kimi-vercel-ai-sdk-provider 0.2.0

package/src/code/kimi-code-settings.ts

@@ -0,0 +1,193 @@
+/**
+ * Settings for the Kimi Code provider.
+ *
+ * @remarks
+ * Kimi Code is compatible with Claude Code and Roo Code agents.
+ * Default settings are based on Roo Code documentation:
+ * - Max Output Tokens: 32768
+ * - Context Window Size: 262144
+ * - Reasoning Effort: Medium
+ *
+ * @see https://www.kimi.com/code/docs/en/more/third-party-agents.html
+ * @module
+ */
+
+import type { ExtendedThinkingConfig, KimiCodeCapabilities, ReasoningEffort } from './kimi-code-types';
+import { z } from 'zod/v4';
+
+// ============================================================================
+// Settings Interface
+// ============================================================================
+
+/**
+ * Settings for creating a Kimi Code model instance.
+ */
+export interface KimiCodeSettings {
+  /**
+   * Override inferred model capabilities.
+   */
+  capabilities?: KimiCodeCapabilities;
+
+  /**
+   * Extended thinking/reasoning configuration.
+   * When enabled, the model will show its reasoning process.
+   *
+   * @example
+   * ```ts
+   * const model = kimiCode('kimi-for-coding', {
+   *   extendedThinking: {
+   *     enabled: true,
+   *     effort: 'high'
+   *   }
+   * });
+   * ```
+   */
+  extendedThinking?: ExtendedThinkingConfig | boolean;
+
+  /**
+   * Whether to include usage in streaming responses.
+   */
+  includeUsageInStream?: boolean;
+
+  /**
+   * Override supported URL patterns.
+   */
+  supportedUrls?: Record<string, RegExp[]>;
+}
+
+// ============================================================================
+// Provider Options Schema
+// ============================================================================
+
+/**
+ * Schema for Kimi Code provider options passed via providerOptions.
+ */
+export const kimiCodeProviderOptionsSchema = z.object({
+  /**
+   * Extended thinking configuration.
+   */
+  extendedThinking: z
+    .union([
+      z.boolean(),
+      z.object({
+        enabled: z.boolean().optional(),
+        effort: z.enum(['low', 'medium', 'high']).optional(),
+        budgetTokens: z.number().optional()
+      })
+    ])
+    .optional(),
+
+  /**
+   * System prompt to prepend.
+   */
+  system: z.string().optional(),
+
+  /**
+   * Custom stop sequences.
+   */
+  stopSequences: z.array(z.string()).optional()
+});
+
+/**
+ * Inferred type from the provider options schema.
+ */
+export type KimiCodeProviderOptions = z.infer<typeof kimiCodeProviderOptionsSchema>;
+
+// ============================================================================
+// Helper Functions
+// ============================================================================
+
+/**
+ * Normalize extended thinking config to a consistent format.
+ *
+ * @param config - Boolean or config object
+ * @returns Normalized config object
+ */
+export function normalizeExtendedThinkingConfig(
+  config: ExtendedThinkingConfig | boolean | undefined
+): ExtendedThinkingConfig | undefined {
+  if (config === undefined) {
+    return undefined;
+  }
+
+  if (typeof config === 'boolean') {
+    return config ? { enabled: true, effort: 'medium' } : { enabled: false };
+  }
+
+  return {
+    enabled: config.enabled ?? true,
+    effort: config.effort ?? 'medium',
+    budgetTokens: config.budgetTokens
+  };
+}
+
+/**
+ * Convert reasoning effort to budget tokens.
+ *
+ * @param effort - Reasoning effort level
+ * @returns Approximate budget tokens for the effort level
+ */
+export function effortToBudgetTokens(effort: ReasoningEffort): number {
+  switch (effort) {
+    case 'low':
+      return 2048;
+    case 'medium':
+      return 8192;
+    case 'high':
+      return 16384;
+    default:
+      return 8192;
+  }
+}
+
+// ============================================================================
+// Anthropic API Conversion Helpers
+// ============================================================================
+
+/**
+ * Convert ExtendedThinkingConfig to Anthropic API thinking parameter format.
+ *
+ * @remarks
+ * This converts our config to the Anthropic API format expected by Kimi Code:
+ * - `{ type: 'enabled', budget_tokens: number }` when enabled
+ * - `{ type: 'disabled' }` when disabled
+ * - `undefined` when not configured
+ *
+ * @param config - Extended thinking configuration (boolean or object)
+ * @returns Anthropic-compatible thinking parameter or undefined
+ *
+ * @example
+ * ```ts
+ * // From boolean
+ * toAnthropicThinking(true)
+ * // => { type: 'enabled', budget_tokens: 8192 }
+ *
+ * // From config object
+ * toAnthropicThinking({ enabled: true, effort: 'high' })
+ * // => { type: 'enabled', budget_tokens: 16384 }
+ *
+ * // With explicit budget
+ * toAnthropicThinking({ enabled: true, budgetTokens: 10000 })
+ * // => { type: 'enabled', budget_tokens: 10000 }
+ * ```
+ */
+export function toAnthropicThinking(
+  config: ExtendedThinkingConfig | boolean | undefined
+): { type: 'enabled'; budget_tokens: number } | { type: 'disabled' } | undefined {
+  const normalized = normalizeExtendedThinkingConfig(config);
+
+  if (normalized === undefined) {
+    return undefined;
+  }
+
+  if (!normalized.enabled) {
+    return { type: 'disabled' };
+  }
+
+  const budgetTokens = normalized.budgetTokens ?? effortToBudgetTokens(normalized.effort ?? 'medium');
+
+  return {
+    type: 'enabled',
+    budget_tokens: budgetTokens
+  };
+}

package/src/code/kimi-code-types.ts

@@ -0,0 +1,354 @@
+/**
+ * Types for the Kimi Code provider.
+ *
+ * Kimi Code is a premium coding service within the Kimi ecosystem that provides:
+ * - High-speed output (up to 100 tokens/s)
+ * - Extended thinking/reasoning support
+ * - Full compatibility with Claude Code and Roo Code
+ * - Anthropic-compatible API format
+ *
+ * @see https://www.kimi.com/code/docs/en/
+ * @module
+ */
+
+import type { LanguageModelV3 } from '@ai-sdk/provider';
+
+// ============================================================================
+// Constants
+// ============================================================================
+
+/**
+ * Default Kimi Code API endpoint (Anthropic-compatible).
+ * Used with Claude Code - set as ANTHROPIC_BASE_URL
+ */
+export const KIMI_CODE_BASE_URL = 'https://api.kimi.com/coding/v1';
+
+/**
+ * OpenAI-compatible API endpoint.
+ * Used with Roo Code - set as OpenAI Compatible entrypoint
+ */
+export const KIMI_CODE_OPENAI_BASE_URL = 'https://api.kimi.com/coding/v1';
+
+/**
+ * Default Kimi Code model ID.
+ * Primary coding model optimized for development tasks.
+ */
+export const KIMI_CODE_DEFAULT_MODEL = 'kimi-for-coding';
+
+/**
+ * Alternative model with enhanced thinking/reasoning.
+ * Can be toggled with Tab key in Claude Code.
+ */
+export const KIMI_CODE_THINKING_MODEL = 'kimi-k2-thinking';
+
+/**
+ * All available Kimi Code models.
+ */
+export const KIMI_CODE_MODELS = [KIMI_CODE_DEFAULT_MODEL, KIMI_CODE_THINKING_MODEL] as const;
+
+/**
+ * Default max output tokens for Kimi Code (per Roo Code docs).
+ */
+export const KIMI_CODE_DEFAULT_MAX_TOKENS = 32768;
+
+/**
+ * Default context window size (per Roo Code docs).
+ */
+export const KIMI_CODE_DEFAULT_CONTEXT_WINDOW = 262144;
+
+/**
+ * Anthropic API version header value.
+ */
+export const KIMI_CODE_ANTHROPIC_VERSION = '2023-06-01';
+
+// ============================================================================
+// Model IDs
+// ============================================================================
+
+/**
+ * Available Kimi Code model IDs.
+ *
+ * @remarks
+ * - `kimi-for-coding` - Primary coding model optimized for development tasks
+ * - `kimi-k2-thinking` - Model with extended thinking for complex reasoning (toggle with Tab in Claude Code)
+ *
+ * @example
+ * ```ts
+ * // Default model
+ * const model = kimiCode('kimi-for-coding');
+ *
+ * // Thinking model
+ * const thinkingModel = kimiCode('kimi-k2-thinking');
+ * ```
+ */
+export type KimiCodeModelId = 'kimi-for-coding' | 'kimi-k2-thinking' | (string & {});
+
+// ============================================================================
+// Model Capabilities
+// ============================================================================
+
+/**
+ * Capabilities specific to Kimi Code models.
+ * Based on Roo Code configuration documentation.
+ */
+export interface KimiCodeCapabilities {
+  /**
+   * Whether the model supports extended thinking/reasoning.
+   * When enabled, use `thinking.type: 'enabled'` with `budget_tokens`.
+   */
+  extendedThinking?: boolean;
+
+  /**
+   * Maximum output tokens.
+   * Default: 32768 (per Roo Code docs)
+   */
+  maxOutputTokens?: number;
+
+  /**
+   * Maximum context window size.
+   * Default: 262144 (per Roo Code docs)
+   */
+  maxContextSize?: number;
+
+  /**
+   * Whether the model supports streaming.
+   * Always true for Kimi Code.
+   */
+  streaming?: boolean;
+
+  /**
+   * Whether the model supports tool/function calling.
+   * Always true for Kimi Code.
+   */
+  toolCalling?: boolean;
+
+  /**
+   * Whether the model supports image inputs.
+   */
+  imageInput?: boolean;
+}
+
+/**
+ * Infer model capabilities from the model ID.
+ *
+ * @param modelId - The model identifier
+ * @returns Inferred capabilities based on model name patterns
+ *
+ * @example
+ * ```ts
+ * const caps = inferKimiCodeCapabilities('kimi-k2-thinking');
+ * // caps.extendedThinking === true
+ * ```
+ */
+export function inferKimiCodeCapabilities(modelId: string): KimiCodeCapabilities {
+  const isThinkingModel = modelId.includes('-thinking') || modelId.includes('k2-thinking');
+
+  return {
+    extendedThinking: isThinkingModel,
+    maxOutputTokens: KIMI_CODE_DEFAULT_MAX_TOKENS,
+    maxContextSize: KIMI_CODE_DEFAULT_CONTEXT_WINDOW,
+    streaming: true,
+    toolCalling: true,
+    imageInput: true
+  };
+}
+
+// ============================================================================
+// Provider Configuration
+// ============================================================================
+
+/**
+ * Configuration for the Kimi Code language model.
+ * @internal
+ */
+export interface KimiCodeConfig {
+  /**
+   * Provider identifier.
+   */
+  provider: string;
+
+  /**
+   * Base URL for the API.
+   */
+  baseURL: string;
+
+  /**
+   * Function to get headers for requests.
+   */
+  headers: () => Record<string, string | undefined>;
+
+  /**
+   * Custom fetch implementation.
+   */
+  fetch?: typeof globalThis.fetch;
+
+  /**
+   * ID generator for tool call fallback IDs.
+   */
+  generateId?: () => string;
+
+  /**
+   * Whether to include usage in streaming responses.
+   */
+  includeUsageInStream?: boolean;
+
+  /**
+   * Override supported URL patterns.
+   */
+  supportedUrls?: LanguageModelV3['supportedUrls'];
+}
+
+// ============================================================================
+// Extended Thinking Configuration
+// ============================================================================
+
+/**
+ * Reasoning effort levels for extended thinking.
+ * Maps to budget_tokens for the thinking parameter.
+ *
+ * Per Roo Code docs: Enable Reasoning Effort: Medium
+ */
+export type ReasoningEffort = 'low' | 'medium' | 'high';
+
+/**
+ * Configuration for extended thinking/reasoning.
+ * Compatible with Anthropic's thinking parameter format.
+ *
+ * @example
+ * ```ts
+ * // Enable thinking with medium effort
+ * const config: ExtendedThinkingConfig = {
+ *   enabled: true,
+ *   effort: 'medium'
+ * };
+ *
+ * // Or specify exact budget tokens
+ * const config: ExtendedThinkingConfig = {
+ *   enabled: true,
+ *   budgetTokens: 10000
+ * };
+ * ```
+ */
+export interface ExtendedThinkingConfig {
+  /**
+   * Enable extended thinking mode.
+   * @default false
+   */
+  enabled?: boolean;
+
+  /**
+   * Reasoning effort level.
+   * Controls how much computation is spent on reasoning.
+   * - low: ~2048 tokens
+   * - medium: ~8192 tokens (default, recommended by Roo Code)
+   * - high: ~16384 tokens
+   * @default 'medium'
+   */
+  effort?: ReasoningEffort;
+
+  /**
+   * Budget tokens for thinking (alternative to effort).
+   * Higher values allow for more complex reasoning.
+   * Takes precedence over effort if both are specified.
+   */
+  budgetTokens?: number;
+}
+
+// ============================================================================
+// API Response Types (Anthropic-compatible)
+// ============================================================================
+
+/**
+ * Token usage from Kimi Code API.
+ * Follows Anthropic's usage format.
+ */
+export interface KimiCodeTokenUsage {
+  input_tokens?: number | null;
+  output_tokens?: number | null;
+  cache_read_input_tokens?: number | null;
+  cache_creation_input_tokens?: number | null;
+}
+
+/**
+ * Response metadata from Kimi Code API.
+ * Follows Anthropic's message format.
+ */
+export interface KimiCodeResponseMetadata {
+  id?: string | null;
+  model?: string | null;
+  type?: string | null;
+  stop_reason?: string | null;
+  stop_sequence?: string | null;
+}
+
+/**
+ * Thinking block in response (for extended thinking).
+ * Appears when thinking is enabled.
+ */
+export interface KimiCodeThinkingBlock {
+  type: 'thinking';
+  thinking: string;
+}
+
+/**
+ * Text block in response.
+ */
+export interface KimiCodeTextBlock {
+  type: 'text';
+  text: string;
+}
+
+/**
+ * Tool use block in response.
+ * Follows Anthropic's tool_use format.
+ */
+export interface KimiCodeToolUseBlock {
+  type: 'tool_use';
+  id: string;
+  name: string;
+  input: Record<string, unknown>;
+}
+
+/**
+ * Content block types in Kimi Code responses.
+ */
+export type KimiCodeContentBlock = KimiCodeThinkingBlock | KimiCodeTextBlock | KimiCodeToolUseBlock;
+
+// ============================================================================
+// Streaming Event Types (Anthropic SSE format)
+// ============================================================================
+
+/**
+ * Streaming event types from Kimi Code API.
+ * Follows Anthropic's SSE format.
+ *
+ * Event sequence:
+ * 1. message_start - Contains message metadata and initial usage
+ * 2. content_block_start - Start of a content block (text, thinking, tool_use)
+ * 3. content_block_delta - Incremental content updates
+ * 4. content_block_stop - End of a content block
+ * 5. message_delta - Final message updates (stop_reason, usage)
+ * 6. message_stop - End of message
+ *
+ * @see https://docs.anthropic.com/claude/reference/streaming
+ */
+export type KimiCodeStreamEventType =
+  | 'message_start'
+  | 'content_block_start'
+  | 'content_block_delta'
+  | 'content_block_stop'
+  | 'message_delta'
+  | 'message_stop'
+  | 'ping'
+  | 'error';
+
+/**
+ * Delta types for content_block_delta events.
+ */
+export type KimiCodeDeltaType = 'text_delta' | 'thinking_delta' | 'input_json_delta';
+
+/**
+ * Stop reasons from Kimi Code API.
+ * Follows Anthropic's stop_reason format.
+ */
+export type KimiCodeStopReason = 'end_turn' | 'max_tokens' | 'stop_sequence' | 'tool_use' | null;

package/src/core/errors.ts

@@ -0,0 +1,140 @@
+/**
+ * Custom error classes for the Kimi provider.
+ * @module
+ */
+
+import { createJsonErrorResponseHandler } from '@ai-sdk/provider-utils';
+import { z } from 'zod/v4';
+
+// ============================================================================
+// Error Schema
+// ============================================================================
+
+/**
+ * Zod schema for Kimi API error responses.
+ * Handles both standard OpenAI-style errors and simple message errors.
+ */
+export const kimiErrorSchema = z.union([
+  z.object({
+    error: z.object({
+      message: z.string(),
+      type: z.string().nullish(),
+      param: z.any().nullish(),
+      code: z.union([z.string(), z.number()]).nullish(),
+      request_id: z.string().nullish()
+    })
+  }),
+  z.object({
+    message: z.string()
+  })
+]);
+
+/**
+ * Type for Kimi API error data.
+ */
+export type KimiErrorData = z.infer<typeof kimiErrorSchema>;
+
+// ============================================================================
+// Error Handler
+// ============================================================================
+
+/**
+ * Failed response handler for Kimi API errors.
+ * Parses error responses and creates appropriate error objects.
+ */
+export const kimiFailedResponseHandler = createJsonErrorResponseHandler({
+  errorSchema: kimiErrorSchema,
+  errorToMessage: (error: KimiErrorData) => {
+    if ('error' in error) {
+      return error.error.message;
+    }
+    return error.message;
+  },
+  isRetryable: (response) =>
+    response.status === 408 || response.status === 409 || response.status === 429 || response.status >= 500
+});
+
+// ============================================================================
+// Custom Error Classes
+// ============================================================================
+
+/**
+ * Base error class for Kimi provider errors.
+ */
+export class KimiError extends Error {
+  readonly code: string;
+  readonly statusCode?: number;
+
+  constructor(message: string, code: string, statusCode?: number) {
+    super(message);
+    this.name = 'KimiError';
+    this.code = code;
+    this.statusCode = statusCode;
+  }
+}
+
+/**
+ * Error thrown when authentication fails.
+ */
+export class KimiAuthenticationError extends KimiError {
+  constructor(message: string = 'Invalid API key or authentication failed') {
+    super(message, 'authentication_error', 401);
+    this.name = 'KimiAuthenticationError';
+  }
+}
+
+/**
+ * Error thrown when rate limit is exceeded.
+ */
+export class KimiRateLimitError extends KimiError {
+  readonly retryAfter?: number;
+
+  constructor(message: string = 'Rate limit exceeded', retryAfter?: number) {
+    super(message, 'rate_limit_error', 429);
+    this.name = 'KimiRateLimitError';
+    this.retryAfter = retryAfter;
+  }
+}
+
+/**
+ * Error thrown when the request is invalid.
+ */
+export class KimiValidationError extends KimiError {
+  constructor(message: string, param?: string) {
+    super(param ? `${message} (param: ${param})` : message, 'validation_error', 400);
+    this.name = 'KimiValidationError';
+  }
+}
+
+/**
+ * Error thrown when a model is not found.
+ */
+export class KimiModelNotFoundError extends KimiError {
+  readonly modelId: string;
+
+  constructor(modelId: string) {
+    super(`Model '${modelId}' not found`, 'model_not_found', 404);
+    this.name = 'KimiModelNotFoundError';
+    this.modelId = modelId;
+  }
+}
+
+/**
+ * Error thrown when content is filtered.
+ */
+export class KimiContentFilterError extends KimiError {
+  constructor(message: string = 'Content was filtered due to policy violation') {
+    super(message, 'content_filter', 400);
+    this.name = 'KimiContentFilterError';
+  }
+}
+
+/**
+ * Error thrown when context length is exceeded.
+ */
+export class KimiContextLengthError extends KimiError {
+  constructor(message: string = 'Context length exceeded') {
+    super(message, 'context_length_exceeded', 400);
+    this.name = 'KimiContextLengthError';
+  }
+}