@fairyhunter13/ai-anthropic 3.0.58-fork.8 → 3.0.74-fork

package/src/{anthropic-messages-options.ts → anthropic-language-model-options.ts}

@@ -1,7 +1,7 @@
 import { z } from 'zod/v4';
 
 // https://docs.claude.com/en/docs/about-claude/models/overview
-export type AnthropicMessagesModelId =
+export type AnthropicModelId =
   | 'claude-3-haiku-20240307'
   | 'claude-haiku-4-5-20251001'
   | 'claude-haiku-4-5'
@@ -17,6 +17,7 @@ export type AnthropicMessagesModelId =
   | 'claude-sonnet-4-5'
   | 'claude-sonnet-4-6'
   | 'claude-opus-4-6'
+  | 'claude-opus-4-7'
   | (string & {});
 
 /**
@@ -84,9 +85,9 @@ export const anthropicLanguageModelOptions = z.object({
         /** for Sonnet 4.6, Opus 4.6, and newer models */
         type: z.literal('adaptive'),
         /**
-         * Controls thinking output visibility.
-         * - 'omitted': thinking happens but output is not returned (faster TTFT)
-         * - 'summarized': a summary of the thinking is returned
+         * Controls whether thinking content is included in the response.
+         * - `"omitted"`: Thinking blocks are present but text is empty (default for Opus 4.7+).
+         * - `"summarized"`: Thinking content is returned. Required to see reasoning output.
          */
         display: z.enum(['omitted', 'summarized']).optional(),
       }),
@@ -94,12 +95,6 @@ export const anthropicLanguageModelOptions = z.object({
         /** for models before Opus 4.6, except Sonnet 4.6 still supports it */
         type: z.literal('enabled'),
         budgetTokens: z.number().optional(),
-        /**
-         * Controls thinking output visibility.
-         * - 'omitted': thinking happens but output is not returned (faster TTFT)
-         * - 'summarized': a summary of the thinking is returned
-         */
-        display: z.enum(['omitted', 'summarized']).optional(),
       }),
       z.object({
         type: z.literal('disabled'),
@@ -124,6 +119,23 @@ export const anthropicLanguageModelOptions = z.object({
     })
     .optional(),
 
+  /**
+   * Metadata to include with the request.
+   *
+   * See https://platform.claude.com/docs/en/api/messages/create for details.
+   */
+  metadata: z
+    .object({
+      /**
+       * An external identifier for the user associated with the request.
+       *
+       * Should be a UUID, hash value, or other opaque identifier.
+       * Must not contain PII (name, email, phone number, etc.).
+       */
+      userId: z.string().optional(),
+    })
+    .optional(),
+
   /**
    * MCP servers to be utilized in this request.
    */
@@ -154,21 +166,29 @@ export const anthropicLanguageModelOptions = z.object({
       id: z.string().optional(),
       skills: z
         .array(
-          z.object({
-            type: z.union([z.literal('anthropic'), z.literal('custom')]),
-            skillId: z.string(),
-            version: z.string().optional(),
-          }),
+          z.discriminatedUnion('type', [
+            z.object({
+              type: z.literal('anthropic'),
+              skillId: z.string(),
+              version: z.string().optional(),
+            }),
+            z.object({
+              type: z.literal('custom'),
+              providerReference: z.record(z.string(), z.string()),
+              version: z.string().optional(),
+            }),
+          ]),
         )
         .optional(),
     })
     .optional(),
 
   /**
-   * Whether to enable tool streaming (and structured output streaming).
-   *
-   * When set to false, the model will return all tool calls and results
-   * at once after a delay.
+   * Whether to enable fine-grained (eager) streaming of tool call inputs
+   * and structured outputs for every function tool in the request. When
+   * true (the default), each function tool receives a default of
+   * `eager_input_streaming: true` unless it explicitly sets
+   * `providerOptions.anthropic.eagerInputStreaming`.
    *
    * @default true
    */
@@ -177,7 +197,22 @@ export const anthropicLanguageModelOptions = z.object({
   /**
    * @default 'high'
    */
-  effort: z.enum(['low', 'medium', 'high', 'max']).optional(),
+  effort: z.enum(['low', 'medium', 'high', 'xhigh', 'max']).optional(),
+
+  /**
+   * Task budget for agentic turns. Informs the model of the total token budget
+   * available for the current task, allowing it to prioritize work and wind down
+   * gracefully as the budget is consumed.
+   *
+   * Advisory only — does not enforce a hard token limit.
+   */
+  taskBudget: z
+    .object({
+      type: z.literal('tokens'),
+      total: z.number().int().min(20000),
+      remaining: z.number().int().min(0).optional(),
+    })
+    .optional(),
 
   /**
    * Enable fast mode for faster inference (2.5x faster output token speeds).
@@ -186,37 +221,41 @@ export const anthropicLanguageModelOptions = z.object({
   speed: z.enum(['fast', 'standard']).optional(),
 
   /**
-   * A set of beta features to enable.
-   * Allow a provider to receive the full `betas` set if it needs it.
+   * Advisor tool configuration. When enabled, Claude consults a cheaper
+   * advisor model before generating its response, improving tool-use
+   * accuracy and reducing cost on complex agentic tasks.
+   *
+   * Requires the `advisor-tool-2026-03-01` beta header (added automatically).
    */
-  anthropicBeta: z.array(z.string()).optional(),
+  advisor: z
+    .object({
+      /** The model ID for the advisor (e.g. "claude-haiku-4-5"). */
+      model: z.string(),
+    })
+    .optional(),
 
   /**
-   * Assistant prefill text. When set, this text is appended as a partial
-   * assistant message at the end of the prompt, guiding the model's output.
-   * The model will continue generating from where the prefill ends.
-   *
-   * Example: Setting prefill to '{"' will make the model start its response with '{"'
+   * Enable 300K output tokens per response (vs default ~8K/32K caps).
+   * When true, the `output-300k-2026-03-24` beta header is sent automatically
+   * and max_tokens capping is raised to 300,000.
    */
-  prefill: z.string().optional(),
+  output300k: z.boolean().optional(),
 
   /**
-   * Request metadata. Use `user_id` to track per-user usage for abuse detection
-   * and rate limit apportioning. Should be a UUID or similar opaque identifier.
-   * See https://docs.anthropic.com/en/api/messages
+   * Controls where model inference runs for this request.
+   *
+   * - `"global"`: Inference may run in any available geography (default).
+   * - `"us"`: Inference runs only in US-based infrastructure.
+   *
+   * See https://platform.claude.com/docs/en/build-with-claude/data-residency
    */
-  metadata: z
-    .object({
-      userId: z.string().optional(),
-    })
-    .optional(),
+  inferenceGeo: z.enum(['us', 'global']).optional(),
 
   /**
-   * Service tier selection.
-   * - 'auto': Allow Anthropic to pick standard or priority based on availability.
-   * See https://docs.anthropic.com/en/api/messages
+   * A set of beta features to enable.
+   * Allow a provider to receive the full `betas` set if it needs it.
    */
-  serviceTier: z.enum(['auto']).optional(),
+  anthropicBeta: z.array(z.string()).optional(),
 
   contextManagement: z
     .object({