feat: widgets and ai chat

apps/api/src/services/agent/context-window.ts

@@ -0,0 +1,27 @@
+/**
+ * Trims conversation message history to stay within token budget.
+ *
+ * Strategy:
+ * - Always keep the first message (often contains the initial request/document context)
+ * - Always keep the last N messages (recent context is most relevant)
+ * - Drop middle messages when the array is too long
+ *
+ * This is a message-count heuristic, not a token counter. Each message averages
+ * ~200-500 tokens, so capping at ~40 messages keeps us under ~20K tokens of
+ * history, leaving room for the system prompt + document text + LLM response.
+ */
+
+const MAX_MESSAGES = 40;
+const KEEP_RECENT = 30;
+
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+export function trimMessageHistory(messages: any[]): any[] {
+  if (messages.length <= MAX_MESSAGES) {
+    return messages;
+  }
+
+  const firstMessage = messages[0];
+  const recentMessages = messages.slice(-KEEP_RECENT);
+
+  return [firstMessage, ...recentMessages];
+}