AIX: Anthropic: minor API changes

src/modules/aix/server/api/aix.wiretypes.ts

@@ -497,8 +497,8 @@ export namespace AixWire_API {
      */
     userGeolocation: z.object({
       city: z.string().optional(),      // free text input for the city of the user, e.g. San Francisco.
-      country: z.string().optional(),   // two-letter ISO country code of the user, e.g. US
       region: z.string().optional(),    // free text input for the reg. of the user the user, e.g. California
+      country: z.string().optional(),   // two-letter ISO country code of the user, e.g. US
       timezone: z.string().optional(),  // IANA timezone of the user, e.g. America/Los_Angeles
     }).optional(),
   });

src/modules/aix/server/dispatch/chatGenerate/adapters/anthropic.messageCreate.ts

@@ -136,6 +136,9 @@ export function aixToAnthropicMessageCreate(model: AixAPI_Model, _chatGenerate:
   const areToolCallsRequired = payload.tool_choice && typeof payload.tool_choice === 'object' && (payload.tool_choice.type === 'any' || payload.tool_choice.type === 'tool');
   const canUseThinking = !areToolCallsRequired || !hotFixDisableThinkingWhenToolsForced;
   if (model.vndAntThinkingBudget !== undefined && canUseThinking) {
+    // NOTE: We don't want to maintain patches here - so just FYI: budget_tokens is deprecated on Opus 4.6 in favor of adaptive thinking
+    // if (typeof model.vndAntThinkingBudget === 'number' && model.id.includes('opus-4-6'))
+    //   console.warn('[Anthropic] budget_tokens is deprecated on Opus 4.6 - use adaptive thinking instead');
     payload.thinking = model.vndAntThinkingBudget === 'adaptive' ? {
       type: 'adaptive',
     } : model.vndAntThinkingBudget !== null ? {
@@ -187,6 +190,10 @@ export function aixToAnthropicMessageCreate(model: AixAPI_Model, _chatGenerate:
         type: 'web_search_20250305',
         name: 'web_search',
         max_uses: 10, // Allow up to 10 progressive searches // FIXME: HARDCODED
+        // Pass user geolocation for location-aware search results
+        ...(model.userGeolocation ? {
+          user_location: { type: 'approximate' as const, ...model.userGeolocation },
+        } : {}),
       });
     }
 

src/modules/llms/server/anthropic/anthropic.access.ts

@@ -52,29 +52,15 @@ const DEFAULT_ANTHROPIC_HEADERS = {
 
 const DEFAULT_ANTHROPIC_BETA_FEATURES: string[] = [
 
-  // NOTE: undocumented: I wonder what this is for
-  // 'claude-code-20250219',
+  // Known SDK beta headers (for reference, not all used):
+  //   prompt-caching-2024-07-31        -- GA: no longer needed
+  //   pdfs-2024-09-25                  -- GA: no longer needed
+  //   token-efficient-tools-2025-02-19 -- not used; disabled for now as side-effects are untested
+  //   extended-cache-ttl-2025-04-11    -- for 1h cache TTL; we support ttl:'1h' in wiretypes already
+  //   interleaved-thinking-2025-05-14  -- for Claude 4/4.5 interleaved thinking (auto on Opus 4.6 adaptive)
+  //   context-management-2025-06-27    -- for context_management edits (e.g. clear_tool_uses)
+  //   model-context-window-exceeded-2025-08-26 -- Sonnet 4.5+ have this by default
 
-  // NOTE: disabled for now, as we don't have tested side-effects for this feature yet
-  // 'token-efficient-tools-2025-02-19', // https://docs.anthropic.com/en/docs/build-with-claude/tool-use/token-efficient-tool-use
-
-  /**
-   * to use the prompt caching feature; adds to any API invocation:
-   *  - message_start.message.usage.cache_creation_input_tokens: number
-   *  - message_start.message.usage.cache_read_input_tokens: number
-   */
-  'prompt-caching-2024-07-31',
-
-  /**
-   * Enables model_context_window_exceeded stop reason for models earlier than Sonnet 4.5
-   * (Sonnet 4.5+ have this by default). This allows requesting max tokens without calculating
-   * input size, and the API will return as much as possible within the context window.
-   * https://docs.claude.com/en/api/handling-stop-reasons#model-context-window-exceeded
-   */
-  // 'model-context-window-exceeded-2025-08-26',
-
-  // now default
-  // 'messages-2023-12-15'
 ] as const;
 
 const PER_MODEL_BETA_FEATURES: { [modelId: string]: string[] } = {