@zhivex-ai/qwen 0.10.2 → 0.11.0

package/dist/index.js

@@ -1,59 +1,7 @@
 import { toJSONSchema } from "zod";
+import { capabilities, embeddingCapabilities, imageGenerationCapabilities, realtimeCapabilities, speechCapabilities, transcriptionCapabilities, videoGenerationCapabilities, } from "./capabilities.js";
 import { CallbackRealtimeSession, ConfigurationError, decodeBase64WithLimit, ProviderHTTPError, ValidationError, assertTrustedEndpoint, readBodyWithLimit, readErrorBodyWithLimit, readJsonWithLimit, resolveAudioResponseLimits, UnsupportedFeatureError, createProviderAdapter, hostedTool, openWebSocketConnection, isCallableToolDefinition, normalizeFinishReason, providerDataPart, streamSSE, toToolSet, withRetry, withTimeoutSignal } from "@zhivex-ai/core";
 export const QWEN_TOKEN_PLAN_BASE_URL = "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1";
-const capabilities = {
-    streaming: true,
-    tools: true,
-    structuredOutput: true,
-    jsonMode: true,
-    toolChoice: true,
-    parallelToolCalls: false,
-    vision: true,
-    files: false,
-    audioInput: false,
-    audioOutput: false,
-    embeddings: true,
-    reasoning: false,
-    webSearch: true,
-    agentCapabilities: {
-        supportTier: "tier-b",
-        toolChoiceNone: true,
-        approvalRequests: false,
-        hostedWebSearch: true,
-        hostedFileSearch: true,
-        remoteMcp: true,
-        computerUse: false,
-        codeExecution: true,
-        webExtraction: true,
-        toolsets: false
-    }
-};
-const embeddingCapabilities = {
-    streaming: false,
-    tools: false,
-    structuredOutput: false,
-    jsonMode: false,
-    toolChoice: false,
-    parallelToolCalls: false,
-    vision: false,
-    files: false,
-    audioInput: false,
-    audioOutput: false,
-    embeddings: true,
-    reasoning: false,
-    webSearch: false,
-    agentCapabilities: {
-        supportTier: "tier-c",
-        toolChoiceNone: false,
-        approvalRequests: false,
-        hostedWebSearch: false,
-        hostedFileSearch: false,
-        remoteMcp: false,
-        computerUse: false,
-        codeExecution: false,
-        toolsets: false
-    }
-};
 const qwenTaskBaseURLFrom = (baseURL) => {
     const url = new URL(baseURL);
     return `${url.protocol}//${url.host}/api/v1`;
@@ -375,8 +323,10 @@ const stripResponsesRequestOptions = (providerOptions) => {
 };
 const modelFamily = (modelId) => modelId.toLowerCase();
 const isQwen38Max = (modelId) => modelFamily(modelId) === "qwen3.8-max";
+const isQwen38Flash = (modelId) => modelFamily(modelId) === "qwen3.8-flash";
 const isQwen38MaxPreview = (modelId) => modelFamily(modelId) === "qwen3.8-max-preview";
-const isQwen38MaxFamily = (modelId) => isQwen38Max(modelId) || isQwen38MaxPreview(modelId);
+const isQwen38ProductionModel = (modelId) => isQwen38Max(modelId) || isQwen38Flash(modelId);
+const isQwen38ReasoningModel = (modelId) => isQwen38ProductionModel(modelId) || isQwen38MaxPreview(modelId);
 const isQwenTokenPlanURL = (url) => {
     const path = url.pathname.replace(/\/+$/, "");
     const tokenPlanURL = new URL(QWEN_TOKEN_PLAN_BASE_URL);
@@ -399,7 +349,7 @@ const validateQwen38BaseURL = (baseURL) => {
         throw new ConfigurationError(`Qwen qwen3.8-max-preview is Token Plan only. Configure baseURL with QWEN_TOKEN_PLAN_BASE_URL (${QWEN_TOKEN_PLAN_BASE_URL}); pay-as-you-go and workspace endpoints cannot serve this model.`);
     }
 };
-const validateQwen38MaxBaseURL = (baseURL) => {
+const validateQwen38ProductionBaseURL = (modelId, baseURL) => {
     let url;
     try {
         url = new URL(baseURL);
@@ -408,7 +358,7 @@ const validateQwen38MaxBaseURL = (baseURL) => {
         return;
     }
     if (isQwenTokenPlanURL(url)) {
-        throw new ConfigurationError("Qwen qwen3.8-max uses standard Model Studio credentials and endpoints; QWEN_TOKEN_PLAN_BASE_URL is reserved for qwen3.8-max-preview.");
+        throw new ConfigurationError(`Qwen ${modelId} uses standard Model Studio credentials and endpoints; QWEN_TOKEN_PLAN_BASE_URL is reserved for qwen3.8-max-preview.`);
     }
 };
 const stripQwen38ReasoningOptions = (providerOptions) => {
@@ -428,7 +378,7 @@ const supportsQwenVision = (modelId) => {
         model.includes("omni") ||
         model.includes("ocr") ||
         model.includes("vision") ||
-        isQwen38Max(modelId) ||
+        isQwen38ProductionModel(modelId) ||
         isQwen38MaxPreview(modelId) ||
         /^qwen3\.7-plus(?:$|-)/.test(model) ||
         /^qwen3\.6-flash(?:$|-)/.test(model));
@@ -440,7 +390,7 @@ const qwenLanguageCapabilities = (modelId) => {
     const tools = supportsQwenTools(modelId);
     const omni = isQwenOmniLanguageModel(modelId);
     const reasoning = supportsQwenReasoning(modelId);
-    const qwen38Max = isQwen38Max(modelId);
+    const qwen38Production = isQwen38ProductionModel(modelId);
     const qwen38MaxPreview = isQwen38MaxPreview(modelId);
     return {
         ...capabilities,
@@ -449,12 +399,12 @@ const qwenLanguageCapabilities = (modelId) => {
         structuredOutput: tools && !omni && !qwen38MaxPreview,
         jsonMode: tools && !omni && !qwen38MaxPreview,
         toolChoice: tools,
-        parallelToolCalls: qwen38Max,
+        parallelToolCalls: qwen38Production,
         webSearch: tools && !omni,
-        files: supportsQwenFiles(modelId) || qwen38Max,
+        files: supportsQwenFiles(modelId) || qwen38Production,
         audioInput: supportsQwenAudioInput(modelId),
         reasoning,
-        reasoningEfforts: qwen38Max
+        reasoningEfforts: qwen38Production
             ? ["none", "minimal", "low", "medium", "high", "xhigh", "max"]
             : qwen38MaxPreview
                 ? ["minimal", "low", "medium", "high", "xhigh", "max"]
@@ -473,53 +423,6 @@ const qwenLanguageCapabilities = (modelId) => {
         }
     };
 };
-const transcriptionCapabilities = {
-    ...embeddingCapabilities,
-    audioInput: true
-};
-const speechCapabilities = {
-    ...embeddingCapabilities,
-    audioOutput: true
-};
-const imageGenerationCapabilities = {
-    ...embeddingCapabilities,
-    imageGeneration: true
-};
-const videoGenerationCapabilities = {
-    ...embeddingCapabilities,
-    videoGeneration: true
-};
-const realtimeCapabilities = {
-    ...capabilities,
-    streaming: false,
-    structuredOutput: false,
-    jsonMode: false,
-    toolChoice: false,
-    parallelToolCalls: false,
-    files: false,
-    audioInput: true,
-    audioOutput: true,
-    embeddings: false,
-    reasoning: false,
-    webSearch: false,
-    agentCapabilities: {
-        ...capabilities.agentCapabilities,
-        toolChoiceNone: true,
-        hostedWebSearch: false,
-        hostedFileSearch: false,
-        remoteMcp: false,
-        codeExecution: false,
-        webExtraction: false
-    },
-    realtime: {
-        sessions: true,
-        audioInput: true,
-        audioOutput: true,
-        imageInput: true,
-        tools: true,
-        browserTokens: false
-    }
-};
 const reasoningContentFromMessage = (message) => message.parts
     .filter((part) => {
     if (part.type !== "provider-data" || part.provider !== "qwen") {
@@ -683,16 +586,29 @@ const mapResponsesToolChoice = (toolChoice, tools) => {
         tools: [{ type: "function", name: toolChoice.toolName }]
     };
 };
-const mapStructuredOutput = (input) => {
+const mapStructuredOutput = (modelId, input) => {
     if (!input.structuredOutput || input.structuredOutput.mode !== "native") {
         return undefined;
     }
+    if (isQwen38Flash(modelId)) {
+        return {
+            type: "json_schema",
+            json_schema: {
+                name: input.structuredOutput.name ?? "response",
+                strict: true,
+                schema: toJSONSchema(input.structuredOutput.schema)
+            }
+        };
+    }
     return { type: "json_object" };
 };
-const withStructuredOutputMessages = (input) => {
+const withStructuredOutputMessages = (modelId, input) => {
     if (!input.structuredOutput || input.structuredOutput.mode !== "native") {
         return input.messages;
     }
+    if (isQwen38Flash(modelId)) {
+        return input.messages;
+    }
     const schema = JSON.stringify(toJSONSchema(input.structuredOutput.schema));
     return [
         {
@@ -766,26 +682,28 @@ const validateQwen38Request = (modelId, input, providerOptions) => {
     if (sharedBudget !== undefined && sharedBudget > 262_144) {
         throw new ConfigurationError(`Qwen ${modelId} reasoning.budgetTokens cannot exceed 262144.`);
     }
-    if (isQwen38Max(modelId) &&
+    if (isQwen38ProductionModel(modelId) &&
         providerOptions.enable_thinking === false &&
         ((effort !== undefined && effort !== "none") || budget !== undefined)) {
-        throw new ConfigurationError("Qwen qwen3.8-max cannot disable thinking while also configuring a reasoning effort or thinking budget.");
+        throw new ConfigurationError(`Qwen ${modelId} cannot disable thinking while also configuring a reasoning effort or thinking budget.`);
     }
-    if (isQwen38Max(modelId) && providerOptions.enable_thinking === true && effort === "none") {
-        throw new ConfigurationError('Qwen qwen3.8-max cannot enable thinking with reasoning effort "none".');
+    if (isQwen38ProductionModel(modelId) &&
+        providerOptions.enable_thinking === true &&
+        effort === "none") {
+        throw new ConfigurationError(`Qwen ${modelId} cannot enable thinking with reasoning effort "none".`);
     }
     const thinkingEnabled = isQwen38MaxPreview(modelId) ||
         !(effort === "none" || providerOptions.enable_thinking === false);
     if (thinkingEnabled && (input.toolChoice === "required" || typeof input.toolChoice === "object")) {
         throw new UnsupportedFeatureError(`Qwen ${modelId} supports toolChoice "required" or a named tool only when thinking is disabled.`);
     }
-    if (isQwen38Max(modelId) &&
+    if (isQwen38ProductionModel(modelId) &&
         input.messages.some((message) => message.parts.some((part) => part.type === "file" && !part.mediaType.toLowerCase().startsWith("video/")))) {
-        throw new UnsupportedFeatureError("Qwen qwen3.8-max FilePart input is reserved for video/* media. Use ImagePart for images.");
+        throw new UnsupportedFeatureError(`Qwen ${modelId} FilePart input is reserved for video/* media. Use ImagePart for images.`);
     }
 };
 const mapChatReasoning = (modelId, input, providerOptions) => {
-    if (isQwen38MaxFamily(modelId)) {
+    if (isQwen38ReasoningModel(modelId)) {
         const rawEffort = input.reasoning?.effort ?? providerOptions.reasoning_effort;
         const thinkingDisabled = rawEffort === "none" || providerOptions.enable_thinking === false;
         if (thinkingDisabled) {
@@ -818,7 +736,9 @@ const mapChatReasoning = (modelId, input, providerOptions) => {
 };
 const mapResponsesReasoning = (modelId, input, providerOptions) => {
     const effort = input.reasoning?.effort ?? providerOptions.reasoning_effort;
-    const qwen38Effort = isQwen38Max(modelId) && providerOptions.enable_thinking === false && effort === undefined
+    const qwen38Effort = isQwen38ProductionModel(modelId) &&
+        providerOptions.enable_thinking === false &&
+        effort === undefined
         ? "none"
         : effort;
     if (!qwen38Effort) {
@@ -828,7 +748,7 @@ const mapResponsesReasoning = (modelId, input, providerOptions) => {
         reasoning: {
             effort: isQwen38MaxPreview(modelId)
                 ? mapQwen38ChatReasoningEffort(modelId, qwen38Effort)
-                : isQwen38Max(modelId)
+                : isQwen38ProductionModel(modelId)
                     ? qwen38Effort
                     : qwen38Effort === "low"
                         ? "minimal"
@@ -860,7 +780,7 @@ const hasVideoInput = (messages) => messages.some((message) => message.parts.som
 const hasNonVideoFileInput = (messages) => messages.some((message) => message.parts.some((part) => part.type === "file" && !part.mediaType.toLowerCase().startsWith("video/")));
 const resolveApiMode = (modelId, requestedMode, input, providerOptions) => {
     const videoInput = hasVideoInput(input.messages);
-    if (videoInput && !isQwen38Max(modelId)) {
+    if (videoInput && !isQwen38ProductionModel(modelId)) {
         throw new UnsupportedFeatureError(`Qwen model "${modelId}" does not support video FilePart input through this adapter.`);
     }
     const needsChat = input.maxTokens !== undefined ||
@@ -891,25 +811,87 @@ const resolveApiMode = (modelId, requestedMode, input, providerOptions) => {
     }
     return needsChat ? "chat" : "responses";
 };
-const parseAssistantMessage = (message) => ({
-    role: "assistant",
-    parts: [
-        ...(typeof message.reasoning_content === "string" && message.reasoning_content
-            ? [providerDataPart("qwen", { type: "reasoning_content", reasoningContent: message.reasoning_content })]
-            : []),
-        ...(typeof message.content === "string" && message.content
-            ? [{ type: "text", text: message.content }]
-            : []),
-        ...((message.tool_calls ?? []).map((call) => ({
-            type: "tool-call",
-            toolCall: {
-                id: call.id,
-                name: call.function.name,
-                input: JSON.parse(call.function.arguments ?? "{}")
-            }
-        })) ?? [])
-    ]
-});
+const classifyToolCallId = (value) => {
+    if (typeof value === "number" && Number.isFinite(value)) {
+        return { kind: "transient" };
+    }
+    if (typeof value !== "string") {
+        return { kind: "missing" };
+    }
+    const trimmed = value.trim();
+    if (!trimmed) {
+        return { kind: "missing" };
+    }
+    if (/^\d+$/.test(trimmed)) {
+        return { kind: "transient" };
+    }
+    return { kind: "stable", id: value };
+};
+const stableToolCallId = (value) => {
+    const classification = classifyToolCallId(value);
+    return classification.kind === "stable" ? classification.id : undefined;
+};
+const qwenResponsesProviderCallId = (value) => typeof value === "string" && value.trim().length > 0 ? value : undefined;
+const existingToolCallIds = (messages) => new Set(messages.flatMap((message) => message.parts.flatMap((part) => part.type === "tool-call" ? [part.toolCall.id] : [])));
+const nextFallbackToolCallGeneration = (surface, input, seenIds) => {
+    const prefix = `qwen-${surface}-tool-`;
+    let generation = input.messages.length;
+    for (const id of seenIds) {
+        if (!id.startsWith(prefix))
+            continue;
+        const separator = id.indexOf("-", prefix.length);
+        if (separator < 0)
+            continue;
+        const generationText = id.slice(prefix.length, separator);
+        if (!generationText)
+            continue;
+        const retainedGeneration = Number(generationText);
+        if (Number.isSafeInteger(retainedGeneration) && retainedGeneration >= generation) {
+            generation = retainedGeneration + 1;
+        }
+    }
+    return generation;
+};
+const fallbackToolCallId = (surface, generation, index) => `qwen-${surface}-tool-${generation}-${index}`;
+class QwenToolCallIdError extends ValidationError {
+    diagnosticCode = "QWEN_DUPLICATE_TOOL_CALL_ID";
+    constructor() {
+        super("Qwen returned a duplicate tool-call id.");
+        this.name = "QwenToolCallIdError";
+    }
+}
+const resolveToolCallId = (candidates, fallback, seenIds) => {
+    const candidate = candidates.map(stableToolCallId).find((value) => value !== undefined);
+    const id = candidate ?? (typeof fallback === "function" ? fallback() : fallback);
+    if (seenIds.has(id)) {
+        throw new QwenToolCallIdError();
+    }
+    seenIds.add(id);
+    return id;
+};
+const parseAssistantMessage = (message, input) => {
+    const seenIds = existingToolCallIds(input.messages);
+    const fallbackGeneration = nextFallbackToolCallGeneration("chat", input, seenIds);
+    return {
+        role: "assistant",
+        parts: [
+            ...(typeof message.reasoning_content === "string" && message.reasoning_content
+                ? [providerDataPart("qwen", { type: "reasoning_content", reasoningContent: message.reasoning_content })]
+                : []),
+            ...(typeof message.content === "string" && message.content
+                ? [{ type: "text", text: message.content }]
+                : []),
+            ...((message.tool_calls ?? []).map((call, index) => ({
+                type: "tool-call",
+                toolCall: {
+                    id: resolveToolCallId([call.id], fallbackToolCallId("chat", fallbackGeneration, index), seenIds),
+                    name: call.function.name,
+                    input: JSON.parse(call.function.arguments ?? "{}")
+                }
+            })) ?? [])
+        ]
+    };
+};
 const getProviderResponseId = (messages) => {
     for (let index = messages.length - 1; index >= 0; index -= 1) {
         const message = messages[index];
@@ -932,11 +914,16 @@ const getProviderResponseId = (messages) => {
 };
 const serializeResponsesToolOutput = (message) => message.parts
     .filter((part) => part.type === "tool-result")
-    .map((part) => ({
-    type: "function_call_output",
-    call_id: part.toolResult.toolCallId,
-    output: JSON.stringify(part.toolResult.isError ? part.toolResult.error : part.toolResult.output ?? null)
-}));
+    .map((part) => {
+    const providerCallId = part.toolResult.providerMetadata?.qwenResponsesCallId;
+    return {
+        type: "function_call_output",
+        call_id: typeof providerCallId === "string" && providerCallId.trim().length > 0
+            ? providerCallId
+            : part.toolResult.toolCallId,
+        output: JSON.stringify(part.toolResult.isError ? part.toolResult.error : part.toolResult.output ?? null)
+    };
+});
 const toResponsesInput = (messages) => {
     const input = [];
     for (const message of messages) {
@@ -1003,8 +990,10 @@ const parseResponsesProviderData = (item) => {
     }
     return item;
 };
-const parseResponsesAssistantMessage = (json) => {
+const parseResponsesAssistantMessage = (json, input) => {
     const parts = [];
+    const seenIds = existingToolCallIds(input.messages);
+    const fallbackGeneration = nextFallbackToolCallGeneration("responses", input, seenIds);
     for (const [index, item] of (json.output ?? []).entries()) {
         if (item?.type === "message") {
             for (const content of item.content ?? []) {
@@ -1015,12 +1004,17 @@ const parseResponsesAssistantMessage = (json) => {
             continue;
         }
         if (item?.type === "function_call") {
+            const providerCallId = qwenResponsesProviderCallId(item.call_id);
+            const id = resolveToolCallId([item.call_id, item.id], fallbackToolCallId("responses", fallbackGeneration, index), seenIds);
             parts.push({
                 type: "tool-call",
                 toolCall: {
-                    id: item.call_id ?? item.id ?? `${item.name}-${index}`,
+                    id,
                     name: item.name,
-                    input: JSON.parse(item.arguments ?? "{}")
+                    input: JSON.parse(item.arguments ?? "{}"),
+                    ...(providerCallId !== undefined && providerCallId !== id
+                        ? { providerMetadata: { qwenResponsesCallId: providerCallId } }
+                        : {})
                 }
             });
             continue;
@@ -1053,9 +1047,30 @@ const normalizeResponsesFinishReason = (status, hasToolCalls) => {
     }
     return normalizeFinishReason(status);
 };
-const streamResponses = async function* (response) {
+const streamResponses = async function* (response, input) {
     const toolBuffers = new Map();
+    const toolBufferAliases = new Map();
+    const seenIds = existingToolCallIds(input.messages);
+    const fallbackGeneration = nextFallbackToolCallGeneration("responses", input, seenIds);
     let sawToolCalls = false;
+    const bufferKey = (outputIndex, candidates, fallbackId) => {
+        const stableIds = candidates
+            .map(stableToolCallId)
+            .filter((value) => value !== undefined);
+        if ((typeof outputIndex === "number" && Number.isSafeInteger(outputIndex) && outputIndex >= 0) ||
+            (typeof outputIndex === "string" && outputIndex.trim().length > 0)) {
+            const key = `output:${outputIndex}`;
+            for (const stableId of stableIds) {
+                toolBufferAliases.set(`id:${stableId}`, key);
+            }
+            return key;
+        }
+        for (const stableId of stableIds) {
+            const identityKey = `id:${stableId}`;
+            return toolBufferAliases.get(identityKey) ?? identityKey;
+        }
+        return `fallback:${fallbackId}`;
+    };
     const emitToolCall = (key) => {
         const toolCall = toolBuffers.get(key);
         if (!toolCall || toolCall.emitted || !toolCall.name) {
@@ -1063,12 +1078,16 @@ const streamResponses = async function* (response) {
         }
         toolCall.emitted = true;
         sawToolCalls = true;
+        const id = resolveToolCallId([toolCall.callId], toolCall.fallbackId, seenIds);
         return {
             type: "tool-call",
             toolCall: {
-                id: toolCall.callId,
+                id,
                 name: toolCall.name,
-                input: JSON.parse(toolCall.args || "{}")
+                input: JSON.parse(toolCall.args || "{}"),
+                ...(toolCall.providerCallId !== undefined && toolCall.providerCallId !== id
+                    ? { providerMetadata: { qwenResponsesCallId: toolCall.providerCallId } }
+                    : {})
             }
         };
     };
@@ -1097,14 +1116,19 @@ const streamResponses = async function* (response) {
         if (type === "response.output_item.added" || type === "response.output_item.done") {
             const item = json.item;
             if (item?.type === "function_call") {
-                const key = item.id ?? json.item_id ?? `${json.output_index ?? toolBuffers.size}`;
+                const outputIndex = json.output_index ?? toolBuffers.size;
+                const fallbackId = fallbackToolCallId("responses", fallbackGeneration, outputIndex);
+                const key = bufferKey(json.output_index, [item.id, json.item_id, item.call_id], fallbackId);
                 const existing = toolBuffers.get(key) ?? {
-                    callId: item.call_id ?? key,
+                    callId: stableToolCallId(item.call_id) ?? stableToolCallId(item.id) ?? stableToolCallId(json.item_id),
+                    providerCallId: qwenResponsesProviderCallId(item.call_id),
+                    fallbackId,
                     name: item.name ?? "",
                     args: "",
                     emitted: false
                 };
-                existing.callId = item.call_id ?? existing.callId;
+                existing.callId = stableToolCallId(item.call_id) ?? existing.callId;
+                existing.providerCallId = qwenResponsesProviderCallId(item.call_id) ?? existing.providerCallId;
                 existing.name ||= item.name ?? "";
                 if (typeof item.arguments === "string") {
                     existing.args = item.arguments;
@@ -1128,25 +1152,35 @@ const streamResponses = async function* (response) {
             continue;
         }
         if (type === "response.function_call_arguments.delta") {
-            const key = json.item_id ?? `${json.output_index ?? toolBuffers.size}`;
+            const outputIndex = json.output_index ?? toolBuffers.size;
+            const fallbackId = fallbackToolCallId("responses", fallbackGeneration, outputIndex);
+            const key = bufferKey(json.output_index, [json.item_id, json.call_id], fallbackId);
             const existing = toolBuffers.get(key) ?? {
-                callId: key,
+                callId: stableToolCallId(json.call_id) ?? stableToolCallId(json.item_id),
+                providerCallId: qwenResponsesProviderCallId(json.call_id),
+                fallbackId,
                 name: "",
                 args: "",
                 emitted: false
             };
+            existing.providerCallId = qwenResponsesProviderCallId(json.call_id) ?? existing.providerCallId;
             existing.args += typeof json.delta === "string" ? json.delta : "";
             toolBuffers.set(key, existing);
             continue;
         }
         if (type === "response.function_call_arguments.done") {
-            const key = json.item_id ?? `${json.output_index ?? toolBuffers.size}`;
+            const outputIndex = json.output_index ?? toolBuffers.size;
+            const fallbackId = fallbackToolCallId("responses", fallbackGeneration, outputIndex);
+            const key = bufferKey(json.output_index, [json.item_id, json.call_id], fallbackId);
             const existing = toolBuffers.get(key) ?? {
-                callId: key,
+                callId: stableToolCallId(json.call_id) ?? stableToolCallId(json.item_id),
+                providerCallId: qwenResponsesProviderCallId(json.call_id),
+                fallbackId,
                 name: "",
                 args: "",
                 emitted: false
             };
+            existing.providerCallId = qwenResponsesProviderCallId(json.call_id) ?? existing.providerCallId;
             if (typeof json.arguments === "string") {
                 existing.args = json.arguments;
             }
@@ -1195,11 +1229,11 @@ class QwenLanguageModel {
         }
         const providerOptions = { ...(input.providerOptions ?? {}) };
         validateQwenToolStream(this.modelId, providerOptions, false);
-        if (isQwen38MaxFamily(this.modelId)) {
+        if (isQwen38ReasoningModel(this.modelId)) {
             validateQwen38Request(this.modelId, input, providerOptions);
         }
-        if (isQwen38Max(this.modelId)) {
-            validateQwen38MaxBaseURL(this.baseURL);
+        if (isQwen38ProductionModel(this.modelId)) {
+            validateQwen38ProductionBaseURL(this.modelId, this.baseURL);
         }
         if (isQwen38MaxPreview(this.modelId)) {
             validateQwen38BaseURL(this.baseURL);
@@ -1211,7 +1245,7 @@ class QwenLanguageModel {
             if (apiMode === "responses") {
                 const previousResponse = getProviderResponseId(input.messages);
                 const baseResponseProviderOptions = stripResponsesRequestOptions(providerOptions);
-                const responseProviderOptions = isQwen38MaxFamily(this.modelId)
+                const responseProviderOptions = isQwen38ReasoningModel(this.modelId)
                     ? stripQwen38ReasoningOptions(baseResponseProviderOptions)
                     : baseResponseProviderOptions;
                 const messages = previousResponse && previousResponse.index < input.messages.length - 1
@@ -1234,7 +1268,7 @@ class QwenLanguageModel {
                     })
                 }), input);
                 const json = await parseJson(response);
-                const assistantMessage = parseResponsesAssistantMessage(json);
+                const assistantMessage = parseResponsesAssistantMessage(json, input);
                 const hasToolCalls = assistantMessage.parts.some((part) => part.type === "tool-call");
                 return {
                     messages: [assistantMessage],
@@ -1249,7 +1283,7 @@ class QwenLanguageModel {
                 };
             }
             const baseChatProviderOptions = stripHeaderOptions(providerOptions);
-            const chatProviderOptions = isQwen38MaxFamily(this.modelId)
+            const chatProviderOptions = isQwen38ReasoningModel(this.modelId)
                 ? stripQwen38ReasoningOptions(baseChatProviderOptions)
                 : baseChatProviderOptions;
             const response = await withRetry(() => this.fetcher(`${this.baseURL}/chat/completions`, {
@@ -1259,13 +1293,13 @@ class QwenLanguageModel {
                 body: JSON.stringify({
                     ...chatProviderOptions,
                     model: this.modelId,
-                    messages: mapMessages(withStructuredOutputMessages(input)),
+                    messages: mapMessages(withStructuredOutputMessages(this.modelId, input)),
                     tools: mapChatTools(input.tools),
                     tool_choice: mapChatToolChoice(input.toolChoice),
-                    response_format: mapStructuredOutput(input),
+                    response_format: mapStructuredOutput(this.modelId, input),
                     temperature: input.temperature,
-                    max_tokens: isQwen38MaxFamily(this.modelId) ? undefined : input.maxTokens,
-                    max_completion_tokens: isQwen38MaxFamily(this.modelId) ? input.maxTokens : undefined,
+                    max_tokens: isQwen38ReasoningModel(this.modelId) ? undefined : input.maxTokens,
+                    max_completion_tokens: isQwen38ReasoningModel(this.modelId) ? input.maxTokens : undefined,
                     stream: false,
                     ...mapChatReasoning(this.modelId, input, providerOptions)
                 })
@@ -1273,7 +1307,7 @@ class QwenLanguageModel {
             const json = await parseJson(response);
             const choice = json.choices?.[0];
             const message = choice?.message ?? {};
-            const assistantMessage = parseAssistantMessage(message);
+            const assistantMessage = parseAssistantMessage(message, input);
             return {
                 messages: [assistantMessage],
                 text: assistantMessage.parts
@@ -1293,11 +1327,11 @@ class QwenLanguageModel {
     async stream(input) {
         const providerOptions = { ...(input.providerOptions ?? {}) };
         validateQwenToolStream(this.modelId, providerOptions, true);
-        if (isQwen38MaxFamily(this.modelId)) {
+        if (isQwen38ReasoningModel(this.modelId)) {
             validateQwen38Request(this.modelId, input, providerOptions);
         }
-        if (isQwen38Max(this.modelId)) {
-            validateQwen38MaxBaseURL(this.baseURL);
+        if (isQwen38ProductionModel(this.modelId)) {
+            validateQwen38ProductionBaseURL(this.modelId, this.baseURL);
         }
         if (isQwen38MaxPreview(this.modelId)) {
             validateQwen38BaseURL(this.baseURL);
@@ -1308,7 +1342,7 @@ class QwenLanguageModel {
         if (apiMode === "responses") {
             const previousResponse = getProviderResponseId(input.messages);
             const baseResponseProviderOptions = stripResponsesRequestOptions(providerOptions);
-            const responseProviderOptions = isQwen38MaxFamily(this.modelId)
+            const responseProviderOptions = isQwen38ReasoningModel(this.modelId)
                 ? stripQwen38ReasoningOptions(baseResponseProviderOptions)
                 : baseResponseProviderOptions;
             const messages = previousResponse && previousResponse.index < input.messages.length - 1
@@ -1332,7 +1366,7 @@ class QwenLanguageModel {
             }), input);
             return (async function* () {
                 try {
-                    yield* streamResponses(response);
+                    yield* streamResponses(response, input);
                 }
                 finally {
                     cleanup();
@@ -1340,7 +1374,7 @@ class QwenLanguageModel {
             })();
         }
         const baseChatProviderOptions = stripHeaderOptions(providerOptions);
-        const chatProviderOptions = isQwen38MaxFamily(this.modelId)
+        const chatProviderOptions = isQwen38ReasoningModel(this.modelId)
             ? stripQwen38ReasoningOptions(baseChatProviderOptions)
             : baseChatProviderOptions;
         const response = await withRetry(() => this.fetcher(`${this.baseURL}/chat/completions`, {
@@ -1351,13 +1385,13 @@ class QwenLanguageModel {
                 ...chatProviderOptions,
                 model: this.modelId,
                 modalities: isQwenOmniLanguageModel(this.modelId) ? providerOptions.modalities ?? ["text"] : undefined,
-                messages: mapMessages(withStructuredOutputMessages(input)),
+                messages: mapMessages(withStructuredOutputMessages(this.modelId, input)),
                 tools: mapChatTools(input.tools),
                 tool_choice: mapChatToolChoice(input.toolChoice),
-                response_format: mapStructuredOutput(input),
+                response_format: mapStructuredOutput(this.modelId, input),
                 temperature: input.temperature,
-                max_tokens: isQwen38MaxFamily(this.modelId) ? undefined : input.maxTokens,
-                max_completion_tokens: isQwen38MaxFamily(this.modelId) ? input.maxTokens : undefined,
+                max_tokens: isQwen38ReasoningModel(this.modelId) ? undefined : input.maxTokens,
+                max_completion_tokens: isQwen38ReasoningModel(this.modelId) ? input.maxTokens : undefined,
                 stream: true,
                 stream_options: { include_usage: true },
                 ...mapChatReasoning(this.modelId, input, providerOptions)
@@ -1366,6 +1400,8 @@ class QwenLanguageModel {
         return (async function* () {
             try {
                 const toolBuffers = new Map();
+                const seenIds = existingToolCallIds(input.messages);
+                const fallbackGeneration = nextFallbackToolCallGeneration("chat", input, seenIds);
                 for await (const event of streamSSE(response)) {
                     if (event.data === "[DONE]") {
                         return;
@@ -1389,12 +1425,13 @@ class QwenLanguageModel {
                     for (const toolCall of delta?.tool_calls ?? []) {
                         const index = Number(toolCall.index ?? 0);
                         const existing = toolBuffers.get(index) ?? {
-                            id: toolCall.id ?? `${index}`,
+                            id: stableToolCallId(toolCall.id),
+                            fallbackId: fallbackToolCallId("chat", fallbackGeneration, index),
                             name: toolCall.function?.name ?? "",
                             args: "",
                             emitted: false
                         };
-                        existing.id = toolCall.id ?? existing.id;
+                        existing.id = stableToolCallId(toolCall.id) ?? existing.id;
                         existing.name ||= toolCall.function?.name ?? "";
                         existing.args += toolCall.function?.arguments ?? "";
                         toolBuffers.set(index, existing);
@@ -1409,7 +1446,7 @@ class QwenLanguageModel {
                                 yield {
                                     type: "tool-call",
                                     toolCall: {
-                                        id: toolCall.id,
+                                        id: resolveToolCallId([toolCall.id], toolCall.fallbackId, seenIds),
                                         name: toolCall.name,
                                         input: JSON.parse(toolCall.args || "{}")
                                     }
@@ -2155,7 +2192,7 @@ class QwenTasksClientImpl {
         }
     }
 }
-const parseRealtimeEvent = (payload) => {
+const parseRealtimeEvent = (payload, fallbackRealtimeToolCallId, seenToolCallIds) => {
     const type = String(payload.type ?? "");
     const metadata = payload;
     const itemId = typeof payload.item_id === "string" ? payload.item_id : undefined;
@@ -2196,7 +2233,7 @@ const parseRealtimeEvent = (payload) => {
         return [{
                 type: "realtime-tool-call",
                 toolCall: {
-                    id: String(payload.call_id ?? payload.item_id ?? "qwen-tool-call"),
+                    id: resolveToolCallId([payload.call_id, payload.item_id], fallbackRealtimeToolCallId, seenToolCallIds),
                     name: String(payload.name ?? ""),
                     input: JSON.parse(typeof payload.arguments === "string" ? payload.arguments : "{}")
                 }
@@ -2290,6 +2327,8 @@ class QwenRealtimeModel {
         assertQwenRealtimeConfig(config);
         const url = appendQuery(this.realtimeURL, { model: this.modelId });
         const connection = await (this.connectionFactory ?? openQwenRealtimeConnection)(url, { authorization: `Bearer ${this.apiKey}` }, options);
+        let fallbackRealtimeToolCallIndex = 0;
+        const seenRealtimeToolCallIds = new Set();
         const session = new CallbackRealtimeSession({
             provider: this.provider,
             modelId: this.modelId,
@@ -2297,7 +2336,7 @@ class QwenRealtimeModel {
             config,
             connection,
             callbacks: {
-                parseEvent: parseRealtimeEvent,
+                parseEvent: (payload) => parseRealtimeEvent(payload, () => `qwen-realtime-tool-${fallbackRealtimeToolCallIndex++}`, seenRealtimeToolCallIds),
                 buildInitialPayloads: (value) => [{ type: "session.update", session: mapRealtimeSession(value) }],
                 buildAudioPayloads: (frame, sessionConfig) => [
                     { type: "input_audio_buffer.append", audio: encodeRealtimeData(frame.data) },