assistant: pick the preferred model, loaded on demand

The daemon answers with whichever model llama-server has loaded and
never asks for one, so it never evicts a model another client is using.
The Assistant settings page now lists the router's models with their
state. Load saves one as the preferred model, in Prefs, and asks the
router to load it now, polling /v1/models until it reports loaded;
Prefer marks the loaded model as preferred without swapping. The swap
happens only on that click, so the daemon's rule holds.

AGENTS.md records the router's load endpoint and its two refusals, both
measured: it lives on the server root, not under /v1.

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

1 files changed, 7 insertions, 0 deletions

diff --git a/desktop/modules/assistant/README.md b/desktop/modules/assistant/README.md
index 347f9ea..07cc7ad 100644
--- a/desktop/modules/assistant/README.md
+++ b/desktop/modules/assistant/README.md
@@ -12,6 +12,13 @@ box, Save and "Use the default". The tool environment (`[env]`) is one row
 per key, with an add row below. A tool's switch renames `tools/<name>.toml`
 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.
+
 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
 does not come back shows the script's status, whose log tail says why.