@ai-sdk/moonshotai 2.0.44 → 2.0.46

package/src/moonshotai-chat-language-model.ts

@@ -1,60 +1,543 @@
-import { OpenAICompatibleChatLanguageModel } from '@ai-sdk/openai-compatible';
-import type { OpenAICompatibleChatConfig } from '@ai-sdk/openai-compatible/internal';
-import type {
-  LanguageModelV3CallOptions,
-  LanguageModelV3GenerateResult,
-  LanguageModelV3StreamPart,
-  LanguageModelV3StreamResult,
+import {
+  InvalidResponseDataError,
+  type APICallError,
+  type LanguageModelV3,
+  type LanguageModelV3CallOptions,
+  type LanguageModelV3Content,
+  type LanguageModelV3FinishReason,
+  type LanguageModelV3GenerateResult,
+  type LanguageModelV3StreamPart,
+  type LanguageModelV3StreamResult,
+  type SharedV3Warning,
 } from '@ai-sdk/provider';
+import {
+  combineHeaders,
+  createEventSourceResponseHandler,
+  createJsonErrorResponseHandler,
+  createJsonResponseHandler,
+  generateId,
+  parseProviderOptions,
+  postJsonToApi,
+  type FetchFunction,
+  type InferSchema,
+  type ParseResult,
+  type ResponseHandler,
+} from '@ai-sdk/provider-utils';
+import { convertToMoonshotAIChatMessages } from './convert-to-moonshotai-chat-messages';
 import { convertMoonshotAIChatUsage } from './convert-moonshotai-chat-usage';
-import type { MoonshotAIChatModelId } from './moonshotai-chat-options';
-
-export class MoonshotAIChatLanguageModel extends OpenAICompatibleChatLanguageModel {
-  constructor(
-    modelId: MoonshotAIChatModelId,
-    config: OpenAICompatibleChatConfig,
-  ) {
-    super(modelId, config);
+import { getResponseMetadata } from './get-response-metadata';
+import { mapMoonshotAIFinishReason } from './map-moonshotai-finish-reason';
+import {
+  moonshotAIChatChunkSchema,
+  moonshotAIChatResponseSchema,
+  moonshotAIErrorSchema,
+  type MoonshotAIChatTokenUsage,
+} from './moonshotai-chat-api-types';
+import {
+  getModelThinkingKeepSupport,
+  moonshotaiLanguageModelOptions,
+  type MoonshotAIChatModelId,
+} from './moonshotai-chat-options';
+import { prepareTools } from './moonshotai-prepare-tools';
+
+export type MoonshotAIChatConfig = {
+  provider: string;
+  headers?: () => Record<string, string | undefined>;
+  url: (options: { modelId: string; path: string }) => string;
+  fetch?: FetchFunction;
+  includeUsage?: boolean;
+  supportsStructuredOutputs?: boolean;
+};
+
+export class MoonshotAIChatLanguageModel implements LanguageModelV3 {
+  readonly specificationVersion = 'v3';
+
+  readonly modelId: MoonshotAIChatModelId;
+
+  // Moonshot AI does not fetch external URLs; the AI SDK downloads and
+  // inlines URL file parts instead. ms:// file references from the Moonshot
+  // Files API are passed through natively.
+  readonly supportedUrls = {
+    'image/*': [/^ms:\/\//],
+    'video/*': [/^ms:\/\//],
+  };
+
+  private readonly config: MoonshotAIChatConfig;
+  private readonly failedResponseHandler: ResponseHandler<APICallError>;
+
+  constructor(modelId: MoonshotAIChatModelId, config: MoonshotAIChatConfig) {
+    this.modelId = modelId;
+    this.config = config;
+
+    this.failedResponseHandler = createJsonErrorResponseHandler({
+      errorSchema: moonshotAIErrorSchema,
+      errorToMessage: error => error.error.message,
+    });
+  }
+
+  get provider(): string {
+    return this.config.provider;
+  }
+
+  private get providerOptionsName(): string {
+    return this.config.provider.split('.')[0].trim();
+  }
+
+  private async getArgs({
+    prompt,
+    maxOutputTokens,
+    temperature,
+    topP,
+    topK,
+    frequencyPenalty,
+    presencePenalty,
+    providerOptions,
+    stopSequences,
+    responseFormat,
+    seed,
+    toolChoice,
+    tools,
+  }: LanguageModelV3CallOptions) {
+    const moonshotOptions =
+      (await parseProviderOptions({
+        provider: this.providerOptionsName,
+        providerOptions,
+        schema: moonshotaiLanguageModelOptions,
+      })) ?? {};
+
+    const messages = convertToMoonshotAIChatMessages(prompt);
+
+    const allWarnings: SharedV3Warning[] = [];
+    if (topK != null) {
+      allWarnings.push({ type: 'unsupported', feature: 'topK' });
+    }
+    if (seed != null) {
+      allWarnings.push({ type: 'unsupported', feature: 'seed' });
+    }
+
+    const {
+      tools: moonshotTools,
+      toolChoice: moonshotToolChoice,
+      toolWarnings,
+    } = prepareTools({ tools, toolChoice });
+
+    // Thinking is configured through explicit provider options only.
+    const thinking = moonshotOptions.thinking;
+
+    // Moonshot has no reasoning_history field; the API silently ignores it
+    // (verified against the live API). Preserved Thinking maps to
+    // thinking.keep, which only accepts 'all' and only on some models
+    // (verified: k2.6, k2.7-code, k3 accept it; k2.5 rejects it). Other
+    // reasoningHistory values use the server default.
+    let keep: 'all' | undefined;
+    if (moonshotOptions.reasoningHistory === 'preserved') {
+      if (getModelThinkingKeepSupport(this.modelId)) {
+        keep = 'all';
+      } else {
+        allWarnings.push({
+          type: 'unsupported',
+          feature: `reasoningHistory 'preserved' is not supported by model "${this.modelId}"`,
+        });
+      }
+    }
+
+    let response_format: Record<string, unknown> | undefined;
+    if (responseFormat?.type === 'json') {
+      if (
+        this.config.supportsStructuredOutputs === true &&
+        responseFormat.schema != null
+      ) {
+        // kimi-k2.5 produces nonsensical output when the top-level `$schema`
+        // keyword injected by the AI SDK is present, even though it otherwise
+        // supports structured outputs. Strip it from the schema sent to
+        // Moonshot; the full original schema is still used for result
+        // validation.
+        const { $schema: _$schema, ...schemaWithoutDollarSchema } =
+          responseFormat.schema;
+        response_format = {
+          type: 'json_schema',
+          json_schema: {
+            name: responseFormat.name ?? 'response',
+            schema: schemaWithoutDollarSchema,
+            ...(responseFormat.description != null && {
+              description: responseFormat.description,
+            }),
+          },
+        };
+      } else {
+        response_format = { type: 'json_object' };
+      }
+    }
+
+    return {
+      args: {
+        model: this.modelId,
+        max_tokens: maxOutputTokens,
+        temperature,
+        top_p: topP,
+        frequency_penalty: frequencyPenalty,
+        presence_penalty: presencePenalty,
+        response_format,
+        stop: stopSequences,
+        messages,
+        tools: moonshotTools,
+        tool_choice: moonshotToolChoice,
+        ...(thinking != null || keep != null
+          ? {
+              thinking: {
+                ...(thinking?.type != null && { type: thinking.type }),
+                ...(thinking?.budgetTokens !== undefined && {
+                  budget_tokens: thinking.budgetTokens,
+                }),
+                ...(keep != null && { keep }),
+              },
+            }
+          : {}),
+        ...(moonshotOptions.reasoningEffort != null && {
+          reasoning_effort: moonshotOptions.reasoningEffort,
+        }),
+        ...(moonshotOptions.promptCacheKey != null && {
+          prompt_cache_key: moonshotOptions.promptCacheKey,
+        }),
+        ...(moonshotOptions.safetyIdentifier != null && {
+          safety_identifier: moonshotOptions.safetyIdentifier,
+        }),
+      },
+      warnings: [...allWarnings, ...toolWarnings],
+    };
   }
 
   async doGenerate(
     options: LanguageModelV3CallOptions,
   ): Promise<LanguageModelV3GenerateResult> {
-    const result = await super.doGenerate(options);
+    const { args, warnings } = await this.getArgs({ ...options });
+
+    const {
+      responseHeaders,
+      value: responseBody,
+      rawValue: rawResponse,
+    } = await postJsonToApi({
+      url: this.config.url({
+        path: '/chat/completions',
+        modelId: this.modelId,
+      }),
+      headers: combineHeaders(this.config.headers?.(), options.headers),
+      body: args,
+      failedResponseHandler: this.failedResponseHandler,
+      successfulResponseHandler: createJsonResponseHandler(
+        moonshotAIChatResponseSchema,
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch,
+    });
+
+    const choice = responseBody.choices[0];
+    const content: Array<LanguageModelV3Content> = [];
 
-    // @ts-expect-error accessing response body from parent result
-    const usage = result.response?.body?.usage;
+    // reasoning content (before text):
+    const reasoning = choice.message.reasoning_content;
+    if (reasoning != null && reasoning.length > 0) {
+      content.push({ type: 'reasoning', text: reasoning });
+    }
+
+    // tool calls:
+    if (choice.message.tool_calls != null) {
+      for (const toolCall of choice.message.tool_calls) {
+        content.push({
+          type: 'tool-call',
+          toolCallId: toolCall.id ?? generateId(),
+          toolName: toolCall.function.name,
+          input: toolCall.function.arguments ?? '',
+        });
+      }
+    }
+
+    // text content:
+    const text = choice.message.content;
+    if (text != null && text.length > 0) {
+      content.push({ type: 'text', text });
+    }
 
     return {
-      ...result,
-      usage: convertMoonshotAIChatUsage(usage),
+      content,
+      finishReason: {
+        unified: mapMoonshotAIFinishReason(choice.finish_reason),
+        raw: choice.finish_reason ?? undefined,
+      },
+      usage: convertMoonshotAIChatUsage(responseBody.usage),
+      request: { body: args },
+      response: {
+        ...getResponseMetadata(responseBody),
+        headers: responseHeaders,
+        body: rawResponse,
+      },
+      warnings,
     };
   }
 
   async doStream(
     options: LanguageModelV3CallOptions,
   ): Promise<LanguageModelV3StreamResult> {
-    const result = await super.doStream(options);
+    const { args, warnings } = await this.getArgs({ ...options });
+
+    const body = {
+      ...args,
+      stream: true,
+      ...(this.config.includeUsage && {
+        stream_options: { include_usage: true },
+      }),
+    };
+
+    const { responseHeaders, value: response } = await postJsonToApi({
+      url: this.config.url({
+        path: '/chat/completions',
+        modelId: this.modelId,
+      }),
+      headers: combineHeaders(this.config.headers?.(), options.headers),
+      body,
+      failedResponseHandler: this.failedResponseHandler,
+      successfulResponseHandler: createEventSourceResponseHandler(
+        moonshotAIChatChunkSchema,
+      ),
+      abortSignal: options.abortSignal,
+      fetch: this.config.fetch,
+    });
+
+    const toolCalls: Array<{
+      id: string;
+      type: 'function';
+      function: { name: string; arguments: string };
+      hasFinished: boolean;
+    }> = [];
+
+    let finishReason: LanguageModelV3FinishReason = {
+      unified: 'other',
+      raw: undefined,
+    };
+    let usage: MoonshotAIChatTokenUsage | undefined = undefined;
+    let isFirstChunk = true;
+    let isActiveReasoning = false;
+    let isActiveText = false;
 
     return {
-      ...result,
-      stream: result.stream.pipeThrough(
+      stream: response.pipeThrough(
         new TransformStream<
-          LanguageModelV3StreamPart,
+          ParseResult<InferSchema<typeof moonshotAIChatChunkSchema>>,
           LanguageModelV3StreamPart
         >({
+          start(controller) {
+            controller.enqueue({ type: 'stream-start', warnings });
+          },
+
           transform(chunk, controller) {
-            if (chunk.type === 'finish' && chunk.usage) {
+            // emit raw chunk if requested (before anything else):
+            if (options.includeRawChunks) {
+              controller.enqueue({ type: 'raw', rawValue: chunk.rawValue });
+            }
+
+            // handle failed chunk parsing / validation:
+            if (!chunk.success) {
+              finishReason = { unified: 'error', raw: undefined };
+              controller.enqueue({ type: 'error', error: chunk.error });
+              return;
+            }
+            const value = chunk.value;
+
+            // handle error chunks:
+            if ('error' in value) {
+              finishReason = { unified: 'error', raw: undefined };
+              controller.enqueue({ type: 'error', error: value.error.message });
+              return;
+            }
+
+            if (isFirstChunk) {
+              isFirstChunk = false;
+
+              controller.enqueue({
+                type: 'response-metadata',
+                ...getResponseMetadata(value),
+              });
+            }
+
+            if (value.usage != null) {
+              usage = value.usage;
+            }
+
+            const choice = value.choices[0];
+
+            if (choice?.finish_reason != null) {
+              finishReason = {
+                unified: mapMoonshotAIFinishReason(choice.finish_reason),
+                raw: choice.finish_reason,
+              };
+            }
+
+            if (choice?.delta == null) {
+              return;
+            }
+
+            const delta = choice.delta;
+
+            // enqueue reasoning before text deltas:
+            const reasoningContent = delta.reasoning_content;
+            if (reasoningContent) {
+              if (!isActiveReasoning) {
+                controller.enqueue({
+                  type: 'reasoning-start',
+                  id: 'reasoning-0',
+                });
+                isActiveReasoning = true;
+              }
+
+              controller.enqueue({
+                type: 'reasoning-delta',
+                id: 'reasoning-0',
+                delta: reasoningContent,
+              });
+            }
+
+            if (delta.content) {
+              if (!isActiveText) {
+                controller.enqueue({ type: 'text-start', id: 'txt-0' });
+                isActiveText = true;
+              }
+
+              // end reasoning when text starts:
+              if (isActiveReasoning) {
+                controller.enqueue({
+                  type: 'reasoning-end',
+                  id: 'reasoning-0',
+                });
+                isActiveReasoning = false;
+              }
+
               controller.enqueue({
-                ...chunk,
-                usage: convertMoonshotAIChatUsage(chunk.usage.raw as any),
+                type: 'text-delta',
+                id: 'txt-0',
+                delta: delta.content,
               });
-            } else {
-              controller.enqueue(chunk);
             }
+
+            if (delta.tool_calls != null) {
+              // end reasoning when tool calls start:
+              if (isActiveReasoning) {
+                controller.enqueue({
+                  type: 'reasoning-end',
+                  id: 'reasoning-0',
+                });
+                isActiveReasoning = false;
+              }
+
+              for (const toolCallDelta of delta.tool_calls) {
+                const index = toolCallDelta.index;
+
+                if (toolCalls[index] == null) {
+                  if (toolCallDelta.id == null) {
+                    throw new InvalidResponseDataError({
+                      data: toolCallDelta,
+                      message: `Expected 'id' to be a string.`,
+                    });
+                  }
+
+                  if (toolCallDelta.function?.name == null) {
+                    throw new InvalidResponseDataError({
+                      data: toolCallDelta,
+                      message: `Expected 'function.name' to be a string.`,
+                    });
+                  }
+
+                  controller.enqueue({
+                    type: 'tool-input-start',
+                    id: toolCallDelta.id,
+                    toolName: toolCallDelta.function.name,
+                  });
+
+                  toolCalls[index] = {
+                    id: toolCallDelta.id,
+                    type: 'function',
+                    function: {
+                      name: toolCallDelta.function.name,
+                      arguments: toolCallDelta.function.arguments ?? '',
+                    },
+                    hasFinished: false,
+                  };
+
+                  const toolCall = toolCalls[index];
+
+                  if (
+                    toolCall.function?.name != null &&
+                    toolCall.function?.arguments != null &&
+                    toolCall.function.arguments.length > 0
+                  ) {
+                    // send delta if the argument text has already started:
+                    controller.enqueue({
+                      type: 'tool-input-delta',
+                      id: toolCall.id,
+                      delta: toolCall.function.arguments,
+                    });
+                  }
+
+                  continue;
+                }
+
+                // existing tool call, merge if not finished
+                const toolCall = toolCalls[index];
+
+                if (toolCall.hasFinished) {
+                  continue;
+                }
+
+                if (toolCallDelta.function?.arguments != null) {
+                  toolCall.function.arguments +=
+                    toolCallDelta.function.arguments;
+                }
+
+                // send delta
+                controller.enqueue({
+                  type: 'tool-input-delta',
+                  id: toolCall.id,
+                  delta: toolCallDelta.function.arguments ?? '',
+                });
+              }
+            }
+          },
+
+          flush(controller) {
+            if (isActiveReasoning) {
+              controller.enqueue({ type: 'reasoning-end', id: 'reasoning-0' });
+            }
+
+            if (isActiveText) {
+              controller.enqueue({ type: 'text-end', id: 'txt-0' });
+            }
+
+            // go through all tool calls and send the ones that are not finished
+            for (const toolCall of toolCalls.filter(
+              toolCall => !toolCall.hasFinished,
+            )) {
+              controller.enqueue({
+                type: 'tool-input-end',
+                id: toolCall.id,
+              });
+
+              controller.enqueue({
+                type: 'tool-call',
+                toolCallId: toolCall.id,
+                toolName: toolCall.function.name,
+                input: toolCall.function.arguments,
+              });
+            }
+
+            controller.enqueue({
+              type: 'finish',
+              finishReason,
+              usage: convertMoonshotAIChatUsage(usage),
+            });
           },
         }),
       ),
+      request: { body },
+      response: { headers: responseHeaders },
     };
   }
 }

package/src/moonshotai-chat-options.ts

@@ -15,9 +15,9 @@ export type MoonshotAIChatModelId =
 
 export const moonshotaiLanguageModelOptions = z.object({
   /**
-   * Reasoning effort for Kimi K3. Currently, only `max` is supported.
+   * Reasoning effort for Kimi K3.
    */
-  reasoningEffort: z.literal('max').optional(),
+  reasoningEffort: z.enum(['low', 'high', 'max']).optional(),
 
   thinking: z
     .object({
@@ -27,8 +27,35 @@ export const moonshotaiLanguageModelOptions = z.object({
     .optional(),
 
   reasoningHistory: z.enum(['disabled', 'interleaved', 'preserved']).optional(),
+
+  /**
+   * Used to cache responses for similar requests to optimize cache hit rates.
+   * Typically a session or task id.
+   */
+  promptCacheKey: z.string().optional(),
+
+  /**
+   * A stable identifier used to help Moonshot detect users violating usage
+   * policies. Recommended to hash the username or email address.
+   */
+  safetyIdentifier: z.string().optional(),
 });
 
 export type MoonshotAILanguageModelOptions = z.infer<
   typeof moonshotaiLanguageModelOptions
 >;
+
+/**
+ * Whether the model accepts `thinking.keep` (Preserved Thinking). Verified
+ * against the live API: kimi-k2.6, kimi-k2.7-code(+highspeed), and kimi-k3
+ * accept `keep: 'all'`; other models reject it with a 400.
+ */
+export function getModelThinkingKeepSupport(
+  modelId: MoonshotAIChatModelId,
+): boolean {
+  return (
+    modelId === 'kimi-k2.6' ||
+    modelId === 'kimi-k3' ||
+    modelId.startsWith('kimi-k2.7-code')
+  );
+}