gitlab-ai-provider 6.12.0 → 6.12.1

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.
 
+## <small>6.12.1 (2026-07-29)</small>
+
+- Merge branch 'ghavenga-cache-breakpoint-placement' into 'main' ([e8d0fe5](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/e8d0fe5))
+- perf(anthropic): place cache breakpoints on final two messages ([5f2e13e](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/5f2e13e))
+
 ## 6.12.0 (2026-07-27)
 
 - Merge branch 'feature-add-opus-5' into 'main' ([9a5f447](https://gitlab.com/vglafirov/gitlab-ai-provider/commit/9a5f447))

package/dist/index.d.mts

@@ -88,11 +88,17 @@ declare class GitLabAnthropicLanguageModel implements LanguageModelV3 {
      *
      * 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.
+     * 2. The last content block of each of the final two messages.
      *
-     * This lets Anthropic cache the system prompt and the accumulated
-     * conversation prefix, so each new turn only pays for the new content.
+     * Two trailing breakpoints (rather than a single one on the penultimate
+     * message) keep a cache write within Anthropic's 20-block lookback window
+     * as an agentic conversation grows several messages per turn (assistant
+     * tool-call → tool-result → …). With a single breakpoint the most recent
+     * write can drift more than 20 blocks behind the current position, so the
+     * next request fails to prefix-match and pays for a fresh cache write
+     * instead of a cheap read. The extra breakpoint costs nothing (breakpoints
+     * themselves are free; you only pay for tokens actually written/read) and
+     * materially raises the cache hit rate in multi-turn tool-using sessions.
      *
      * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
      */

package/dist/index.d.ts

@@ -88,11 +88,17 @@ declare class GitLabAnthropicLanguageModel implements LanguageModelV3 {
      *
      * 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.
+     * 2. The last content block of each of the final two messages.
      *
-     * This lets Anthropic cache the system prompt and the accumulated
-     * conversation prefix, so each new turn only pays for the new content.
+     * Two trailing breakpoints (rather than a single one on the penultimate
+     * message) keep a cache write within Anthropic's 20-block lookback window
+     * as an agentic conversation grows several messages per turn (assistant
+     * tool-call → tool-result → …). With a single breakpoint the most recent
+     * write can drift more than 20 blocks behind the current position, so the
+     * next request fails to prefix-match and pays for a fresh cache write
+     * instead of a cheap read. The extra breakpoint costs nothing (breakpoints
+     * themselves are free; you only pay for tokens actually written/read) and
+     * materially raises the cache hit rate in multi-turn tool-using sessions.
      *
      * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
      */

package/dist/index.js

@@ -929,11 +929,17 @@ var GitLabAnthropicLanguageModel = class {
    *
    * 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.
+   * 2. The last content block of each of the final two messages.
    *
-   * This lets Anthropic cache the system prompt and the accumulated
-   * conversation prefix, so each new turn only pays for the new content.
+   * Two trailing breakpoints (rather than a single one on the penultimate
+   * message) keep a cache write within Anthropic's 20-block lookback window
+   * as an agentic conversation grows several messages per turn (assistant
+   * tool-call → tool-result → …). With a single breakpoint the most recent
+   * write can drift more than 20 blocks behind the current position, so the
+   * next request fails to prefix-match and pays for a fresh cache write
+   * instead of a cheap read. The extra breakpoint costs nothing (breakpoints
+   * themselves are free; you only pay for tokens actually written/read) and
+   * materially raises the cache hit rate in multi-turn tool-using sessions.
    *
    * @see https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
    */
@@ -1022,10 +1028,11 @@ ${message.content}` : message.content;
         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];
+    const breakpointCount = Math.min(2, messages.length);
+    for (let i = messages.length - breakpointCount; i < messages.length; i++) {
+      const message = messages[i];
+      if (Array.isArray(message.content) && message.content.length > 0) {
+        const lastBlock = message.content[message.content.length - 1];
         lastBlock.cache_control = {
           type: "ephemeral"
         };
@@ -2519,7 +2526,7 @@ var import_node_async_hooks = require("async_hooks");
 var import_isomorphic_ws = __toESM(require("isomorphic-ws"));
 
 // src/version.ts
-var VERSION = true ? "6.11.1" : "0.0.0-dev";
+var VERSION = true ? "6.12.0" : "0.0.0-dev";
 
 // src/gitlab-workflow-client.ts
 var WS_CONNECT_TIMEOUT_MS = 3e4;