fix(conductor): restore standalone model selection

frontends/conductor.html

@@ -540,7 +540,12 @@
           messagesEl.innerHTML=''; (data.chat||[]).forEach(addMsg);
           renderLog(data.log||[]);
           llmEl.innerHTML=(data.llms||[]).map(x=>`<option value="${x[0]}">${esc(x[1])}</option>`).join(''); llmEl.value=data.llm; llmEl.disabled=!llmEl.options.length;
-        } else if(data.type==='subagents'){ renderCards(data.items||[]); }
+        } else if(data.type==='model_selected'){ llmEl.value=String(data.llm); }
+        else if(data.type==='model_error'){
+          if(data.llm!==undefined) llmEl.value=String(data.llm);
+          showLimitToast(data.error||'Failed to select model');
+        }
+        else if(data.type==='subagents'){ renderCards(data.items||[]); }
         else if(data.type==='chat'){ addMsg(data.item); }
         else if(data.type==='chat_read'){ markAllRead(); }
         else if(data.type==='log'){ appendLog(data.item); }

frontends/conductor.py

@@ -1,5 +1,6 @@
 import os, sys, re, time, json, uuid, queue, asyncio, threading, argparse, base64, secrets
 from dataclasses import dataclass, field
+from pathlib import Path
 from typing import Dict, Any, Optional, List
 from contextlib import asynccontextmanager, suppress
 
@@ -8,6 +9,10 @@ from fastapi.responses import FileResponse, PlainTextResponse, JSONResponse
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 
+FRONTENDS_DIR = os.path.dirname(os.path.abspath(__file__))
+if FRONTENDS_DIR not in sys.path: sys.path.insert(0, FRONTENDS_DIR)
+from desktop_settings import DesktopSettingsError, update_settings
+
 def _resolve_ga_root() -> str:
     """Use the external core selected by the package-owned desktop bridge when valid."""
     value = (os.environ.get("GA_ROOT", "") or "").strip()
@@ -67,15 +72,28 @@ HOST = args.host
 PORT = args.port
 
 
+SETTINGS_PATH = Path.home() / ".ga_desktop_settings.json"
+
+
 def _settings_doc() -> dict:
     try:
-        from pathlib import Path
-        doc = json.loads((Path.home() / ".ga_desktop_settings.json").read_text(encoding="utf-8"))
+        doc = json.loads(SETTINGS_PATH.read_text(encoding="utf-8"))
         return doc if isinstance(doc, dict) else {}
     except Exception:
         return {}
 
 
+def _persist_conductor_llm_no(llm_no: int) -> None:
+    def mutate(document: dict) -> None:
+        section = document.get("conductor")
+        if not isinstance(section, dict):
+            section = {}
+            document["conductor"] = section
+        section["llmNo"] = llm_no
+
+    update_settings(SETTINGS_PATH, mutate)
+
+
 def _conductor_llm_no() -> Optional[int]:
     """Read the model index bound to the conductor session.
     Falls back to the legacy desktop default ui.llmNo for existing installs."""
@@ -180,6 +198,24 @@ def _apply_desktop_model(agent: "GenericAgent") -> dict:
             "fallbackReason": "no_models", "current": None}
 
 
+def _selected_conductor_llm_no(agent: "GenericAgent") -> int:
+    configured = _conductor_llm_no()
+    return configured if _usable_model(agent, configured) else agent.llm_no
+
+
+def _set_conductor_llm_no(agent: "GenericAgent", value: Any) -> int:
+    """Persist the standalone UI binding without mutating an in-flight client.
+
+    The Conductor loop applies this binding at its existing idle task boundary, so
+    a selection made immediately before sending a message controls that next turn.
+    """
+    no = _parse_model_no(value)
+    if no is None or not _usable_model(agent, no):
+        raise ValueError(f"llm index out of range or unavailable: {value}")
+    _persist_conductor_llm_no(no)
+    return no
+
+
 def _select_llm(agent: "GenericAgent", llm: Any) -> bool:
     if llm is None or str(llm).strip() == "": return False
     q = str(llm).strip()
@@ -800,9 +836,22 @@ async def websocket(ws: WebSocket):
         await ws.send_json({"type": "hello", "subagents": pool.snapshot(), "chat": chat_messages,
                             "log": conductor.log, "running": running,
                             "model": conductor.model_snapshot(),
-                            "llms": conductor.agent.list_llms(), "llm": conductor.agent.llm_no})
+                            "llms": conductor.agent.list_llms(),
+                            "llm": _selected_conductor_llm_no(conductor.agent)})
         while True:
             data = await ws.receive_json()
+            if "llm" in data:
+                try:
+                    llm_no = _set_conductor_llm_no(conductor.agent, data["llm"])
+                except (TypeError, ValueError, OSError, DesktopSettingsError) as error:
+                    await ws.send_json({
+                        "type": "model_error",
+                        "error": str(error),
+                        "llm": _selected_conductor_llm_no(conductor.agent),
+                    })
+                else:
+                    await broadcast({"type": "model_selected", "llm": llm_no})
+                continue
             msg = (data.get("msg") or "").strip()
             if not msg: continue
             add_chat(msg, role="user", files=data.get("files") or [], images=data.get("images") or [])

frontends/tests/test_conductor_model_routing.py

@@ -164,6 +164,50 @@ def test_explicit_numeric_worker_model_rejects_out_of_range():
     assert agent.next_llm_calls == []
 
 
+def test_standalone_model_selection_persists_without_mutating_live_client(tmp_path, monkeypatch):
+    settings_path = tmp_path / "settings.json"
+    settings_path.write_text('{"theme":"dark","ui":{"llmNo":0}}', encoding="utf-8")
+    monkeypatch.setattr(conductor, "SETTINGS_PATH", settings_path)
+    agent = FakeAgent(["zero", "one", "two"])
+
+    selected = conductor._set_conductor_llm_no(agent, 2)
+
+    assert selected == 2
+    assert agent.llm_no == 0
+    assert agent.next_llm_calls == []
+    assert conductor._settings_doc() == {
+        "theme": "dark",
+        "ui": {"llmNo": 0},
+        "conductor": {"llmNo": 2},
+    }
+
+
+def test_standalone_model_selection_rejects_unavailable_model(tmp_path, monkeypatch):
+    settings_path = tmp_path / "settings.json"
+    settings_path.write_text('{"theme":"dark"}', encoding="utf-8")
+    monkeypatch.setattr(conductor, "SETTINGS_PATH", settings_path)
+    agent = FakeAgent(["zero", None])
+
+    with pytest.raises(ValueError, match="unavailable"):
+        conductor._set_conductor_llm_no(agent, 1)
+
+    assert conductor._settings_doc() == {"theme": "dark"}
+
+
+def test_standalone_model_selection_applies_at_next_task_boundary(tmp_path, monkeypatch):
+    settings_path = tmp_path / "settings.json"
+    monkeypatch.setattr(conductor, "SETTINGS_PATH", settings_path)
+    agent = FakeAgent(["zero", "one", "two"])
+
+    conductor._set_conductor_llm_no(agent, 2)
+    state = conductor._apply_desktop_model(agent)
+
+    assert agent.next_llm_calls == [2]
+    assert state["configured"] == 2
+    assert state["effective"] == 2
+    assert state["fallbackReason"] is None
+
+
 def test_runtime_model_snapshot_is_broadcast_with_running_state(monkeypatch):
     instance = conductor.Conductor()
     payloads: list[dict] = []