferredToolNames.add(name);
+					}
+				}
+
 				if (hasImages && model.input.includes("image")) {
 					for (const block of toolMsg.content) {
 						if (isImageContentBlock(block)) {
@@ -1089,6 +1128,18 @@ export function convertMessages(
 			} else {
 				lastRole = "toolResult";
 			}
+
+			if (deferredToolNames.size > 0) {
+				const deferredTools = getToolsByName(context.tools, deferredToolNames);
+				if (deferredTools.length > 0) {
+					const kimiToolMessage: KimiToolSystemMessageParam = {
+						role: "system",
+						tools: convertTools(deferredTools, compat),
+					};
+					// Kimi accepts a system message with tools but omits the standard content field.
+					params.push(kimiToolMessage as unknown as ChatCompletionMessageParam);
+				}
+			}
 			continue;
 		}
 
@@ -1256,6 +1307,7 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 		supportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,
 		cacheControlFormat,
 		sendSessionAffinityHeaders: false,
+		deferredToolsMode: undefined,
 		sessionAffinityFormat: isOpenRouter ? "openrouter" : "openai",
 		supportsLongCacheRetention: !(
 			isTogether ||
@@ -1296,6 +1348,7 @@ function getCompat(model: Model<"openai-completions">): ResolvedOpenAICompletion
 		supportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,
 		cacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,
 		sendSessionAffinityHeaders: model.compat.sendSessionAffinityHeaders ?? detected.sendSessionAffinityHeaders,
+		deferredToolsMode: model.compat.deferredToolsMode ?? detected.deferredToolsMode,
 		sessionAffinityFormat: model.compat.sessionAffinityFormat ?? detected.sessionAffinityFormat,
 		supportsLongCacheRetention: model.compat.supportsLongCacheRetention ?? detected.supportsLongCacheRetention,
 	};

packages/ai/src/types.ts

@@ -524,6 +524,8 @@ export interface OpenAICompletionsCompat {
 	cacheControlFormat?: "anthropic";
 	/** Whether to send session-affinity data from `options.sessionId`. Default: false. */
 	sendSessionAffinityHeaders?: boolean;
+	/** Provider-specific deferred tool serialization mode. */
+	deferredToolsMode?: "kimi";
 	/** Session-affinity header format: `openai` sends `session_id`, `x-client-request-id`, and `x-session-affinity`; `openai-nosession` sends `x-client-request-id` and `x-session-affinity`; `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param, which is governed by cache retention. Default: auto-detected. */
 	sessionAffinityFormat?: SessionAffinityFormat;
 	/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: "24h"` or Anthropic-style `cache_control.ttl: "1h"`, depending on format). Default: true. */