Fix startup warm-up to use user model settings instead of env var

Startup no longer auto-warms Config.OLLAMA_MODEL. Instead it queries
all distinct default_model values from user settings, cross-references
with Ollama's installed models, and warms only the intersection.

Models that users have selected but not yet installed are skipped with
an info log — they are never auto-pulled. The embedding model pull
behaviour is unchanged.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

src/fabledassistant/app.py

@@ -117,33 +117,88 @@ def create_app() -> Quart:
         start_log_retention_loop()
         start_notification_loop()
 
+        async def _warm_model(model: str) -> None:
+            """Warm an already-installed model into VRAM (no pull)."""
+            try:
+                async with httpx.AsyncClient(timeout=300.0) as client:
+                    await client.post(
+                        f"{Config.OLLAMA_URL}/api/generate",
+                        json={"model": model, "prompt": "", "keep_alive": "30m"},
+                    )
+                logger.info("Warmed model '%s' into VRAM", model)
+            except Exception:
+                logger.warning("Failed to warm model '%s'", model, exc_info=True)
+
+        async def _warm_user_models() -> None:
+            """
+            Warm whichever chat model(s) users have selected in Settings.
+
+            Only warms models that are already installed in Ollama — never auto-pulls.
+            Falls back silently if no user preferences exist or Ollama is unreachable.
+            """
+            from sqlalchemy import select as sa_select, distinct
+
+            from fabledassistant.models import async_session
+            from fabledassistant.models.setting import Setting
+
+            # 1. Collect all distinct default_model values users have saved.
+            try:
+                async with async_session() as session:
+                    rows = await session.execute(
+                        sa_select(distinct(Setting.value)).where(
+                            Setting.key == "default_model",
+                            Setting.value.isnot(None),
+                            Setting.value != "",
+                        )
+                    )
+                    user_models: set[str] = {r for (r,) in rows}
+            except Exception:
+                logger.debug("Could not read user model preferences from DB", exc_info=True)
+                return
+
+            if not user_models:
+                logger.debug("No user model preferences found; skipping warm-up")
+                return
+
+            # 2. Ask Ollama which models are currently installed.
+            try:
+                async with httpx.AsyncClient(timeout=10.0) as client:
+                    resp = await client.get(f"{Config.OLLAMA_URL}/api/tags")
+                    resp.raise_for_status()
+                    installed: set[str] = {m["name"] for m in resp.json().get("models", [])}
+            except Exception:
+                logger.debug("Could not reach Ollama to check installed models", exc_info=True)
+                return
+
+            # 3. Warm only the intersection (installed AND user-preferred).
+            for model in user_models:
+                base = model.removesuffix(":latest")
+                if model in installed or f"{base}:latest" in installed or base in installed:
+                    await _warm_model(model)
+                else:
+                    logger.info(
+                        "User-preferred model '%s' is not installed; skipping warm-up "
+                        "(install it via Settings → Models to enable auto-warm)",
+                        model,
+                    )
+
         async def _pull_model(model: str, warm: bool = False) -> None:
             try:
                 await ensure_model(model)
             except Exception:
                 logger.warning(
-                    "Failed to ensure model '%s' — chat may not work until model is available",
+                    "Failed to ensure model '%s'",
                     model,
                     exc_info=True,
                 )
                 return
             if warm:
-                try:
+                await _warm_model(model)
-                    async with httpx.AsyncClient(timeout=300.0) as client:
-                        await client.post(
-                            f"{Config.OLLAMA_URL}/api/generate",
-                            json={"model": model, "prompt": "", "keep_alive": "30m"},
-                        )
-                    logger.info("Warmed model '%s' into VRAM", model)
-                except Exception:
-                    logger.warning("Failed to warm model '%s'", model, exc_info=True)
 
-        # Pull and warm the main model into VRAM at startup so the first request is fast.
+        # Warm user-preferred chat models that are already installed.
-        asyncio.create_task(_pull_model(Config.OLLAMA_MODEL, warm=True))
+        # Also ensure the embedding model is pulled (no warm needed).
-        models_to_warm = {Config.OLLAMA_MODEL}
+        asyncio.create_task(_warm_user_models())
-        # Also pull the embedding model (nomic-embed-text by default), but no need to warm it.
+        asyncio.create_task(_pull_model(Config.EMBEDDING_MODEL, warm=False))
-        if Config.EMBEDDING_MODEL not in models_to_warm:
-            asyncio.create_task(_pull_model(Config.EMBEDDING_MODEL, warm=False))
 
         # After models are pulled, backfill embeddings for existing notes.
         # Runs in the background so it never blocks the server from accepting requests.