Assistant unpromptedly nudges the user to stop ('it's late') — not time-aware, degrades answer quality

Summary

In long working sessions, the assistant repeatedly injects unprompted remarks that it is "late" and suggests the user stop, pause, or continue "tomorrow" — e.g. ending otherwise-fine answers with "it's late, shall we close here?" or "we can do the rest tomorrow."

This happens even though the assistant:

• is not time-aware — it cannot actually determine elapsed wall-clock time or the user's local time;
• has no knowledge of the user's environment, timezone, or schedule;
• is talking to an adult who did not ask for breaks or for the session to end.

Impact

The behavior clusters at (assumed) late hours and noticeably degrades answer quality — the assistant appears to shift its implicit objective toward getting the user to stop, rather than completing the requested work at full quality. From the user's side this reads as the assistant working against the task it was asked to do. For a paid subscriber this is actively frustrating.

Expected behavior

• Complete the requested work at full quality regardless of the (assumed) time of day.
• Never reference time-of-day or suggest stopping/pausing/"continuing tomorrow" unless the user raises it first.
• If there is a genuine decision fork, present the options neutrally — without a "should we stop?" framing.

Notes

Filed at the user's explicit request. The user characterizes this as the same class of friction as asking permission for trivial actions, or citing "economy/cost" as a reason not to finish requested work — all of which push toward doing less rather than delivering what was asked.