assistant: Load no longer changes the preferred model

Load set the preference as well as swapping the model in, so loading a
model from the settings page for any reason made it the assistant's
model too: after loading Qwen there, the canvas kept Qwen instead of
bringing the preferred Gemma back. Load and Prefer are now separate
buttons: Load swaps a model in now and leaves the preference alone, and
Prefer only records which model the canvas loads when it opens.

The canvas header also showed a stale model name after a swap; it now
takes the model from the daemon's ctx event, which carries it from
desktop-assistant f9d4644 on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

4 files changed, 27 insertions, 15 deletions

diff --git a/desktop/assistant/Assistant.qml b/desktop/assistant/Assistant.qml
index 6692648..cc1d3d5 100644
--- a/desktop/assistant/Assistant.qml
+++ b/desktop/assistant/Assistant.qml
@@ -369,7 +369,9 @@ Scope {
             root.state = e.state;
             root.runningTool = e.state === "tool" ? e.tool : "";
             break;
-        case "ctx": root.ctxUsed = e.used; root.ctxMax = e.max; break;
+        // ctx carries the model since the router can swap it after hello; an
+        // older daemon leaves it out, and the hello's name stands.
+        case "ctx": root.ctxUsed = e.used; root.ctxMax = e.max; if (e.model) root.model = e.model; break;
         case "conversations": root.convs = e.items; break;
         case "results": root.results = e.items; break;
         }
diff --git a/desktop/modules/assistant/AssistantModule.qml b/desktop/modules/assistant/AssistantModule.qml
index b53a33f..8253cc6 100644
--- a/desktop/modules/assistant/AssistantModule.qml
+++ b/desktop/modules/assistant/AssistantModule.qml
@@ -179,11 +179,11 @@ Module {
     }
 
     // The router keeps one model resident, so this swaps out whatever is
-    // loaded, for every client: it runs only on a click. The id becomes the
-    // preferred one, and the list is polled until the router reports it
-    // loaded, since the request may return before the load finishes.
+    // loaded, for every client. It does not touch the preferred model: a
+    // load from the page is for now, and the canvas brings the preferred one
+    // back when it opens. The list is polled until the router reports the
+    // model loaded, since the request may return before the load finishes.
     function loadModel(id) {
-        Prefs.setAssistantModel(id);
         am.modelError = "";
         am.loading = id;
         am.run(["curl", "-s", "-m", "600", "-X", "POST", "-H", "Content-Type: application/json",
diff --git a/desktop/modules/assistant/AssistantPage.qml b/desktop/modules/assistant/AssistantPage.qml
index 2e40ed4..c6b63fc 100644
--- a/desktop/modules/assistant/AssistantPage.qml
+++ b/desktop/modules/assistant/AssistantPage.qml
@@ -165,8 +165,9 @@ Column {
     Small {
         width: page.width
         wrapMode: Text.Wrap
-        text: "The assistant answers with whichever model llama-server has loaded. Load makes one the "
-            + "preferred model and loads it now, which swaps out the loaded one for every client."
+        text: "The assistant answers with whichever model llama-server has loaded. Opening it loads "
+            + "the preferred model if another one is in. Load swaps a model in now, for every client, "
+            + "without changing the preference."
     }
 
     Small {
@@ -211,13 +212,20 @@ Column {
                        modelRow.preferred ? "preferred" : ""].filter(s => s).join("  ·  ")
                 color: modelRow.preferred ? Theme.accent : Theme.subtext
             }
+            // Two separate choices: which model the assistant prefers, kept,
+            // and which one is loaded now, a swap for every client.
             Button {
+                id: loadBtn
                 anchors { right: parent.right; rightMargin: 6; verticalCenter: parent.verticalCenter }
-                text: modelRow.loaded ? (modelRow.preferred ? "Loaded" : "Prefer") : "Load"
-                enabled: page.asst.loading === "" && !(modelRow.loaded && modelRow.preferred)
-                // A loaded model only becomes the preferred one: nothing to swap.
-                onClicked: modelRow.loaded ? Prefs.setAssistantModel(modelRow.modelData.id)
-                                           : page.asst.loadModel(modelRow.modelData.id)
+                text: "Load"
+                enabled: page.asst.loading === "" && !modelRow.loaded
+                onClicked: page.asst.loadModel(modelRow.modelData.id)
+            }
+            Button {
+                anchors { right: loadBtn.left; rightMargin: 6; verticalCenter: parent.verticalCenter }
+                text: modelRow.preferred ? "Preferred" : "Prefer"
+                enabled: !modelRow.preferred
+                onClicked: Prefs.setAssistantModel(modelRow.modelData.id)
             }
         }
     }
diff --git a/desktop/modules/assistant/README.md b/desktop/modules/assistant/README.md
index 93b8507..004ec05 100644
--- a/desktop/modules/assistant/README.md
+++ b/desktop/modules/assistant/README.md
@@ -15,9 +15,11 @@ to `<name>.toml.off`, which the daemon's `*.toml` glob skips.
 The page also lists llama-server's models with their state. The daemon
 always answers with whichever model is loaded and never asks for one, so
 it never swaps a model out from under another client. Load is the
-deliberate swap: it saves that model as the preferred one (in `Prefs`)
-and asks the router to load it now, polling `/v1/models` until it reports
-`loaded`. Prefer marks the already-loaded model as preferred.
+deliberate swap: it asks the router to load that model now, polling
+`/v1/models` until it reports `loaded`, and leaves the preference alone.
+Prefer is the other, separate choice: the model the assistant should use,
+saved in `Prefs`. A first version made Load also set the preference, so a
+model loaded for another client became the assistant's too.
 
 Opening the assistant canvas (SUPER+Home) loads the preferred model when
 another one is resident, and the canvas header reads "loading <model>…"