fix: raise Ollama timeout to 300s, add model pre-warming, improve health check

- OllamaBackend enforces _MIN_TIMEOUT=300s (overrides OLLAMA_TIMEOUT env var)
- warm_model() background task loads activeblue-chat into VRAM at startup
- health/detailed reports "warming" vs "ok" via Ollama ps() API
- README updated with May 2026 changes and test coverage details

agent_service/llm/ollama_backend.py

@@ -6,10 +6,12 @@ logger = logging.getLogger(__name__)
 
 
 class OllamaBackend:
-    def __init__(self, url, model, timeout=120, max_concurrent=2):
+    _MIN_TIMEOUT = 300  # activeblue-chat needs ~124s to load from disk
+
+    def __init__(self, url, model, timeout=300, max_concurrent=2):
         self._url = url
         self._model = model
-        self._timeout = timeout
+        self._timeout = max(timeout, self._MIN_TIMEOUT)
         self._semaphore = asyncio.Semaphore(max_concurrent)
         self._active = 0
 
@@ -78,5 +80,24 @@ class OllamaBackend:
         except Exception as exc:
             raise OllamaUnavailableError(f'Ollama ping failed: {exc}') from exc
 
+    async def warm_model(self) -> None:
+        """Pre-load the configured model into VRAM via a minimal inference call."""
+        import ollama
+        logger.info('ollama warm_model=%s starting (timeout=%ds)', self._model, self._timeout)
+        t0 = time.monotonic()
+        try:
+            client = ollama.AsyncClient(host=self._url)
+            await asyncio.wait_for(
+                client.chat(model=self._model, messages=[{'role': 'user', 'content': 'hi'}]),
+                timeout=self._timeout,
+            )
+            ms = int((time.monotonic() - t0) * 1000)
+            logger.info('ollama warm_model=%s ready in %dms', self._model, ms)
+        except asyncio.TimeoutError:
+            logger.warning('ollama warm_model=%s timed out after %ds — model may still be loading',
+                           self._model, self._timeout)
+        except Exception as exc:
+            logger.warning('ollama warm_model=%s failed: %s', self._model, exc)
+
     @property
     def active_count(self): return self._active