agent window: screen directions, pointing ends the turn

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.

studio/agent.js

@@ -251,6 +251,15 @@
     ["z", "rot", "name", "tag", "props", "levels", "locked", "hidden"].forEach(function (k) { if (o[k] !== undefined && o[k] !== null && o[k] !== "") c[k] = o[k]; });
     return c;
   }
+  /** World directions of the user's screen right and up, for "left of", "above", shapes meant to be read. */
+  function screenAxes(rot) {
+    function dir(sx, sy) {
+      var a = IsoKit.unproject(rot, sx, sy, 0), o = IsoKit.unproject(rot, 0, 0, 0), x = a[0] - o[0], y = a[1] - o[1], l = Math.hypot(x, y) || 1;
+      return "(" + +(x / l).toFixed(3) + ", " + +(y / l).toFixed(3) + ")";
+    }
+    return "On the user's screen, right is the world direction " + dir(1, 0) + " and up is " + dir(0, -1) +
+      " (unit vectors in x, y). The view is isometric: a metre along screen-up looks half as long as a metre along screen-right, so a shape that should read correctly on screen (a letter, a logo) needs twice the depth in world metres along screen-up.";
+  }
   function system() {
     var s = st(), S = s.S, sel = S.sel && s.obj(S.sel), v = s.view;
     return {
@@ -259,9 +268,9 @@
         "You are the agent inside the isokit Studio, the page where the user edits an isometric scene. You change the scene with the tools; every edit appears on the user's map at once, and the user can take any edit back with Undo (Ctrl+Z).",
         A.instructions,
         "How to work: look before you build (list_objects shows what is where with world boxes; list_prefabs with a type gives its props). Place things inside the site, on free ground, and use the boxes to avoid overlaps. Put related edits into one apply_ops call so they undo as one step. After bigger changes, check with render_view or check_project. Use select_object to show the user what you mean. Ask back when a request is unclear instead of guessing at big changes.",
-        "Answer in the user's language, short and plain, without tables: say what you changed and give object ids in backticks (the user can click them).",
+        "Answer in the language of the user's latest message, short and plain, without tables: say what you changed and give object ids in backticks (the user can click them). To show the result, write the answer first and call select_object in the same step: that ends your turn.",
         "The Studio right now: project \"" + S.meta.title + "\", " + (S.doc.objects || []).length + " objects; view level \"" + v.getLevel() + "\", rotation " + v.getRotation() + "; snap step " + (s.snapStep() || "off") + ". " +
-          (sel ? "Selected (\"this\", \"it\"): " + JSON.stringify(compact(sel)) : "Nothing is selected.")
+          (sel ? "Selected (\"this\", \"it\"): " + JSON.stringify(compact(sel)) : "Nothing is selected.") + " " + screenAxes(v.getRotation())
       ].filter(Boolean).join("\n\n")
     };
   }
@@ -290,6 +299,9 @@
           return null;
         }
         pending = out.tool_calls.slice();
+        // pointing at things (select_object, set_view) after a written answer ends the turn: asked
+        // again with nothing left to say, models ramble (Qwen even switches language)
+        var gesture = !!m.content.trim() && out.tool_calls.every(function (c) { return PAGE[c["function"].name]; });
         return out.tool_calls.reduce(function (p, c) {
           return p.then(function () {
             if (ctl.signal.aborted) return;
@@ -300,7 +312,7 @@
               pending.shift();
             });
           });
-        }, Promise.resolve()).then(follow).then(function () { if (!ctl.signal.aborted) return turn(); });
+        }, Promise.resolve()).then(follow).then(function () { if (!ctl.signal.aborted && !gesture) return turn(); });
       }, function (e) {
         view.done(null);
         throw e;