fix: guard composite leaf resolution

Author: Mateusz

src/core/services/backend_model_resolver.py

@@ -37,6 +37,7 @@
 from src.core.services.composite_routing_state import (
     COMPOSITE_LEAF_RESOLUTION_EXTRA_BODY_KEY,
     COMPOSITE_LEAF_RESOLUTION_FLAG,
+    is_composite_selector,
     resolve_composite_routing_surface,
 )
 from src.core.services.replacement_compatibility_bridge import (
@@ -470,16 +471,20 @@ def _require_explicit_backend_for_surface(surface: RoutingSurface) -> bool:
         return surface is RoutingSurface.REPLACEMENT_BRIDGE
 
     @staticmethod
-    def _is_composite_leaf_resolution(
-        *,
-        context: RequestContext | None,
-        request: ChatRequest,
-    ) -> bool:
-        if context is None:
-            extra_body = request.extra_body
-            if isinstance(extra_body, dict):
-                return bool(extra_body.get(COMPOSITE_LEAF_RESOLUTION_EXTRA_BODY_KEY))
-            return False
+    def _is_composite_leaf_resolution(
+        *,
+        context: RequestContext | None,
+        request: ChatRequest,
+    ) -> bool:
+        model = request.model
+        if is_composite_selector(model):
+            return False
+
+        if context is None:
+            extra_body = request.extra_body
+            if isinstance(extra_body, dict):
+                return bool(extra_body.get(COMPOSITE_LEAF_RESOLUTION_EXTRA_BODY_KEY))
+            return False
         if bool(context.extensions.get(COMPOSITE_LEAF_RESOLUTION_FLAG)):
             return True
         extra_body = request.extra_body

tests/unit/core/services/test_backend_model_resolver_composite_entrypoint.py

@@ -499,6 +499,30 @@ async def test_composite_failover_across_numbered_backend_instances() -> None:
     assert result.uri_params == {"top_p": "0.9"}
 
 
+@pytest.mark.asyncio
+async def test_stale_leaf_resolution_flag_does_not_bypass_composite_failover() -> None:
+    resolver = _build_resolver_with_real_composite(
+        unavailable_backends={"openai-codex"}
+    )
+    request = ChatRequest(
+        model=(
+            "openai-codex:gpt-5.3-codex?reasoning_effort=low|"
+            "anthropic:claude-3-5-sonnet?reasoning_effort=medium"
+        ),
+        messages=[ChatMessage(role="user", content="hello")],
+        extra_body={"_composite_leaf_resolution": True},
+    )
+
+    result = await resolver.resolve_target(
+        request,
+        context=_context("main"),
+    )
+
+    assert result.backend == "anthropic"
+    assert result.model == "claude-3-5-sonnet"
+    assert result.uri_params == {"reasoning_effort": "medium"}
+
+
 @pytest.mark.asyncio
 async def test_composite_weighted_numbered_instances_respects_deterministic_random() -> (
     None