fix: stuck greeting ROLLBACK for tiny models overwhelmed by tool injection context

Author: Brendan Gray

main/agenticChat.js

@@ -1090,6 +1090,7 @@ function register(ctx) {
       let sessionJustRotated = false; // Flag to rebuild prompt after rotation
       let overflowResponseBudgetReduced = false; // Flag: already tried reducing response budget on first-turn overflow
       let forcedToolFunctions = null; // Set by PILLAR 3 refusal recovery to force grammar on next iteration
+      let forceChatModeForRetry = false; // Set on stuck_greeting ROLLBACK — strips tools and uses chat preamble for retry
       let consecutiveEmptyGrammarRetries = 0; // Track grammar failures for text-mode fallback
 
       // ── Execution State Tracking (ported from Pocket Guide) ──
@@ -1384,7 +1385,12 @@ function register(ctx) {
           : modelTier.tier === 'large' ? 5 : 2;
         const useNativeFunctions = (taskType !== 'chat') && iteration <= grammarIterLimit;
         let nativeFunctions = null;
-        if (consecutiveEmptyGrammarRetries >= 1) {
+        if (forceChatModeForRetry) {
+          // stuck_greeting recovery: strip all tools, use plain text with chat preamble
+          nativeFunctions = null;
+          forcedToolFunctions = null;
+          forceChatModeForRetry = false; // One-shot
+        } else if (consecutiveEmptyGrammarRetries >= 1) {
           // Grammar-to-text fallback: model can't produce grammar output, degrade gracefully.
           // Threshold lowered to 1 — the second native function call attempt can hang at the
           // C++ level and never return. One failure is enough to switch to text mode safely.
@@ -1703,7 +1709,16 @@ function register(ctx) {
           }
 
           // Escalating retry strategy
-          if (rollbackRetries === 1) {
+          if (responseVerdict.reason === 'stuck_greeting') {
+            // Model outputted its trained greeting despite an agentic task.
+            // Retry with the simple chat preamble and NO tools — removes all tool pressure.
+            currentPrompt = {
+              systemContext: buildStaticPrompt('chat'),
+              userMessage: message.substring(0, 500),
+            };
+            forceChatModeForRetry = true;
+            console.log('[AI Chat] Stuck greeting detected — retrying in chat mode (tool-free)');
+          } else if (rollbackRetries === 1) {
             // First retry: same prompt, slightly lower temperature for focus
             if (context?.params) context.params.temperature = Math.max((context.params.temperature || 0.7) - 0.2, 0.1);
           } else if (rollbackRetries === 2) {

main/agenticChatHelpers.js

@@ -563,6 +563,14 @@ function evaluateResponse(responseText, functionCalls, taskType, iteration, hasR
     return { verdict: 'ROLLBACK', reason: 'described_not_executed' };
   }
 
+  // ── STUCK GREETING: generic greeting response on an agentic (non-chat) task ──
+  // Tiny models overwhelmed by tool injection context fall back to their trained greeting.
+  // Detect the exact pattern and ROLLBACK so the retry can strip tool pressure.
+  const STUCK_GREETING_RX = /^(Hello!?\s+How\s+can\s+I\s+(assist|help)\s+(you\s+)?(today)?[!?.]*\s*|Hi!?\s+How\s+can\s+I\s+(assist|help)[!?.]*\s*)$/i;
+  if (taskType !== 'chat' && iteration === 1 && STUCK_GREETING_RX.test(text)) {
+    return { verdict: 'ROLLBACK', reason: 'stuck_greeting' };
+  }
+
   // ── DEFAULT: Accept ──
   return { verdict: 'COMMIT', reason: 'default' };
 }