From db1559f71b8278bfd122cd6d58eafdad5a794620 Mon Sep 17 00:00:00 2001
From: euvre <93761161+euvre@users.noreply.github.com>
Date: Thu, 23 Jul 2026 09:58:24 +0800
Subject: [PATCH] fix: prevent thinking content from leaking into visible
 answer in shared chat (Go mode) (#17225)

---
 internal/service/bot_completion.go | 19 ++++++++++------
 internal/service/think_tag.go      |  7 ++++++
 internal/service/think_tag_test.go | 36 ++++++++++++++++++++++++++++++
 3 files changed, 55 insertions(+), 7 deletions(-)

diff --git a/internal/service/bot_completion.go b/internal/service/bot_completion.go
index 6bd066f43c..c181dc5bb9 100644
--- a/internal/service/bot_completion.go
+++ b/internal/service/bot_completion.go
@@ -429,11 +429,14 @@ func (s *BotService) ChatbotCompletion(
 	}
 
 	// 5. Translate pipeline results into chatbot frames. The
-	// pipeline streams deltas; the python iframe contract sends the
-	// full accumulated answer in every frame, so we accumulate here
-	// (the front-end accepts both shapes, but full-text frames are
-	// byte-parity with python). The final pipeline result carries
-	// the decorated full answer plus the retrieval reference.
+	// pipeline streams deltas; the python iframe contract also yields
+	// deltas (async_chat yields the value from _stream_with_think_delta).
+	// We accumulate the full answer locally for persistence and for the
+	// final frame, but forward each result's own delta in the SSE frame
+	// so the front-end can build the <think>-wrapped answer correctly.
+	// Sending accumulated full text on every frame interacts badly with
+	// the front-end's start_to_think/end_to_think marker append, causing
+	// reasoning content to leak into the visible answer.
 	out := make(chan ChatbotSSEFrame, 16)
 	go func() {
 		defer close(out)
@@ -468,8 +471,10 @@ func (s *BotService) ChatbotCompletion(
 				continue
 			}
 			if res.StartToThink || res.EndToThink {
+				// Marker frames carry no text; the front-end appends
+				// <think> / </think> to the accumulated answer.
 				out <- ChatbotSSEFrame{
-					Data:         fullAnswer,
+					Data:         "",
 					Reference:    map[string]any{},
 					SessionID:    session.ID,
 					StartToThink: res.StartToThink,
@@ -489,7 +494,7 @@ func (s *BotService) ChatbotCompletion(
 				finalRef = res.Reference
 			}
 			out <- ChatbotSSEFrame{
-				Data:      fullAnswer,
+				Data:      res.Answer,
 				Reference: map[string]any{},
 				SessionID: session.ID,
 			}
diff --git a/internal/service/think_tag.go b/internal/service/think_tag.go
index db8736a793..f0fdb71466 100644
--- a/internal/service/think_tag.go
+++ b/internal/service/think_tag.go
@@ -53,22 +53,29 @@ type ThinkStreamState struct {
 }
 
 // EnterReasoning marks the start of a reasoning block (model-level, not tag-based).
+// It also writes a <think> marker into fullText so that ExtractVisibleAnswer can
+// strip the reasoning segment from the final answer when the model exposes
+// reasoning through a separate channel (e.g. delta.reasoning_content).
 // Returns true when this is a new transition (reasoning was not already active).
 func (s *ThinkStreamState) EnterReasoning() bool {
 	if s.inReasoning {
 		return false
 	}
 	s.inReasoning = true
+	s.fullText += "<think>"
 	return true
 }
 
 // ExitReasoning marks the end of a reasoning block (model-level, not tag-based).
+// It writes a matching </think> marker into fullText so the reasoning segment is
+// closed before the visible answer text begins.
 // Returns true when this is a new transition (reasoning was active).
 func (s *ThinkStreamState) ExitReasoning() bool {
 	if !s.inReasoning {
 		return false
 	}
 	s.inReasoning = false
+	s.fullText += "</think>"
 	return true
 }
 
diff --git a/internal/service/think_tag_test.go b/internal/service/think_tag_test.go
index e87008b58a..ff85c4c563 100644
--- a/internal/service/think_tag_test.go
+++ b/internal/service/think_tag_test.go
@@ -301,3 +301,39 @@ func TestStreamThinkTagDelta_DeferredClose(t *testing.T) {
 		t.Errorf("second marker = %q", markers[1])
 	}
 }
+
+// TestReasoningChannelWrapsThinkTags verifies that model-level reasoning
+// (arriving through a separate reason channel rather than <think> tags in
+// the text) is bracketed by <think></think> in fullText. This prevents the
+// reasoning content from leaking into the final visible answer.
+func TestReasoningChannelWrapsThinkTags(t *testing.T) {
+	state := &ThinkStreamState{}
+
+	if !state.EnterReasoning() {
+		t.Fatal("EnterReasoning should return true on first call")
+	}
+	// Simulate the chat pipeline feeding reasoning chunks through NextThinkDelta.
+	deltas := NextThinkDelta(state, "reasoning step 1", 0)
+	if len(deltas) != 1 || deltas[0].Value != "reasoning step 1" {
+		t.Fatalf("unexpected reasoning deltas: %+v", deltas)
+	}
+	deltas = NextThinkDelta(state, " reasoning step 2", 0)
+	if len(deltas) != 1 || deltas[0].Value != " reasoning step 2" {
+		t.Fatalf("unexpected reasoning deltas: %+v", deltas)
+	}
+
+	if state.ExitReasoning() != true {
+		t.Fatal("ExitReasoning should return true when reasoning was active")
+	}
+
+	// Visible answer arrives after reasoning ends.
+	deltas = BufferAnswerDelta(state, "visible answer", 0)
+	if len(deltas) != 1 || deltas[0].Value != "visible answer" {
+		t.Fatalf("unexpected answer deltas: %+v", deltas)
+	}
+
+	wantVisible := "visible answer"
+	if got := ExtractVisibleAnswer(state.fullText); got != wantVisible {
+		t.Errorf("ExtractVisibleAnswer(%q) = %q, want %q", state.fullText, got, wantVisible)
+	}
+}