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, 6 insertions, 0 deletions

diff --git a/AGENTS.md b/AGENTS.md
index 047adf7..79b8916 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -354,6 +354,12 @@ changing that component. The ones that generalise:
   GPU. Pass `autoload=false`, which answers 400 `model is not loaded` instead.
   Checking `/models` for `loaded` first is not enough on its own, because the
   model can unload between the two requests.
+- **llama-server's router loads a model with `POST /models/load`** and a
+  `{"model": "<id>"}` body, on the server root, not under `/v1`. Asking for
+  the loaded model answers 400 `model is already running`, and an unknown
+  id 404; both measured. With `--models-max 1` a load swaps out the
+  resident model for every client, so the assistant page does it only on a
+  click.
 - **`busctl get-property` escapes non-ASCII as octal.** A name with an
   accented letter reads `s "Do\303\250r"`-style, which no shell builtin
   decodes. `busctl --json=short` returns real UTF-8; `userctl` pipes it