4 files changed, 31 insertions, 12 deletions
diff --git a/desktop/assistant/Card.qml b/desktop/assistant/Card.qml
index b9fe4e5..810920f 100644
--- a/desktop/assistant/Card.qml
+++ b/desktop/assistant/Card.qml
@@ -465,7 +465,7 @@ Rectangle {
             text: ""
             font.family: Theme.iconFamily
             onClicked: {
-                card.canvas.historyOpen = false;
+                card.canvas.historyOpen = card.canvas.modelsOpen = false;
                 card.canvas.pickerOpen = !card.canvas.pickerOpen;
             }
         }
diff --git a/desktop/assistant/README.md b/desktop/assistant/README.md
index 58fbb14..4c76a66 100644
--- a/desktop/assistant/README.md
+++ b/desktop/assistant/README.md
@@ -18,6 +18,7 @@ what is typed. The protocol is specified in that repo,
 | `Segments.qml` | markdown with a copy button per code block |
 | `History.qml` | conversation list and full-text search |
 | `Picker.qml` | the in-card file picker behind the paperclip |
+| `Models.qml` | the model panel: local and cloud models, the chat's pick |
 | `Md.js` | splits a message into prose and fenced code |
 | `Files.js` | a path's kind, and a dropped URL to a path |
 | `Presets.js` | the Ask row's canned questions and which kinds they fit |
@@ -38,9 +39,16 @@ model transcribe it and answers the words. The mic is greyed out while the
 model in use reports no `audio` input (`AssistantModule.hears`, from the
 router's `/models`, read on every open).
 
+The header's model name (the Model button outside chat, or Ctrl+M) opens
+the model panel: the router's models, then each cloud provider's. A pick
+applies to this chat from its next question and is stored with it by the
+daemon; a new chat starts on the preferred model. Moving a chat with turns
+to a cloud model says once that its history goes to that provider. Voice is
+transcribed by the local model in a cloud chat too.
+
 Keys: Enter sends, Shift+Enter is a newline, Escape cancels a recording,
 then closes the picker, the history panel, then the card. Ctrl+N starts a
-new chat, Ctrl+H toggles the history panel. Stop replaces
+new chat, Ctrl+H toggles the history panel, Ctrl+M the model panel. Stop replaces
 Send while a reply streams.
 
 Notes:
diff --git a/desktop/modules/assistant/README.md b/desktop/modules/assistant/README.md
index 3c35dc6..9ee9736 100644
--- a/desktop/modules/assistant/README.md
+++ b/desktop/modules/assistant/README.md
@@ -15,6 +15,14 @@ The system prompt also has "Use the default". The tool environment
 saved, renames `tools/<name>.toml`
 to `<name>.toml.off`, which the daemon's `*.toml` glob skips.
 
+**Cloud** edits the daemon's `[[cloud]]` tables: name, URL and key
+command per provider. Fetch saves the page, then runs `assistant.py
+--cloud-models <name>`, which runs the key command and lists the
+provider's ids; the switches choose which the canvas offers. The key
+never reaches this page. A table holding more than the page writes (a
+multi-line array, another key) is left alone, with a note to edit it by
+hand. The saved cloud models also get Prefer in the Model list.
+
 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
@@ -24,12 +32,12 @@ 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>…"
-until it is in. This desktop runs one model and one client at a time, so
-opening the assistant is when its model should come back; the daemon still
-never asks for a model itself. With no preferred model set, nothing is
-swapped.
+Each chat uses the model picked in the canvas's model menu, else the
+preferred one, else whatever is loaded. Opening the canvas, or picking a
+local model for its chat, loads that model when another one is resident,
+and the canvas header reads "loading <model>…" until it is in. This
+desktop runs one model and one client at a time, so that is when the
+model should come back. A cloud model loads nothing.
 
 The daemon reads all of this only at startup, so after a change the page
 says so and offers Restart, which runs `assistant.sh restart`; a daemon that
diff --git a/desktop/shell.qml b/desktop/shell.qml
index 4b4f47f..8aaead3 100644
--- a/desktop/shell.qml
+++ b/desktop/shell.qml
@@ -126,17 +126,20 @@ ShellRoot {
 
     // The assistant canvas on SUPER+Home, a view over desktop-assistant's
     // socket. See assistant/README.md.
-    Assistant { id: assistant; swapping: asstSettings.loading; hears: asstSettings.hears }
+    Assistant { id: assistant; swapping: asstSettings.loading; hears: asstSettings.hears
+                preferred: Prefs.assistantModel }
 
     // The desktop companion, bottom-right of the focused monitor, on SUPER+\.
     // She reads the assistant's state. See companion/README.md.
     Companion { asst: assistant }
 
-    // Opening the assistant loads its preferred model if another one is in
-    // VRAM. See modules/assistant/README.md.
+    // Opening the assistant, or picking another model for its chat, loads
+    // that model if it is a local one and another is in VRAM. See
+    // modules/assistant/README.md.
     Connections {
         target: assistant
-        function onOpenChanged() { if (assistant.open) asstSettings.ensurePreferred(); }
+        function onOpenChanged() { if (assistant.open) asstSettings.ensureModel(assistant.wanted); }
+        function onWantedChanged() { if (assistant.open) asstSettings.ensureModel(assistant.wanted); }
     }
 
     // The waybar launcher and the deep-link keybinds all reach this: