diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 9605f4e6c..3e776eeba 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -5993,8 +5993,17 @@ def call_llm(
         # When the provider returns a 429 rate-limit (not billing), fall
         # back to an alternative provider instead of exhausting retries
         # against the same rate-limited endpoint.
+        #
+        # ── Auth error fallback (#21165) ─────────────────────────────
+        # When the resolved provider returns 401 and neither the Nous
+        # refresh path nor explicit provider credential refresh applies,
+        # fall back to an alternative provider instead of dropping the
+        # auxiliary task on the floor (silent compression failure /
+        # message loss). Auth is NOT a capacity error: it only bypasses
+        # the explicit-provider gate when the user is in auto mode.
         should_fallback = (
-            _is_payment_error(first_err)
+            _is_auth_error(first_err)
+            or _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
@@ -6024,7 +6033,9 @@ def call_llm(
             or _is_invalid_aux_response_error(first_err)
         )
         if should_fallback and (is_auto or is_capacity_error):
-            if _is_payment_error(first_err):
+            if _is_auth_error(first_err):
+                reason = "auth error"
+            elif _is_payment_error(first_err):
                 reason = "payment error"
                 # Resolve the actual provider label (resolved_provider may be
                 # "auto"; the client's base_url tells us which backend got the
@@ -6473,8 +6484,13 @@ async def async_call_llm(
                         raise
 
         # ── Payment / connection / rate-limit fallback (mirrors sync call_llm) ──
+        # Auth error fallback (#21165): a 401 that survived the refresh path
+        # falls back in auto mode just like the sync call_llm() path. Auth is
+        # NOT a capacity error, so on an explicit provider it still respects
+        # the user's choice (handled by the is_auto/is_capacity_error gate).
         should_fallback = (
-            _is_payment_error(first_err)
+            _is_auth_error(first_err)
+            or _is_payment_error(first_err)
             or _is_connection_error(first_err)
             or _is_rate_limit_error(first_err)
             or _is_model_incompatible_error(first_err)
@@ -6496,7 +6512,9 @@ async def async_call_llm(
             or _is_invalid_aux_response_error(first_err)
         )
         if should_fallback and (is_auto or is_capacity_error):
-            if _is_payment_error(first_err):
+            if _is_auth_error(first_err):
+                reason = "auth error"
+            elif _is_payment_error(first_err):
                 reason = "payment error"
                 _mark_provider_unhealthy(
                     _recoverable_pool_provider(resolved_provider, client) or resolved_provider
diff --git a/tests/agent/test_auxiliary_client.py b/tests/agent/test_auxiliary_client.py
index 7637b06d9..2100b5dcd 100644
--- a/tests/agent/test_auxiliary_client.py
+++ b/tests/agent/test_auxiliary_client.py
@@ -1842,6 +1842,62 @@ class TestCallLlmPaymentFallback:
         # Fallback client should have been used
         assert fallback_client.chat.completions.create.called
 
+    def test_401_auth_error_triggers_fallback_in_auto_mode(self, monkeypatch):
+        """401 auth errors should trigger fallback in auto mode (#21165).
+
+        When refresh is unavailable/fails and the user is on the auto chain,
+        a 401 must fall back instead of silently dropping the aux task
+        (which caused compression message loss).
+        """
+        monkeypatch.setenv("OPENROUTER_API_KEY", "or-key")
+
+        primary_client = MagicMock()
+        primary_client.base_url = "https://api.minimax.chat/v1"
+        primary_client.chat.completions.create.side_effect = _AuxAuth401("expired key")
+
+        fallback_client = MagicMock()
+        fallback_client.chat.completions.create.return_value = _DummyResponse("fallback auth response")
+
+        with patch("agent.auxiliary_client._get_cached_client",
+                   return_value=(primary_client, "minimax/minimax-m2.7")), \
+             patch("agent.auxiliary_client._resolve_task_provider_model",
+                   return_value=("auto", "minimax/minimax-m2.7", None, None, None)), \
+             patch("agent.auxiliary_client._try_payment_fallback",
+                   return_value=(fallback_client, "fallback-model", "openrouter")) as mock_fb:
+            result = call_llm(
+                task="compression",
+                messages=[{"role": "user", "content": "hello"}],
+            )
+
+        assert result.choices[0].message.content == "fallback auth response"
+        assert fallback_client.chat.completions.create.called
+        # Labelled as an auth error, not mis-tagged as a connection error.
+        assert mock_fb.call_args.kwargs.get("reason") == "auth error"
+
+    def test_401_auth_error_no_fallback_with_explicit_provider(self, monkeypatch):
+        """401 on an explicitly-configured provider must NOT silently switch.
+
+        Auth is not a capacity error: the explicit-provider gate means a 401
+        respects the user's choice and raises instead of falling back. This
+        guards the deliberate design at the should_fallback/is_capacity gate.
+        """
+        primary_client = MagicMock()
+        primary_client.base_url = "https://api.minimax.chat/v1"
+        primary_client.chat.completions.create.side_effect = _AuxAuth401("expired key")
+
+        with patch("agent.auxiliary_client._get_cached_client",
+                   return_value=(primary_client, "minimax/minimax-m2.7")), \
+             patch("agent.auxiliary_client._resolve_task_provider_model",
+                   return_value=("minimax", "minimax/minimax-m2.7", None, None, None)), \
+             patch("agent.auxiliary_client._refresh_provider_credentials", return_value=False), \
+             patch("agent.auxiliary_client._try_payment_fallback") as mock_fb:
+            with pytest.raises(_AuxAuth401):
+                call_llm(
+                    task="compression",
+                    messages=[{"role": "user", "content": "hello"}],
+                )
+        mock_fb.assert_not_called()
+
 
 class TestAuxiliaryFallbackLayering:
     """Explicit-provider users get layered fallback: configured_chain → main agent → warn."""