fix: document turn duration fallback

api/streaming.py

@@ -2052,7 +2052,10 @@ def _run_agent_streaming(
             # Pass personality via ephemeral_system_prompt (agent's own mechanism)
             if _personality_prompt:
                 agent.ephemeral_system_prompt = _personality_prompt
-            _turn_started_at = getattr(s, 'pending_started_at', None) or time.time()
+            _pending_started_at = getattr(s, 'pending_started_at', None)
+            # Normal chat-start sets pending_started_at before spawning this thread;
+            # fallback to now only for recovered/legacy flows where that marker is absent.
+            _turn_started_at = _pending_started_at if _pending_started_at is not None else time.time()
             _previous_messages = list(s.messages or [])
             _previous_context_messages = list(_session_context_messages(s))
             _pre_compression_count = getattr(

tests/test_turn_duration_display.py

@@ -21,6 +21,14 @@ def test_streaming_done_payload_includes_backend_turn_duration():
         "Turn duration should be measured from the persisted pending_started_at "
         "start time, not only from browser-local state."
     )
+    assert "if _pending_started_at is not None else time.time()" in STREAMING_PY, (
+        "The fallback should preserve explicit timestamp values and only use now "
+        "when pending_started_at is absent."
+    )
+    assert "recovered/legacy flows" in STREAMING_PY, (
+        "The missing-start fallback should be documented so it is not mistaken "
+        "for the primary timing path."
+    )
     assert "_turnDuration" in STREAMING_PY, (
         "The measured duration should be persisted on the assistant message so "
         "it survives reload after the SSE stream settles."