Merge pull request #12021 from ollama/drifkin/thinking-double-emit

thinking: fix double emit when no opening tag

thinking/parser.go

@@ -103,7 +103,9 @@ func eat(s *Parser) (string, string, bool) {
 			// note that we use the original content, not the trimmed one because we
 			// don't want to eat any whitespace in the real content if there were no
 			// thinking tags
-			return "", s.acc.String(), false
+			untrimmed := s.acc.String()
+			s.acc.Reset()
+			return "", untrimmed, false
 		}
 	case thinkingState_ThinkingStartedEatingWhitespace:
 		trimmed := strings.TrimLeftFunc(s.acc.String(), unicode.IsSpace)

thinking/parser_test.go

@@ -58,6 +58,15 @@ func TestThinkingStreaming(t *testing.T) {
 					wantContent:    "  abc",
 					wantStateAfter: thinkingState_ThinkingDone,
 				},
+				// regression test for a bug where we were transitioning directly to
+				// ThinkingDone without clearing the buffer. This would cuase the first
+				// step to be outputted twice
+				{
+					input:          "def",
+					wantThinking:   "",
+					wantContent:    "def",
+					wantStateAfter: thinkingState_ThinkingDone,
+				},
 			},
 		},
 		{