blueprints/conversation/__init__.py

@@ -619,7 +619,7 @@ async def stream_query():
         _log_summary()
         yield "data: [DONE]\n\n"
 
-    return await make_response(
+    response = await make_response(
         event_generator(),
         200,
         {
@@ -628,6 +628,10 @@ async def stream_query():
             "X-Accel-Buffering": "no",
         },
     )
+    # Quart's default 60s response deadline cancels even active SSE streams.
+    # Silent model calls are bounded separately by MODEL_STALL_TIMEOUT.
+    response.timeout = None
+    return response
 
 
 @conversation_blueprint.route("/<conversation_id>")

tests/unit/test_stream_stall.py

@@ -86,7 +86,7 @@ def _final(text):
     }
 
 
-async def _run(owner, monkeypatch, script):
+async def _run(owner, monkeypatch, script, *, response_timeout=60):
     user, conversation = owner
     monkeypatch.setattr(conv, "MODEL_STALL_TIMEOUT", LIMIT)
     monkeypatch.setattr(conv, "get_agent_for", lambda u: _TimedAgent(script))
@@ -97,6 +97,7 @@ async def _run(owner, monkeypatch, script):
     monkeypatch.setattr(conv, "build_system_prompt_with_memories", _prompt)
 
     app = Quart(__name__)
+    app.config["RESPONSE_TIMEOUT"] = response_timeout
     app.config["JWT_SECRET_KEY"] = "test-secret"
     JWTManager(app)
     app.register_blueprint(conv.conversation_blueprint)
@@ -122,6 +123,22 @@ def test_stall_timeout_only_while_a_model_call_is_in_flight(monkeypatch):
     assert conv.stall_timeout(100.0, 200.0) == 1.0  # floor
 
 
+async def test_active_stream_outlasts_response_timeout(owner, monkeypatch):
+    script = [
+        (0, _start("m1")),
+        (0, _chunk("m1", "<think>")),
+        (0.15, _chunk("m1", "working")),
+        (0.15, _chunk("m1", "</think>Answer")),
+        (0, _end("m1")),
+        (0, _final("Answer")),
+    ]
+
+    frames = await _run(owner, monkeypatch, script, response_timeout=0.1)
+
+    assert "error" not in [f["type"] for f in frames]
+    assert frames[-1] == {"type": "response", "message": "Answer"}
+
+
 async def test_long_thinking_is_not_a_stall(owner, monkeypatch):
     # The whole call takes ~4x the limit, but output never pauses for longer
     # than half of it; the old time-to-first-content deadline killed this.