grabowski
84309a7a08
Deploy web Studio / deploy (push) Successful in 3m33s
Found while recording a demo with Qwen3.8 on the inference server:
- When the model wrote its answer and then called select_object, it was
asked again with nothing left to say and rambled ("the user hasn't sent
a message yet"), sometimes in Chinese. A step of page tools only
(select_object, set_view) after a written answer now ends the turn; the
system prompt says so, and to answer in the language of the latest message.
- The system prompt names the world directions of the user's screen right
and up, and the 2:1 foreshortening, so "left of the hall" and shapes meant
to be read on screen come out right (asked for a giant W of people, the
model now draws it upright and readable).
- mock_llm.py "show me X" + check: answer and select_object in one step end
the turn without another request, and the history stays valid.