fix(governor): classify multi-word greetings ('hey there!', 'good morning') and short verb-less messages as conversational so they don't trigger the agent loop

app/api/chat/route.ts

@@ -113,14 +113,26 @@ function classifyTurnIntent(message: string): TurnIntent {
   )
     return "status_check";
 
-  // Conversational fallback — ONLY when the entire message is a greeting or a
-  // bare acknowledgement. Previously `/^(ok|...)/` matched "ok" as a prefix of
-  // "okay <real request>", misclassifying real work as chat (budget 1) and
-  // causing round_cap cut-offs. Require the whole message to be the ack.
+  // Conversational: greetings, acknowledgements, and short pleasantries.
+  // By this point we've already matched action verbs, diagnostics, and
+  // questions, so anything that *starts* with a greeting/ack, or is a very
+  // short verb-less leftover, is chat — reply in text, don't run the agent
+  // loop. Note we still avoid the old `/^(ok)/` prefix bug: "okay is the auth
+  // connected up?" already returned `status_check` above (ends with "?").
+  const wordCount = m.split(/\s+/).filter(Boolean).length;
   if (
-    /^(hi|hey|hello|yo|thanks|thank you|ok|okay|kk|k|yes|yep|yeah|yup|no|nope|nah|sure|cool|nice|great|awesome|perfect)[\s!.?]*$/.test(
+    // Greetings (incl. multi-word: "hey there!", "good morning")
+    /^(hi|hey|hello|heya|hiya|yo|sup|howdy|greetings|good (morning|afternoon|evening|day))\b/.test(
       m,
-    )
+    ) ||
+    // Acknowledgements / pleasantries
+    /^(thanks|thank you|ty|ok|okay|kk|k|yes|yep|yeah|yup|no|nope|nah|sure|cool|nice|great|awesome|perfect|got it|sounds good|will do|nvm|never mind)\b/.test(
+      m,
+    ) ||
+    /(what'?s up|how are you|how'?s it going|good to see you)/.test(m) ||
+    // Very short, verb-less, non-question leftovers are ambiguous → treat as
+    // chat and let the model ask, rather than spinning up a 40-round build.
+    (wordCount <= 4 && m.length <= 30 && !m.endsWith("?"))
   )
     return "conversational";