fix

backend/app/core/llm.py

@@ -79,7 +79,15 @@ class LlmClient:
         tool_calls: List[Dict[str, Any]] = []
         usage: Dict[str, int] = {}
         try:
-            async with httpx.AsyncClient(timeout=self.timeout) as client:
+            # Use explicit timeout config so connect/read/write/pool timeouts
+            # are all visible — a bare `timeout=N` hides WHICH stage failed.
+            timeout = httpx.Timeout(
+                connect=10.0,           # 10s to establish TCP connection
+                read=float(self.timeout),  # full request timeout
+                write=10.0,
+                pool=5.0,
+            )
+            async with httpx.AsyncClient(timeout=timeout) as client:
                 resp = await client.post(url, json=payload, headers=self._headers())
                 resp.raise_for_status()
                 data = resp.json()
@@ -88,9 +96,30 @@ class LlmClient:
             text = msg.get("content") or ""
             tool_calls = msg.get("tool_calls") or []
             usage = data.get("usage") or {}
+        except httpx.ConnectError as e:
+            err = f"ConnectError: {e}"
+            # Surface the URL + cause so the operator can see WHY (DNS, refused, etc.)
+            cause = getattr(e, "__cause__", None) or getattr(e, "__context__", None)
+            log.error(
+                "llm_connect_failed",
+                purpose=purpose,
+                url=url,
+                base_url=self.base_url,
+                model=self.model,
+                error=err,
+                cause=str(cause) if cause else None,
+            )
+            raise
         except Exception as e:
             err = f"{type(e).__name__}: {e}"
-            log.error("llm_call_failed", purpose=purpose, error=err)
+            log.error(
+                "llm_call_failed",
+                purpose=purpose,
+                url=url,
+                base_url=self.base_url,
+                model=self.model,
+                error=err,
+            )
             raise
         finally:
             latency_ms = int((time.monotonic() - started) * 1000)