@myvillage/cli 1.50.0 → 1.51.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@myvillage/cli",
-  "version": "1.50.0",
+  "version": "1.51.0",
   "description": "MyVillageOS CLI for community developers",
   "type": "module",
   "bin": {

package/src/agent-runtime/loop.js

@@ -245,16 +245,38 @@ Guidelines:
       // "find any README under the workspace" can do
       // list_allowed_directories → list_directory → search_files → ...
       // until it's actually done.
+      //
+      // Anthropic prompt caching: we mark the system block as ephemeral so
+      // everything before it (tool schemas + system prompt) is cached.
+      // For Anthropic, cached reads are ~10x cheaper than fresh input —
+      // critical because MCP tool schemas alone can be 20–40k tokens that
+      // would otherwise be re-billed at every step of every loop iteration.
+      // The breakpoint is namespaced under `anthropic`, so OpenAI calls
+      // silently ignore it. Cache TTL is ~5 min; agents that loop more
+      // often than that stay warm and pay full price only on the first hit.
       const result = await generateText({
         model,
-        system: systemPrompt,
-        prompt: context,
+        messages: [
+          {
+            role: 'system',
+            content: systemPrompt,
+            providerOptions: {
+              anthropic: { cacheControl: { type: 'ephemeral' } },
+            },
+          },
+          { role: 'user', content: context },
+        ],
         tools,
         stopWhen: stepCountIs(10),
         maxOutputTokens: maxTokens,
       });
 
-      // Log LLM response
+      // Log LLM response. Cache metrics come from the Anthropic provider
+      // metadata — surfacing them lets `agent logs` show whether caching is
+      // landing. `cacheReadInputTokens` should grow on the 2nd+ iteration in
+      // a 5-min window; if it stays 0, the prefix isn't matching (usually
+      // because something upstream of the breakpoint is varying per call).
+      const anthropicMeta = result.providerMetadata?.anthropic || {};
       logActivity(agentDir, {
         type: 'llm_response',
         text: (result.text || '').slice(0, 500),
@@ -262,6 +284,8 @@ Guidelines:
           prompt: result.usage?.inputTokens || 0,
           completion: result.usage?.outputTokens || 0,
           total: result.usage?.totalTokens || 0,
+          cacheCreate: anthropicMeta.cacheCreationInputTokens || 0,
+          cacheRead: anthropicMeta.cacheReadInputTokens || 0,
         },
       });