fix: issue properly basing diffs when tagged on pr

agents/ollama.ts

@@ -169,17 +169,18 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
       });
     }
 
-    // After select_mode, explicitly tell the model to act on the returned guidance.
-    // Without this nudge, smaller models tend to treat the mode instructions as
-    // informational and stop rather than executing the workflow steps.
-    if (calledSelectMode) {
+    // After the FIRST select_mode call, nudge the model to act on the guidance.
+    // Only nudge once — repeated nudging causes a loop where the model keeps
+    // re-calling select_mode instead of executing the workflow.
+    if (calledSelectMode && !pendingModeNudge) {
       pendingModeNudge = true;
       messages.push({
         role: "user",
         content:
-          "You have selected a mode and received your workflow instructions. " +
-          "Now execute the first step of that workflow immediately by calling the appropriate tool. " +
-          "Do not describe what you will do — just call the tool.",
+          "Good. You have selected a mode and received the workflow. " +
+          "Do NOT call select_mode again. " +
+          "Call the next tool to execute the workflow now (e.g. checkout_pr, get_issue, create_pull_request_review). " +
+          "Start immediately.",
       });
     }
   }