gitlab-ai-provider 6.5.0 → 6.6.0

package/CHANGELOG.md

@@ -2,6 +2,11 @@
 
 All notable changes to this project will be documented in this file. See [Conventional Commits](https://conventionalcommits.org) for commit guidelines.
 
+## 6.6.0 (2026-04-17)
+
+- Merge branch 'feat/anthropic-prompt-caching' into 'main' ([d2125eb](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/d2125eb)), closes [#6](https://gitlab.com/vglafirov/gitlab-ai-provider/issues/6)
+- feat(anthropic): enable prompt caching via cache_control on system prompt ([376556f](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/376556f))
+
 ## 6.5.0 (2026-04-17)
 
 - Merge branch 'add-opus-4-7-model-mapping' into 'main' ([f95e21a](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/f95e21a))

package/dist/index.d.mts

@@ -79,7 +79,17 @@ declare class GitLabAnthropicLanguageModel implements LanguageModelV3 {
      */
     private convertToolChoice;
     /**
-     * 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
      */
     private convertPrompt;
     /**

package/dist/index.d.ts

@@ -79,7 +79,17 @@ declare class GitLabAnthropicLanguageModel implements LanguageModelV3 {
      */
     private convertToolChoice;
     /**
-     * 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
      */
     private convertPrompt;
     /**

package/dist/index.js

@@ -353,7 +353,17 @@ 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;
@@ -425,7 +435,23 @@ ${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
@@ -1654,7 +1680,7 @@ var GitLabOpenAILanguageModel = class {
 var import_isomorphic_ws = __toESM(require("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) => {