fix(auxiliary): keep auto model resolved from runtime pair

agent/auxiliary_client.py

@@ -3420,13 +3420,14 @@ def resolve_provider_client(
     #      with grok-4.3 configured gets grok-4.3 for title generation
     #      instead of silently dropping to whatever Step-2 fallback (#31845).
     #
-    # Each provider branch below sees a non-empty ``model`` whenever the
-    # user has *anything* configured — no provider-specific empty-model
-    # guards needed.  When the user has NOTHING configured (fresh install,
-    # main_model also empty), the branches still hit their own
-    # missing-credentials returns and ``_resolve_auto`` falls through to
-    # the Step-2 chain as before.
-    if not model:
+    # Each explicit provider branch below sees a non-empty ``model`` whenever
+    # the user has *anything* configured — no provider-specific empty-model
+    # guards needed.  The ``auto`` branch is the exception: it must let
+    # ``_resolve_auto(main_runtime=...)`` choose the matched provider/model pair
+    # from the live runtime.  Falling back to ``_read_main_model()`` here can
+    # cross a stale config/default model (e.g. Opus) onto a newly failed-over
+    # provider client (e.g. Codex), producing provider+model mismatches.
+    if provider != "auto" and not model:
         model = _get_aux_model_for_provider(provider) or _read_main_model() or model
 
     def _needs_codex_wrap(client_obj, base_url_str: str, model_str: str) -> bool:

tests/agent/test_auxiliary_main_first.py

@@ -540,3 +540,61 @@ def test_runtime_provider_without_model_does_not_borrow_global_model(self):
         # Opus model was NOT borrowed onto the codex provider.
         mock_resolve.assert_not_called()
         assert client is None
+
+
+class TestResolveProviderClientAutoRuntimeModel:
+    """The outer resolver must not reintroduce stale main-config models."""
+
+    def test_auto_provider_does_not_let_stale_config_model_override_runtime_model(self):
+        """``provider=auto`` uses ``_resolve_auto``'s runtime pair, not ``_read_main_model``.
+
+        Regression for a production compression failure after mid-session
+        fallback: ``_resolve_auto(main_runtime=...)`` correctly selected the
+        Codex runtime pair (openai-codex / gpt-5.5), but
+        ``resolve_provider_client`` had already filled ``model`` from stale
+        process/config state (claude-opus-4-8) and then returned Codex client +
+        stale Opus model. Codex rejected that wire shape with HTTP 400.
+        """
+        codex_client = MagicMock()
+
+        with patch(
+            "agent.auxiliary_client._get_aux_model_for_provider", return_value=None,
+        ), patch(
+            "agent.auxiliary_client._read_main_model", return_value="claude-opus-4-8",
+        ), patch(
+            "agent.auxiliary_client._resolve_auto", return_value=(codex_client, "gpt-5.5"),
+        ) as mock_resolve_auto:
+            from agent.auxiliary_client import resolve_provider_client
+
+            client, model = resolve_provider_client(
+                "auto",
+                None,
+                False,
+                main_runtime={"provider": "openai-codex", "model": "gpt-5.5"},
+            )
+
+        assert client is codex_client
+        mock_resolve_auto.assert_called_once_with(
+            main_runtime={"provider": "openai-codex", "model": "gpt-5.5"}
+        )
+        assert model == "gpt-5.5"
+        assert model != "claude-opus-4-8"
+
+    def test_auto_provider_preserves_explicit_model_override(self):
+        """A real caller-supplied model still wins over the auto-resolved model."""
+        codex_client = MagicMock()
+
+        with patch(
+            "agent.auxiliary_client._resolve_auto", return_value=(codex_client, "gpt-5.5"),
+        ):
+            from agent.auxiliary_client import resolve_provider_client
+
+            client, model = resolve_provider_client(
+                "auto",
+                "gpt-5.4-mini",
+                False,
+                main_runtime={"provider": "openai-codex", "model": "gpt-5.5"},
+            )
+
+        assert client is codex_client
+        assert model == "gpt-5.4-mini"