diff --git a/ai-hub/app/core/orchestration/architect.py b/ai-hub/app/core/orchestration/architect.py
index e219cb9..e6d5a2c 100644
--- a/ai-hub/app/core/orchestration/architect.py
+++ b/ai-hub/app/core/orchestration/architect.py
@@ -1,14 +1,38 @@
 import logging
 import queue
 import time
+import traceback
 from typing import List, Dict, Any, Optional
 
+import litellm
+
 from app.db import models
 from .memory import ContextManager
 from .stream import StreamProcessor
 from .body import ToolExecutor
 from .guards import SafetyGuard
 
+_PROVIDER_ERROR_TYPES = (
+    litellm.ServiceUnavailableError,
+    litellm.InternalServerError,
+    litellm.RateLimitError,
+    litellm.APIConnectionError,
+    litellm.APIError,
+    litellm.Timeout,
+    litellm.AuthenticationError,
+)
+
+def _is_provider_error(e: Exception) -> bool:
+    """True for transient or user-actionable provider errors — not code bugs."""
+    if isinstance(e, _PROVIDER_ERROR_TYPES):
+        return True
+    # MidStreamFallbackError wraps another exception — match by name to avoid version drift
+    if type(e).__name__ in ("MidStreamFallbackError", "ContextWindowExceededError"):
+        return True
+    msg = str(e).lower()
+    return any(tok in msg for tok in ("503", "502", "429", "unavailable", "rate limit", "timeout", "overloaded"))
+
+
 class Architect:
     """
     The Master-Architect Orchestrator. 
@@ -113,10 +137,20 @@
                 yield {"type": "status", "content": f"Task complete in {elapsed:.1f}s"}
 
         except Exception as e:
-            import traceback
-            logging.error(f"[Architect] CRITICAL FAULT:\n{traceback.format_exc()}")
-            yield {"type": "status", "content": "Fatal Orchestration Error"}
-            yield {"type": "content", "content": f"\n\n> **🚨 Core Orchestrator Fault:** `{str(e)}`"}
+            if _is_provider_error(e):
+                logging.warning(f"[Architect] Provider error (non-fatal): {e}")
+                model_hint = getattr(llm_provider, 'model_name', 'the AI model')
+                yield {"type": "status", "content": "AI provider temporarily unavailable"}
+                yield {"type": "content", "content": (
+                    f"\n\n> ⚠️ **`{model_hint}` is temporarily unavailable.**\n"
+                    f">\n"
+                    f"> The provider returned a service error (likely overloaded or a preview endpoint outage).\n"
+                    f"> **Please try your message again** — or go to **Settings → AI Provider** to switch to a stable model."
+                )}
+            else:
+                logging.error(f"[Architect] CRITICAL FAULT:\n{traceback.format_exc()}")
+                yield {"type": "status", "content": "Fatal Orchestration Error"}
+                yield {"type": "content", "content": f"\n\n> **🚨 Core Orchestrator Fault:** `{str(e)}`"}
         finally:
             if registry and user_id: registry.unsubscribe_user(user_id, mesh_bridge)
             # --- M7: Automatic Terminal Task Cancellation ---
@@ -219,9 +253,13 @@
         if tools:
             kwargs["tools"] = tools
             kwargs["tool_choice"] = "auto"
+        # Let provider errors propagate so the top-level handler can show a user-friendly message.
+        # Only swallow true unknowns (e.g. missing provider config) which return None to exit cleanly.
         try:
             return await llm_provider.acompletion(messages=messages, timeout=60, **kwargs)
         except Exception as e:
+            if _is_provider_error(e):
+                raise
             logging.error(f"[Architect] LLM Exception: {e}")
             return None
 
diff --git a/ai-hub/app/core/providers/llm/general.py b/ai-hub/app/core/providers/llm/general.py
index 906510b..4b1329a 100644
--- a/ai-hub/app/core/providers/llm/general.py
+++ b/ai-hub/app/core/providers/llm/general.py
@@ -60,10 +60,17 @@
         }
         try:
             return await self._acompletion_with_retry(request)
+        except litellm.AuthenticationError as e:
+            raise litellm.AuthenticationError(
+                f"API key rejected by {self.model_name}. Please update your API key in Settings → AI Provider.",
+                model=self.model_name, llm_provider=""
+            ) from e
         except Exception as e:
             err_msg = str(e)
             if "authentication" in err_msg.lower() or "401" in err_msg:
-                raise RuntimeError(f"Authentication failed for {self.model_name}. Check your API key.")
-            
-            # If we still fail after retries, wrap it in a cleaner runtime error
-            raise RuntimeError(f"Core Orchestrator Fault: {err_msg}")
\ No newline at end of file
+                raise litellm.AuthenticationError(
+                    f"Authentication failed for {self.model_name}. Check your API key in Settings → AI Provider.",
+                    model=self.model_name, llm_provider=""
+                ) from e
+            # Re-raise with original type preserved so callers can classify provider vs code errors
+            raise
\ No newline at end of file