llmist 17.1.0 → 17.2.1

package/dist/index.d.cts

@@ -9779,10 +9779,28 @@ declare class OpenRouterProvider extends OpenAICompatibleProvider<OpenRouterConf
     constructor(client: OpenAI, config?: OpenRouterConfig);
     getModelSpecs(): ModelSpec[];
     /**
-     * Override buildApiRequest to inject reasoning parameters.
-     * OpenRouter normalizes reasoning into the standard OpenAI format.
+     * Override buildApiRequest to inject reasoning parameters and cache_control breakpoints.
+     * OpenRouter normalizes reasoning into the standard OpenAI format,
+     * and supports cache_control on message content blocks for both
+     * Anthropic Claude and Google Gemini models.
      */
     protected buildApiRequest(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): Parameters<OpenAI["chat"]["completions"]["create"]>[0];
+    /** Minimal shape for messages in the already-built OpenAI-compatible request. */
+    private static readonly CACHE_CONTROL;
+    /**
+     * Add cache_control breakpoints to the last system message and last user message.
+     * This enables OpenRouter's prompt caching for supported providers (Anthropic, Gemini).
+     *
+     * Operates on the already-built request object. We cast through `unknown` because
+     * OpenAI's `ChatCompletionMessageParam` union is too narrow to assign content arrays
+     * with the non-standard `cache_control` property.
+     */
+    private injectCacheBreakpoints;
+    /**
+     * Return a new content array with cache_control on the last block.
+     * String content is promoted to a single-element text block array.
+     */
+    private withCacheControl;
     /**
      * Get custom headers for OpenRouter analytics.
      */

package/dist/index.d.ts

@@ -9779,10 +9779,28 @@ declare class OpenRouterProvider extends OpenAICompatibleProvider<OpenRouterConf
     constructor(client: OpenAI, config?: OpenRouterConfig);
     getModelSpecs(): ModelSpec[];
     /**
-     * Override buildApiRequest to inject reasoning parameters.
-     * OpenRouter normalizes reasoning into the standard OpenAI format.
+     * Override buildApiRequest to inject reasoning parameters and cache_control breakpoints.
+     * OpenRouter normalizes reasoning into the standard OpenAI format,
+     * and supports cache_control on message content blocks for both
+     * Anthropic Claude and Google Gemini models.
      */
     protected buildApiRequest(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): Parameters<OpenAI["chat"]["completions"]["create"]>[0];
+    /** Minimal shape for messages in the already-built OpenAI-compatible request. */
+    private static readonly CACHE_CONTROL;
+    /**
+     * Add cache_control breakpoints to the last system message and last user message.
+     * This enables OpenRouter's prompt caching for supported providers (Anthropic, Gemini).
+     *
+     * Operates on the already-built request object. We cast through `unknown` because
+     * OpenAI's `ChatCompletionMessageParam` union is too narrow to assign content arrays
+     * with the non-standard `cache_control` property.
+     */
+    private injectCacheBreakpoints;
+    /**
+     * Return a new content array with cache_control on the last block.
+     * String content is promoted to a single-element text block array.
+     */
+    private withCacheControl;
     /**
      * Get custom headers for OpenRouter analytics.
      */

package/dist/index.js

@@ -9942,7 +9942,7 @@ var init_openai_compatible_provider = __esm({
             inputTokens: chunk.usage.prompt_tokens,
             outputTokens: chunk.usage.completion_tokens,
             totalTokens: chunk.usage.total_tokens,
-            cachedInputTokens: 0,
+            cachedInputTokens: usageDetails?.prompt_tokens_details?.cached_tokens ?? 0,
             reasoningTokens: usageDetails?.completion_tokens_details?.reasoning_tokens
           } : void 0;
           if (finishReason || usage) {
@@ -11923,7 +11923,7 @@ var init_openrouter = __esm({
       high: "high",
       maximum: "xhigh"
     };
-    OpenRouterProvider = class extends OpenAICompatibleProvider {
+    OpenRouterProvider = class _OpenRouterProvider extends OpenAICompatibleProvider {
       providerId = "openrouter";
       providerAlias = "or";
       constructor(client, config = {}) {
@@ -11933,8 +11933,10 @@ var init_openrouter = __esm({
         return OPENROUTER_MODELS;
       }
       /**
-       * Override buildApiRequest to inject reasoning parameters.
-       * OpenRouter normalizes reasoning into the standard OpenAI format.
+       * Override buildApiRequest to inject reasoning parameters and cache_control breakpoints.
+       * OpenRouter normalizes reasoning into the standard OpenAI format,
+       * and supports cache_control on message content blocks for both
+       * Anthropic Claude and Google Gemini models.
        */
       buildApiRequest(options, descriptor, spec, messages) {
         const request = super.buildApiRequest(options, descriptor, spec, messages);
@@ -11944,8 +11946,49 @@ var init_openrouter = __esm({
             effort: OPENROUTER_EFFORT_MAP[options.reasoning.effort ?? "medium"]
           };
         }
+        const cachingEnabled = options.caching?.enabled !== false;
+        if (cachingEnabled) {
+          this.injectCacheBreakpoints(request);
+        }
         return request;
       }
+      /** Minimal shape for messages in the already-built OpenAI-compatible request. */
+      static CACHE_CONTROL = { type: "ephemeral" };
+      /**
+       * Add cache_control breakpoints to the last system message and last user message.
+       * This enables OpenRouter's prompt caching for supported providers (Anthropic, Gemini).
+       *
+       * Operates on the already-built request object. We cast through `unknown` because
+       * OpenAI's `ChatCompletionMessageParam` union is too narrow to assign content arrays
+       * with the non-standard `cache_control` property.
+       */
+      injectCacheBreakpoints(request) {
+        const msgs = request.messages;
+        let lastSystemIdx = -1;
+        let lastUserIdx = -1;
+        for (let i = 0; i < msgs.length; i++) {
+          if (msgs[i].role === "system") lastSystemIdx = i;
+          if (msgs[i].role === "user") lastUserIdx = i;
+        }
+        if (lastSystemIdx >= 0) {
+          msgs[lastSystemIdx].content = this.withCacheControl(msgs[lastSystemIdx].content);
+        }
+        if (lastUserIdx >= 0) {
+          msgs[lastUserIdx].content = this.withCacheControl(msgs[lastUserIdx].content);
+        }
+      }
+      /**
+       * Return a new content array with cache_control on the last block.
+       * String content is promoted to a single-element text block array.
+       */
+      withCacheControl(content) {
+        if (typeof content === "string") {
+          return [{ type: "text", text: content, cache_control: _OpenRouterProvider.CACHE_CONTROL }];
+        }
+        return content.map(
+          (block, i) => i === content.length - 1 ? { ...block, cache_control: _OpenRouterProvider.CACHE_CONTROL } : block
+        );
+      }
       /**
        * Get custom headers for OpenRouter analytics.
        */