bugfix: improve qwen25 tool-call parsing robustness.

Author: yingxudeng

xllm/api_service/utils.h

@@ -98,16 +98,22 @@ inline ToolCallResult process_tool_calls(
     return result;
   }
 
-  if (finish_reason == "stop") {
-    result.finish_reason = "tool_calls";
-  } else {
-    result.finish_reason = std::move(finish_reason);
-  }
-
   try {
     auto [parsed_text, call_info_list] = parser.parse_non_stream(text);
     result.text = std::move(parsed_text);
 
+    if (call_info_list.empty()) {
+      result.text = std::move(text);
+      result.finish_reason = std::move(finish_reason);
+      return result;
+    }
+
+    if (finish_reason == "stop") {
+      result.finish_reason = "tool_calls";
+    } else {
+      result.finish_reason = std::move(finish_reason);
+    }
+
     google::protobuf::RepeatedPtrField<proto::ToolCall> tool_calls;
 
     for (const auto& call_info : call_info_list) {
@@ -154,4 +160,4 @@ inline nlohmann::json struct_to_json(
 }
 
 }  // namespace api_service
-}  // namespace xllm
+}  // namespace xllm

xllm/function_call/qwen25_detector.cpp

@@ -110,7 +110,9 @@ StreamingParseResult Qwen25Detector::detect_and_parse(
 
     try {
       std::string json_content(trimmed_content);
-      auto json_obj = nlohmann::json::parse(json_content);
+      auto [json_obj, consumed_len] =
+          partial_json_loads(json_content, Allow::ALL);
+      (void)consumed_len;
       auto parsed_calls = parse_base_json(json_obj, tools);
 
       calls.insert(calls.end(),

xllm/function_call/qwen25_detector_test.cpp

@@ -353,6 +353,39 @@ TEST_F(Qwen25DetectorTest, PerformanceWithManyToolCalls) {
   }
 }
 
+// Regression test: malformed qwen3 payload in message.content should still be
+// recovered into structured tool_calls for non-stream output.
+TEST_F(Qwen25StreamingTest, NonStreamMalformedToolPayloadRecoversToolCall) {
+  nlohmann::json search_recall_params = {
+      {"type", "object"},
+      {"properties",
+       {{"attribute_selection", {{"type", "object"}}},
+        {"query_list", {{"type", "array"}}},
+        {"search_mode", {{"type", "string"}}}}},
+      {"required", {"attribute_selection", "query_list", "search_mode"}}};
+  JsonTool search_recall_tool(
+      "function",
+      JsonFunction("SearchRecall", "SearchRecall tool", search_recall_params));
+
+  FunctionCallParser parser({search_recall_tool}, "qwen25");
+
+  // Real user sample: malformed JSON inside <tool_call> content.
+  std::string malformed_content =
+      "<tool_call>\n"
+      "{\"name\": \"SearchRecall\", \"arguments\": {\"attribute_selection\": "
+      "{\"color: [\"}, \"query_list\": [{\"query\": \"kids wiggle car\"}], "
+      "\"search_mode\": \"single\"}}\n"
+      "</tool_call>";
+
+  auto [normal_text, calls] = parser.parse_non_stream(malformed_content);
+
+  EXPECT_TRUE(normal_text.empty());
+  ASSERT_EQ(calls.size(), 1);
+  ASSERT_TRUE(calls[0].name.has_value());
+  EXPECT_EQ(calls[0].name.value(), "SearchRecall");
+  EXPECT_FALSE(calls[0].parameters.empty());
+}
+
 // Test basic streaming functionality
 TEST_F(Qwen25StreamingTest, BasicStreamingParsing) {
   FunctionCallParser parser(tools_, "qwen25");
@@ -541,5 +574,52 @@ TEST_F(Qwen25StreamingTest, PartialTokenHandling) {
   EXPECT_GT(accumulated_calls.size(), 0);
 }
 
+// Regression test: even with malformed payload from message.content, streaming
+// parser may still emit intermediate tool-call events (e.g. tool name token).
+TEST_F(Qwen25StreamingTest,
+       StreamingMalformedToolPayloadStillEmitsToolNameEvent) {
+  nlohmann::json search_recall_params = {
+      {"type", "object"},
+      {"properties",
+       {{"attribute_selection", {{"type", "object"}}},
+        {"query_list", {{"type", "array"}}},
+        {"search_mode", {{"type", "string"}}}}},
+      {"required", {"attribute_selection", "query_list", "search_mode"}}};
+  JsonTool search_recall_tool(
+      "function",
+      JsonFunction("SearchRecall", "SearchRecall tool", search_recall_params));
+
+  FunctionCallParser parser({search_recall_tool}, "qwen25");
+
+  std::vector<std::string> chunks = {
+      "<tool_call>\n",
+      "{\"name\": \"SearchRecall\", \"arguments\": {\"attribute_selection\": "
+      "{\"color: [\"}, ",
+      "\"query_list\": [{\"query\": \"kids wiggle car\"}], \"search_mode\": "
+      "\"single\"}}\n",
+      "</tool_call>"};
+
+  std::string streamed_normal_text;
+  std::vector<ToolCallItem> streamed_calls;
+  for (const auto& chunk : chunks) {
+    auto result = parser.parse_streaming_increment(chunk);
+    streamed_normal_text += result.normal_text;
+    streamed_calls.insert(
+        streamed_calls.end(), result.calls.begin(), result.calls.end());
+  }
+
+  bool found_search_recall_name = false;
+  for (const auto& call : streamed_calls) {
+    if (call.name.has_value() && call.name.value() == "SearchRecall") {
+      found_search_recall_name = true;
+      break;
+    }
+  }
+
+  EXPECT_FALSE(streamed_calls.empty());
+  EXPECT_TRUE(found_search_recall_name);
+  EXPECT_TRUE(streamed_normal_text.empty());
+}
+
 }  // namespace function_call
-}  // namespace xllm
+}  // namespace xllm