fix(gateway): stamp turn_duration on last AI message only in /messages/page (#4755)

_enrich_thread_message_page inlined its own turn_duration loop that
stamped EVERY AI message of a run, re-introducing #4152 on the
/messages/page endpoint (the legacy /messages and /history endpoints
already route through stamp_turn_duration_on_last_ai after #4163, but
that fix missed the page path introduced earlier in #4065). A
multi-step turn thus rendered the same run lifetime beside every
intermediate AI message, reading as repeated thinking latency.

Replace the inline loop with the shared stamp_turn_duration_on_last_ai
helper so all three message endpoints agree: the run's wall-clock
duration lands on its final visible AI message only.

backend/app/gateway/routers/thread_runs.py

@@ -1329,9 +1329,14 @@ async def _enrich_thread_message_page(
                     "comment": feedback.get("comment"),
                 }
 
-        content = row.get("content")
-        if isinstance(content, dict) and content.get("type") == "ai" and run_id in run_durations:
-            content.setdefault("additional_kwargs", {})["turn_duration"] = run_durations[run_id]
+    # ``turn_duration`` is the run's wall-clock lifetime, not model thinking
+    # time — stamp it on the run's LAST visible AI message only so the UI does
+    # not repeat the same number on every intermediate AI message of a
+    # multi-step turn (#4152). The legacy ``GET /messages`` and ``/history``
+    # endpoints already use ``stamp_turn_duration_on_last_ai``; the page
+    # endpoint was inlining the equivalent loop but stamping every AI row,
+    # which #4163 fixed for the other paths and missed here.
+    stamp_turn_duration_on_last_ai(data, run_durations)
     return data
 
 

backend/tests/test_thread_messages_page.py

@@ -443,6 +443,37 @@ def test_thread_page_batch_hydrates_duration_for_old_runs():
     assert response.json()["data"][0]["content"]["additional_kwargs"]["turn_duration"] == 7
 
 
+def test_thread_page_stamps_turn_duration_on_last_ai_message_only():
+    # #4152 regression: a run with multiple AI messages (multi-step turn) must
+    # carry turn_duration on the LAST AI message only, not on every one.
+    store = MemoryRunEventStore()
+
+    async def seed():
+        await _put_message(store, "run-multi", "ai", "first-answer")
+        await _put_message(store, "run-multi", "ai", "second-answer")
+
+    asyncio.run(seed())
+    record = RunRecord(
+        run_id="run-multi",
+        thread_id="thread-1",
+        assistant_id=None,
+        status="success",
+        on_disconnect="cancel",
+        created_at="2026-01-01T00:00:00Z",
+        updated_at="2026-01-01T00:00:07Z",
+    )
+    app = _make_app(store, records={"run-multi": record})
+    with TestClient(app) as client:
+        response = client.get("/api/threads/thread-1/messages/page")
+
+    data = response.json()["data"]
+    # Messages are returned newest-first; the last AI message is data[0].
+    first_row = next(row for row in data if row["content"]["id"] == "first-answer")
+    second_row = next(row for row in data if row["content"]["id"] == "second-answer")
+    assert second_row["content"]["additional_kwargs"]["turn_duration"] == 7
+    assert "turn_duration" not in first_row["content"].get("additional_kwargs", {})
+
+
 def test_thread_page_preserves_tool_and_subagent_wrapper_metadata():
     store = MemoryRunEventStore()
     asyncio.run(