From 263763e15a7bfc512789c05333c9aaa95c1cb3c8 Mon Sep 17 00:00:00 2001
From: Susheel Daswani <Sdaswani@cs.stanford.edu>
Date: Tue, 28 Jul 2026 21:58:46 -0700
Subject: [PATCH] fix: stop Thinking chip from trapping answers in reasoning
 middleware

- model.startWithReasoning means show/enable thinking; middleware
  startWithReasoning means treat content as reasoning until </think>
- Ollama thinking models emit a native reasoning field and plain answer
  content with no closing tag, so wiring the chip flag through left the
  spinner stuck with the reply in the muted Thinking pane
- show a Thinking label while streaming with no active tool

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 src/ai/fetch.ts                               | 11 +++++-
 src/components/chat/reasoning-group-title.tsx | 36 +++++++++----------
 src/components/chat/reasoning-group.test.tsx  |  3 ++
 3 files changed, 31 insertions(+), 19 deletions(-)

diff --git a/src/ai/fetch.ts b/src/ai/fetch.ts
index 2453ab56b..c5a4044d7 100644
--- a/src/ai/fetch.ts
+++ b/src/ai/fetch.ts
@@ -708,9 +708,18 @@ export const aiFetchStreamingResponse = async ({
       providerId: model.provider,
       model: baseModel,
       middleware: [
+        // Extract `<think>...</think>` from content when present. Do NOT pass
+        // `startWithReasoning` from `model.startWithReasoning` — that DB flag
+        // means “model can think / show Thinking chip”, while the middleware
+        // option means “treat all content as reasoning until </think>”.
+        // Ollama thinking models emit a native `reasoning` field and plain
+        // answer `content` with no closing tag; wiring the chip flag through
+        // here classifies the answer as reasoning forever and never emits
+        // reasoning-end (stuck spinner with the reply stuck in the muted
+        // Thinking pane).
         extractReasoningMiddleware({
           tagName: 'think',
-          startWithReasoning: Boolean(model.startWithReasoning),
+          startWithReasoning: false,
         }),
       ],
     })
diff --git a/src/components/chat/reasoning-group-title.tsx b/src/components/chat/reasoning-group-title.tsx
index 5372f5e19..886d23c12 100644
--- a/src/components/chat/reasoning-group-title.tsx
+++ b/src/components/chat/reasoning-group-title.tsx
@@ -34,30 +34,30 @@ const activeToolLabel = (tool: ToolOrDynamicToolUIPart, mcpTools?: UIMessageMeta
 export const ReasoningGroupTitle = ({ totalDuration, isGroupReasoning, tools, mcpTools }: ReasoningGroupTitleProps) => {
   const activeIndex = tools.length - 1
   const activeTool = tools[activeIndex]
-  const loadingLabel = activeTool ? activeToolLabel(activeTool, mcpTools) : null
+  // While streaming with no tool yet, still show a label — an empty title next
+  // to the spinner reads as a hung chat (common for thinking-only turns).
+  const loadingLabel = activeTool ? activeToolLabel(activeTool, mcpTools) : 'Thinking'
 
   return (
     <div className="relative">
       <AnimatePresence mode="wait">
         {isGroupReasoning ? (
-          loadingLabel ? (
-            <m.div
-              key={`tool-${activeIndex}`}
-              // Skip entrance animation for tools already in progress (e.g., when switching back to a chat with active streaming)
-              initial={activeTool?.state === 'input-streaming' ? { y: 20, opacity: 0 } : { y: 0, opacity: 1 }}
-              animate={{ y: 0, opacity: 1, scale: 1 }}
-              exit={{ y: -20, opacity: 0 }}
-              transition={{ duration: 0.3, ease: [0.4, 0, 0.2, 1] }}
-              className="w-full"
+          <m.div
+            key={activeTool ? `tool-${activeIndex}` : 'thinking'}
+            // Skip entrance animation for tools already in progress (e.g., when switching back to a chat with active streaming)
+            initial={activeTool?.state === 'input-streaming' ? { y: 20, opacity: 0 } : { y: 0, opacity: 1 }}
+            animate={{ y: 0, opacity: 1, scale: 1 }}
+            exit={{ y: -20, opacity: 0 }}
+            transition={{ duration: 0.3, ease: [0.4, 0, 0.2, 1] }}
+            className="w-full"
+          >
+            <span
+              data-testid="tool-status"
+              className="text-xs text-muted-foreground italic animate-pulse truncate min-w-0"
             >
-              <span
-                data-testid="tool-status"
-                className="text-xs text-muted-foreground italic animate-pulse truncate min-w-0"
-              >
-                {loadingLabel}
-              </span>
-            </m.div>
-          ) : null
+              {loadingLabel}
+            </span>
+          </m.div>
         ) : (
           <m.div
             key="completed"
diff --git a/src/components/chat/reasoning-group.test.tsx b/src/components/chat/reasoning-group.test.tsx
index e1726326e..a05b65f13 100644
--- a/src/components/chat/reasoning-group.test.tsx
+++ b/src/components/chat/reasoning-group.test.tsx
@@ -254,6 +254,9 @@ describe('ReasoningGroup', () => {
       // ReasoningDisplay should render the reasoning text when streaming
       // It might be in a specific container, so check the container
       expect(container.textContent).toContain('Let me think about this...')
+      // Thinking-only turns must keep a visible status label (empty title + spinner
+      // reads as a hung chat).
+      expect(screen.getByTestId('tool-status')).toHaveTextContent('Thinking')
     })
 
     it('should not render ReasoningDisplay when hasTextPart is true', () => {