From 63e14677f618255103e740ab4ca88b61e8e3e1e2 Mon Sep 17 00:00:00 2001
From: GyeongChan Jang <cham9994@gmail.com>
Date: Mon, 13 Jul 2026 10:13:32 +0900
Subject: [PATCH] fix(stream): emit a single finish_reason for tool-call chat
 completions

sse_translate_chat emitted finish_reason "tool_calls" for each completed
function_call item and then an additional finish_reason "stop" chunk from
the response.completed handler. Per the OpenAI streaming format a choice
carries exactly one finish_reason; clients that honor the last one classify
a mixed text + tool_call response as a plain text turn and silently drop
the accumulated tool calls.

Mark the stop chunk as already sent when the tool_calls finish chunk goes
out, so response.completed no longer appends the spurious "stop".

Adds a streaming regression test that fails without the fix.
---
 chatmock/utils.py    |  2 ++
 tests/test_routes.py | 61 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 63 insertions(+)

diff --git a/chatmock/utils.py b/chatmock/utils.py
index c04b997..0d180a0 100644
--- a/chatmock/utils.py
+++ b/chatmock/utils.py
@@ -599,6 +599,7 @@ def _merge_from(src):
                             ],
                         }
                         yield f"data: {json.dumps(finish_chunk)}\n\n".encode("utf-8")
+                        sent_stop_chunk = True
                 except Exception:
                     pass
 
@@ -682,6 +683,7 @@ def _merge_from(src):
                             "choices": [{"index": 0, "delta": {}, "finish_reason": "tool_calls"}],
                         }
                         yield f"data: {json.dumps(finish_chunk)}\n\n".encode("utf-8")
+                        sent_stop_chunk = True
             elif kind == "response.reasoning_summary_part.added":
                 if compat in ("think-tags", "o3"):
                     if saw_any_summary:
diff --git a/tests/test_routes.py b/tests/test_routes.py
index acf135d..5fb3c9a 100644
--- a/tests/test_routes.py
+++ b/tests/test_routes.py
@@ -97,6 +97,67 @@ def test_chat_completions(self, mock_start) -> None:
         self.assertEqual(body["choices"][0]["message"]["content"], "hello")
         self.assertEqual(body["model"], "gpt5.4-mini")
 
+    @patch("chatmock.routes_openai.start_upstream_request")
+    def test_chat_completions_stream_single_finish_reason_with_tool_call(self, mock_start) -> None:
+        """A streamed response that mixes assistant text with a function call must
+        emit exactly one finish_reason ("tool_calls"), not a trailing "stop" chunk.
+
+        Clients that honor the last finish_reason otherwise treat the response as a
+        plain text turn and drop the accumulated tool calls entirely.
+        """
+        mock_start.return_value = (
+            FakeUpstream(
+                [
+                    {"type": "response.output_text.delta", "delta": "Let me look that up."},
+                    {
+                        "type": "response.output_item.done",
+                        "item": {
+                            "type": "function_call",
+                            "call_id": "call_1",
+                            "name": "get_weather",
+                            "arguments": "{\"city\": \"Seoul\"}",
+                        },
+                    },
+                    {"type": "response.completed", "response": {"id": "resp-tool-stream"}},
+                ]
+            ),
+            None,
+        )
+        response = self.client.post(
+            "/v1/chat/completions",
+            json={
+                "model": "gpt-5.4",
+                "messages": [{"role": "user", "content": "weather in seoul?"}],
+                "stream": True,
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "get_weather",
+                            "parameters": {
+                                "type": "object",
+                                "properties": {"city": {"type": "string"}},
+                            },
+                        },
+                    }
+                ],
+            },
+        )
+        self.assertEqual(response.status_code, 200)
+        finish_reasons: list[str] = []
+        saw_tool_call_delta = False
+        for line in response.get_data(as_text=True).splitlines():
+            if not line.startswith("data: ") or line == "data: [DONE]":
+                continue
+            chunk = json.loads(line[len("data: ") :])
+            for choice in chunk.get("choices", []):
+                if choice.get("finish_reason"):
+                    finish_reasons.append(choice["finish_reason"])
+                if (choice.get("delta") or {}).get("tool_calls"):
+                    saw_tool_call_delta = True
+        self.assertTrue(saw_tool_call_delta)
+        self.assertEqual(finish_reasons, ["tool_calls"])
+
     @patch("chatmock.routes_openai.start_upstream_request")
     def test_chat_completions_honors_debug_model_override(self, mock_start) -> None:
         app = create_app(debug_model="gpt-5.4")