workspace out of cache; welcome.md

gateway/src/llm/middleware.ts

@@ -42,13 +42,14 @@ export class AnthropicCachingMiddleware implements ModelMiddleware {
       addCacheControl(systemMsg);
     }
 
-    // 2. Tag the last cacheable message that isn't the current user input.
-    // The current user message is always the last element; we want the one before it.
-    // We look backwards for the last AIMessage or HumanMessage (excluding the final message).
+    // 2. Tag the last AIMessage (assistant turn) before the current user input.
+    // Workspace state is injected as a HumanMessage just before the current user input;
+    // by matching only 'ai' we ensure the cache boundary is the last assistant response,
+    // leaving workspace context and the new user input uncached (always fresh).
     const candidates = result.slice(0, -1);
     for (let i = candidates.length - 1; i >= 0; i--) {
       const t = candidates[i]._getType();
-      if (t === 'ai' || t === 'human') {
+      if (t === 'ai') {
         addCacheControl(candidates[i]);
         break;
       }