feat: improved compatability layer and responses api handing

tests/core/test_byom_user_aware_helpers.py

@@ -791,7 +791,7 @@ class TestContinuationPreservesByomScope:
         )
 
         cont_service = MagicMock()
-        cont_service.load_state.return_value = {
+        cont_service.claim_state.return_value = {
             "messages": [],
             "pending_tool_calls": [],
             "tools_dict": {},

tests/core/test_model_registry_yaml.py

@@ -90,8 +90,11 @@ def _make_settings(**overrides):
 def _reset_registry(monkeypatch):
     ModelRegistry.reset()
     # openai_compatible catalogs read their key directly from os.environ,
-    # so clear any host-set keys to keep these tests deterministic.
-    monkeypatch.delenv("DEEPSEEK_API_KEY", raising=False)
+    # so clear every built-in/operator catalog key to keep these tests
+    # deterministic even when a developer has an ignored ``*_internal.yaml``.
+    for catalog in load_model_yamls([BUILTIN_MODELS_DIR]):
+        if catalog.api_key_env:
+            monkeypatch.delenv(catalog.api_key_env, raising=False)
     yield
     ModelRegistry.reset()
 
@@ -118,7 +121,17 @@ class TestYAMLLoader:
     def test_each_provider_has_expected_ids(self):
         grouped = _by_provider(load_model_yamls([BUILTIN_MODELS_DIR]))
         for provider, expected in EXPECTED_IDS.items():
-            actual = {m.id for c in grouped[provider] for m in c.models}
+            # Local ``*_internal.yaml`` files are intentionally gitignored
+            # operator extensions, not part of the canonical catalog snapshot.
+            canonical = [
+                c
+                for c in grouped[provider]
+                if not (
+                    c.source_path
+                    and c.source_path.stem.endswith("_internal")
+                )
+            ]
+            actual = {m.id for c in canonical for m in c.models}
             assert actual == expected, f"{provider}: expected {expected}, got {actual}"
 
     def test_attachment_alias_image_expands_to_five_mime_types(self):

tests/core/test_openai_compatible.py

@@ -16,6 +16,7 @@ import pytest
 
 from application.core.model_registry import ModelRegistry
 from application.core.model_settings import ModelProvider
+from application.core.model_yaml import BUILTIN_MODELS_DIR, load_model_yamls
 
 
 def _make_settings(**overrides):
@@ -81,7 +82,9 @@ def _reset_registry(monkeypatch):
     # so a key leaked by another test (or present in the dev .env) doesn't add
     # extra models to the exact-match assertions below. Mirrors the delenv
     # guard in test_model_registry_yaml.py.
-    monkeypatch.delenv("DEEPSEEK_API_KEY", raising=False)
+    for catalog in load_model_yamls([BUILTIN_MODELS_DIR]):
+        if catalog.api_key_env:
+            monkeypatch.delenv(catalog.api_key_env, raising=False)
     ModelRegistry.reset()
     yield
     ModelRegistry.reset()
@@ -211,6 +214,33 @@ class TestYAMLCompatibleProvider:
         for v in d.values():
             assert v != "sk"
 
+    def test_materialization_preserves_upstream_model_id(
+        self, tmp_path, monkeypatch
+    ):
+        path = _write_mistral_yaml(tmp_path)
+        path.write_text(
+            dedent("""
+                provider: openai_compatible
+                display_provider: mistral
+                api_key_env: MISTRAL_API_KEY
+                base_url: https://api.mistral.ai/v1
+                models:
+                  - id: mistral-large-high
+                    upstream_model_id: mistral-large-latest
+                    display_name: Mistral Large (High)
+            """)
+        )
+        monkeypatch.setenv("MISTRAL_API_KEY", "sk")
+
+        s = _make_settings(MODELS_CONFIG_DIR=str(tmp_path))
+        with patch("application.core.settings.settings", s):
+            reg = ModelRegistry()
+
+        assert (
+            reg.get_model("mistral-large-high").upstream_model_id
+            == "mistral-large-latest"
+        )
+
 
 # ── Legacy OPENAI_BASE_URL fallback ──────────────────────────────────────
 

tests/core/test_registry_user_layer.py

@@ -227,6 +227,8 @@ class TestLLMCreatorDispatchUsesUpstreamModelId:
                 "supports_structured_output": False,
                 "attachments": [],
                 "context_window": 8192,
+                "api_flavor": "responses",
+                "reasoning_effort": "high",
             },
         )
 
@@ -262,6 +264,8 @@ class TestLLMCreatorDispatchUsesUpstreamModelId:
         assert caps.supports_tools is False
         assert caps.supports_structured_output is False
         assert caps.supported_attachment_types == []
+        assert caps.api_flavor == "responses"
+        assert caps.reasoning_effort == "high"
 
     def test_byom_image_alias_expands_to_mime_types(self, pg_conn):
         """A BYOM stored with ``attachments: ["image"]`` (the alias the