@ai-sdk/alibaba 1.0.53 → 1.0.55

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # @ai-sdk/alibaba
 
+## 1.0.55
+
+### Patch Changes
+
+- def3cdf: feat(alibaba): preserve reasoning in multi-turn requests by default on supported models
+
+## 1.0.54
+
+### Patch Changes
+
+- 4bd2599: fix(deepseek): preserve reasoning streams across empty tool-call deltas
+
 ## 1.0.53
 
 ### Patch Changes

package/README.md

@@ -63,6 +63,60 @@ console.log('Reasoning:', reasoningText);
 console.log('Answer:', text);
 ```
 
+## Preserved Thinking Example (Multi-Turn Reasoning)
+
+For models that support preserved thinking, the AI SDK sends reasoning from
+previous assistant messages back as Alibaba `reasoning_content` by default
+(`preserve_thinking`), so the model can build on its earlier thought process:
+
+```ts
+import { alibaba } from '@ai-sdk/alibaba';
+import { generateText } from 'ai';
+
+const providerOptions = {
+  alibaba: {
+    enableThinking: true,
+    thinkingBudget: 2048,
+  },
+};
+
+const opening = {
+  role: 'user' as const,
+  content: 'Is Kafka or RocketMQ a better fit for transactional messages?',
+};
+
+const first = await generateText({
+  model: alibaba('qwen3.7-max'),
+  messages: [opening],
+  providerOptions,
+});
+
+const second = await generateText({
+  model: alibaba('qwen3.7-max'),
+  messages: [
+    opening,
+    ...first.responseMessages, // append unchanged to keep the reasoning parts
+    { role: 'user', content: 'Which tradeoff mattered most?' },
+  ],
+  providerOptions,
+});
+```
+
+When continuing the conversation, append `responseMessages` unchanged so the
+reasoning parts survive to be serialized as `reasoning_content`. Set the
+`preserveThinking` provider option to `false` to opt out. Keep in mind:
+
+- `preserveThinking` does not enable thinking by itself.
+- It is enabled by default only for models that Alibaba documents as supporting
+  preserved thinking; for other models the option is not sent unless you set it
+  explicitly. See Alibaba's
+  [preserved-thinking documentation](https://docs.qwencloud.com/developer-guides/text-generation/thinking#preserve-thinking-in-multi-turn).
+- Reasoning from the current tool-call round is always sent back with tool
+  results, as Alibaba recommends.
+- Preserved reasoning increases input token usage and billing.
+- Historical reasoning remains separate from visible assistant text; it is never
+  merged into `content`.
+
 ## Embedding Model Example
 
 ```ts

package/dist/index.d.mts

@@ -6,6 +6,7 @@ type AlibabaChatModelId = 'qwen3.7-max' | 'qwen3-max' | 'qwen3-max-preview' | 'q
 declare const alibabaLanguageModelOptions: z.ZodObject<{
     enableThinking: z.ZodOptional<z.ZodBoolean>;
     thinkingBudget: z.ZodOptional<z.ZodNumber>;
+    preserveThinking: z.ZodOptional<z.ZodBoolean>;
     parallelToolCalls: z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>;
 type AlibabaLanguageModelOptions = z.infer<typeof alibabaLanguageModelOptions>;

package/dist/index.d.ts

@@ -6,6 +6,7 @@ type AlibabaChatModelId = 'qwen3.7-max' | 'qwen3-max' | 'qwen3-max-preview' | 'q
 declare const alibabaLanguageModelOptions: z.ZodObject<{
     enableThinking: z.ZodOptional<z.ZodBoolean>;
     thinkingBudget: z.ZodOptional<z.ZodNumber>;
+    preserveThinking: z.ZodOptional<z.ZodBoolean>;
     parallelToolCalls: z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>;
 type AlibabaLanguageModelOptions = z.infer<typeof alibabaLanguageModelOptions>;

package/dist/index.js

@@ -50,6 +50,15 @@ var alibabaLanguageModelOptions = import_v4.z.object({
    * Maximum number of reasoning tokens to generate.
    */
   thinkingBudget: import_v4.z.number().positive().optional(),
+  /**
+   * Whether to preserve reasoning from previous assistant messages.
+   *
+   * When enabled, historical reasoning is sent separately as
+   * `reasoning_content` and Alibaba receives `preserve_thinking: true`.
+   * Defaults to `true` for models that support preserved thinking; set to
+   * `false` to opt out.
+   */
+  preserveThinking: import_v4.z.boolean().optional(),
   /**
    * Whether to enable parallel function calling during tool use.
    *
@@ -101,11 +110,21 @@ function formatImageUrl({
 }
 function convertToAlibabaChatMessages({
   prompt,
-  cacheControlValidator
+  cacheControlValidator,
+  preserveThinking = false
 }) {
   var _a, _b;
   const messages = [];
+  let lastUserMessageIndex = -1;
+  for (let i = prompt.length - 1; i >= 0; i--) {
+    if (prompt[i].role === "user") {
+      lastUserMessageIndex = i;
+      break;
+    }
+  }
+  let index = -1;
   for (const { role, content, ...message } of prompt) {
+    index++;
     const messageCacheControl = cacheControlValidator == null ? void 0 : cacheControlValidator.getCacheControl(
       message.providerOptions
     );
@@ -130,9 +149,9 @@ function convertToAlibabaChatMessages({
       case "user": {
         messages.push({
           role: "user",
-          content: content.map((part, index) => {
+          content: content.map((part, index2) => {
             var _a2;
-            const isLastPart = index === content.length - 1;
+            const isLastPart = index2 === content.length - 1;
             const partCacheControl = (_a2 = cacheControlValidator == null ? void 0 : cacheControlValidator.getCacheControl(part.providerOptions)) != null ? _a2 : isLastPart ? messageCacheControl : void 0;
             switch (part.type) {
               case "text": {
@@ -165,6 +184,7 @@ function convertToAlibabaChatMessages({
       }
       case "assistant": {
         let text = "";
+        let reasoningContent = "";
         const toolCalls = [];
         for (const part of content) {
           switch (part.type) {
@@ -184,14 +204,20 @@ function convertToAlibabaChatMessages({
               break;
             }
             case "reasoning": {
-              text += part.text;
+              if (preserveThinking || index > lastUserMessageIndex) {
+                reasoningContent += part.text;
+              }
               break;
             }
           }
         }
+        if (text.length === 0 && toolCalls.length === 0 && reasoningContent.length === 0) {
+          break;
+        }
         messages.push({
           role: "assistant",
-          content: messageCacheControl ? [{ type: "text", text, cache_control: messageCacheControl }] : text || null,
+          content: text.length === 0 && toolCalls.length === 0 && reasoningContent.length > 0 ? null : messageCacheControl ? [{ type: "text", text, cache_control: messageCacheControl }] : text || null,
+          ...reasoningContent.length > 0 ? { reasoning_content: reasoningContent } : {},
           tool_calls: toolCalls.length > 0 ? toolCalls : void 0
         });
         break;
@@ -277,6 +303,29 @@ var CacheControlValidator = class {
   }
 };
 
+// src/supports-preserved-thinking.ts
+var preservedThinkingModelIds = /* @__PURE__ */ new Set([
+  "kimi-k2.7-code",
+  "qwen3.6-max-preview",
+  "qwen3.6-plus",
+  "qwen3.6-plus-2026-04-02",
+  "qwen3.7-flash",
+  "qwen3.7-flash-2026-07-15",
+  "qwen3.7-max",
+  "qwen3.7-max-2026-05-17",
+  "qwen3.7-max-2026-05-20",
+  "qwen3.7-max-2026-06-08",
+  "qwen3.7-max-preview",
+  "qwen3.7-plus",
+  "qwen3.7-plus-2026-05-26",
+  "qwen3.8-flash",
+  "qwen3.8-max",
+  "qwen3.8-max-0902"
+]);
+function supportsPreservedThinking(modelId) {
+  return preservedThinkingModelIds.has(modelId);
+}
+
 // src/alibaba-chat-language-model.ts
 var AlibabaLanguageModel = class {
   constructor(modelId, config) {
@@ -309,7 +358,7 @@ var AlibabaLanguageModel = class {
     tools,
     toolChoice
   }) {
-    var _a;
+    var _a, _b;
     const warnings = [];
     const cacheControlValidator = new CacheControlValidator();
     const alibabaOptions = await (0, import_provider_utils3.parseProviderOptions)({
@@ -320,6 +369,7 @@ var AlibabaLanguageModel = class {
     if (frequencyPenalty != null) {
       warnings.push({ type: "unsupported", feature: "frequencyPenalty" });
     }
+    const preserveThinking = (_a = alibabaOptions == null ? void 0 : alibabaOptions.preserveThinking) != null ? _a : supportsPreservedThinking(this.modelId) ? true : void 0;
     const baseArgs = {
       model: this.modelId,
       max_tokens: maxOutputTokens,
@@ -333,17 +383,19 @@ var AlibabaLanguageModel = class {
         type: "json_schema",
         json_schema: {
           schema: responseFormat.schema,
-          name: (_a = responseFormat.name) != null ? _a : "response",
+          name: (_b = responseFormat.name) != null ? _b : "response",
           description: responseFormat.description
         }
       } : { type: "json_object" } : void 0,
       // Alibaba-specific options
       ...(alibabaOptions == null ? void 0 : alibabaOptions.enableThinking) != null ? { enable_thinking: alibabaOptions.enableThinking } : {},
       ...(alibabaOptions == null ? void 0 : alibabaOptions.thinkingBudget) != null ? { thinking_budget: alibabaOptions.thinkingBudget } : {},
+      ...preserveThinking != null ? { preserve_thinking: preserveThinking } : {},
       // Convert messages with cache control support
       messages: convertToAlibabaChatMessages({
         prompt,
-        cacheControlValidator
+        cacheControlValidator,
+        preserveThinking: preserveThinking != null ? preserveThinking : false
       })
     };
     const {
@@ -513,7 +565,7 @@ var AlibabaLanguageModel = class {
                 delta: delta.content
               });
             }
-            if (delta.tool_calls != null) {
+            if (delta.tool_calls != null && delta.tool_calls.length > 0) {
               if (activeReasoningId != null) {
                 controller.enqueue({
                   type: "reasoning-end",
@@ -1312,7 +1364,7 @@ var AlibabaVideoModel = class {
 };
 
 // src/version.ts
-var VERSION = true ? "1.0.53" : "0.0.0-test";
+var VERSION = true ? "1.0.55" : "0.0.0-test";
 
 // src/alibaba-provider.ts
 function createAlibaba(options = {}) {