From 6a57d3795ab435f29250576a7ea490c87745aca1 Mon Sep 17 00:00:00 2001
From: wasimysaid <wasimysdev@gmail.com>
Date: Fri, 19 Jun 2026 21:54:50 +0200
Subject: [PATCH] Avoid synthetic finish after passthrough errors

---
 studio/backend/routes/inference.py            | 16 +++++++-
 .../tests/test_openai_tool_passthrough.py     | 39 +++++++++++++++++++
 2 files changed, 53 insertions(+), 2 deletions(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index fb5426047b..365b1c3f3a 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -9640,6 +9640,7 @@ async def _openai_passthrough_stream(
             monitor_done = False
             saw_finish_reason = False
             saw_done = False
+            saw_stream_error = False
             saw_tool_call_delta = False
             last_chunk_id = completion_id
             last_chunk_model = model_name
@@ -9676,7 +9677,11 @@ async def _openai_passthrough_stream(
                     data_text = raw_line[6:].strip()
                     if data_text == "[DONE]":
                         saw_done = True
-                        if not saw_finish_reason and not cancel_event.is_set():
+                        if (
+                            not saw_finish_reason
+                            and not saw_stream_error
+                            and not cancel_event.is_set()
+                        ):
                             finish_line = _synthetic_finish_line()
                             _monitor_openai_sse_line(
                                 monitor_id,
@@ -9725,13 +9730,20 @@ async def _openai_passthrough_stream(
                         raw_line,
                         llama_backend.context_length,
                     )
+                    if monitor_event == "error":
+                        saw_stream_error = True
                     # Relay verbatim to preserve llama-server's native id,
                     # finish_reason, delta.tool_calls, and usage chunks.
                     yield raw_line + "\n\n"
                     if monitor_event == "done":
                         monitor_done = True
                         break
-                if not saw_done and not saw_finish_reason and not cancel_event.is_set():
+                if (
+                    not saw_done
+                    and not saw_finish_reason
+                    and not saw_stream_error
+                    and not cancel_event.is_set()
+                ):
                     finish_line = _synthetic_finish_line()
                     _monitor_openai_sse_line(
                         monitor_id,
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 190bfd6a36..1a649f09fe 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -2289,6 +2289,45 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_error_done_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"error":{"message":"boom","type":"server_error"}}',
+                    "data: [DONE]",
+                ],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_error_eof_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                ['data: {"error":{"message":"boom","type":"server_error"}}'],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" not in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod