gitlab-ai-provider 6.5.0 → 6.6.0

package/dist/index.mjs

@@ -282,7 +282,17 @@ var GitLabAnthropicLanguageModel = class {
     }
   }
   /**
-   * Convert AI SDK prompt to Anthropic messages format
+   * Convert AI SDK prompt to Anthropic messages format.
+   *
+   * Cache breakpoints (`cache_control: { type: "ephemeral" }`) are placed on:
+   * 1. The system prompt content block — static across all turns.
+   * 2. The last content block of the second-to-last message — the boundary
+   *    between conversation history and the current turn.
+   *
+   * This lets Anthropic cache the system prompt and the accumulated
+   * conversation prefix, so each new turn only pays for the new content.
+   *
+   * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
    */
   convertPrompt(prompt) {
     let systemMessage;
@@ -354,7 +364,23 @@ ${message.content}` : message.content;
         }
       }
     }
-    return { system: systemMessage, messages };
+    const system = systemMessage ? [
+      {
+        type: "text",
+        text: systemMessage,
+        cache_control: { type: "ephemeral" }
+      }
+    ] : void 0;
+    if (messages.length >= 2) {
+      const penultimate = messages[messages.length - 2];
+      if (Array.isArray(penultimate.content)) {
+        const lastBlock = penultimate.content[penultimate.content.length - 1];
+        lastBlock.cache_control = {
+          type: "ephemeral"
+        };
+      }
+    }
+    return { system, messages };
   }
   /**
    * Convert Anthropic finish reason to AI SDK format
@@ -1583,7 +1609,7 @@ var GitLabOpenAILanguageModel = class {
 import WebSocket from "isomorphic-ws";
 
 // src/version.ts
-var VERSION = true ? "6.4.2" : "0.0.0-dev";
+var VERSION = true ? "6.5.0" : "0.0.0-dev";
 
 // src/gitlab-workflow-types.ts
 var WorkflowType = /* @__PURE__ */ ((WorkflowType2) => {