@compilr-dev/agents 0.3.27 → 0.3.28

package/dist/providers/claude.d.ts

@@ -107,6 +107,18 @@ export declare class ClaudeProvider implements LLMProvider {
      * reducing token costs by up to 90% on subsequent requests.
      */
     private wrapSystemPromptWithCache;
+    /**
+     * Add cache_control breakpoint to conversation messages.
+     *
+     * Caches the conversation history prefix (all messages except the most recent turn).
+     * This avoids re-processing the entire history on each API call.
+     * Only applies when there are enough messages to benefit (>= 4 messages = 2+ turns).
+     *
+     * Strategy: place cache_control on the last content block of the second-to-last
+     * user message. This caches system + tools + all messages up to that point.
+     * Only the most recent user message is uncached (and processed at full cost).
+     */
+    private addCacheControlToMessages;
     /**
      * Add cache_control to the last tool definition.
      *

package/dist/providers/claude.js

@@ -66,7 +66,9 @@ export class ClaudeProvider {
                 model: options?.model ?? this.defaultModel,
                 max_tokens: options?.maxTokens ?? this.defaultMaxTokens,
                 system: shouldCache && systemPrompt ? this.wrapSystemPromptWithCache(systemPrompt) : systemPrompt,
-                messages: anthropicMessages,
+                messages: shouldCache
+                    ? this.addCacheControlToMessages(anthropicMessages)
+                    : anthropicMessages,
                 tools: tools.length > 0
                     ? shouldCache
                         ? this.addCacheControlToLastTool(tools)
@@ -255,6 +257,67 @@ export class ClaudeProvider {
             },
         ];
     }
+    /**
+     * Add cache_control breakpoint to conversation messages.
+     *
+     * Caches the conversation history prefix (all messages except the most recent turn).
+     * This avoids re-processing the entire history on each API call.
+     * Only applies when there are enough messages to benefit (>= 4 messages = 2+ turns).
+     *
+     * Strategy: place cache_control on the last content block of the second-to-last
+     * user message. This caches system + tools + all messages up to that point.
+     * Only the most recent user message is uncached (and processed at full cost).
+     */
+    addCacheControlToMessages(messages) {
+        // Need at least 4 messages (2 turns) to benefit from caching
+        if (messages.length < 4)
+            return messages;
+        // Find the second-to-last user message
+        let targetIndex = -1;
+        let userCount = 0;
+        for (let i = messages.length - 1; i >= 0; i--) {
+            if (messages[i].role === 'user') {
+                userCount++;
+                if (userCount === 2) {
+                    targetIndex = i;
+                    break;
+                }
+            }
+        }
+        if (targetIndex < 0)
+            return messages;
+        // Clone messages to avoid mutating originals
+        const result = messages.map((msg, idx) => {
+            if (idx !== targetIndex)
+                return msg;
+            // Add cache_control to the last content block of this message
+            const content = msg.content;
+            if (typeof content === 'string') {
+                return {
+                    ...msg,
+                    content: [
+                        {
+                            type: 'text',
+                            text: content,
+                            cache_control: { type: 'ephemeral' },
+                        },
+                    ],
+                };
+            }
+            if (Array.isArray(content) && content.length > 0) {
+                const lastBlock = content[content.length - 1];
+                return {
+                    ...msg,
+                    content: [
+                        ...content.slice(0, -1),
+                        { ...lastBlock, cache_control: { type: 'ephemeral' } },
+                    ],
+                };
+            }
+            return msg;
+        });
+        return result;
+    }
     /**
      * Add cache_control to the last tool definition.
      *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@compilr-dev/agents",
-  "version": "0.3.27",
+  "version": "0.3.28",
   "description": "Lightweight multi-LLM agent library for building CLI AI assistants",
   "type": "module",
   "main": "dist/index.js",