fix: preserve consumer slot compatibility

scripts/api_client.py

@@ -199,7 +199,7 @@ def fetch_pull_request_diff(
         url,
         token,
         payload=None,
-        accept="application/vnd.github.diff",
+        accept="application/vnd.github.v3.diff",
         **_retry_kwargs(retry_attempts, retry_backoff),
     )
     return response.text

templates/consumer-repo/scripts/api_client.py

@@ -199,7 +199,7 @@ def fetch_pull_request_diff(
         url,
         token,
         payload=None,
-        accept="application/vnd.github.diff",
+        accept="application/vnd.github.v3.diff",
         **_retry_kwargs(retry_attempts, retry_backoff),
     )
     return response.text

templates/consumer-repo/tools/ci_failure_triage.py

@@ -297,20 +297,19 @@ def _get_llm_client() -> tuple[_LLMClient, str] | None:
 
     if github_token:
         model = configured_model_for_provider("github-models")
-        if not model:
-            return None
-        return (
-            cast(
-                _LLMClient,
-                ChatOpenAI(
-                    model=model,
-                    base_url=GITHUB_MODELS_BASE_URL,
-                    api_key=cast(Any, github_token),
-                    temperature=0.1,
+        if model:
+            return (
+                cast(
+                    _LLMClient,
+                    ChatOpenAI(
+                        model=model,
+                        base_url=GITHUB_MODELS_BASE_URL,
+                        api_key=cast(Any, github_token),
+                        temperature=0.1,
+                    ),
                 ),
-            ),
-            "github-models",
-        )
+                "github-models",
+            )
     if openai_token is None:
         return None
     model = configured_model_for_provider("openai")

templates/consumer-repo/tools/llm_registry.py

@@ -332,7 +332,12 @@ def load_slot_config(*, github_default_model: str = "") -> list[SlotDefinition]:
                 select_model_for_profile(provider=provider, profile=profile, registry=registry)
                 or ""
             )
-        if provider and explicit_model and explicit_model != model:
+        explicit_entry = registry_entry_for(provider, explicit_model, registry=registry)
+        if provider and explicit_model and not configured_profile and explicit_entry and not (
+            explicit_entry.blocked or explicit_entry.lifecycle != "current"
+        ):
+            model = explicit_model
+        elif provider and explicit_model and explicit_model != model:
             logger.warning(
                 "Ignoring slot model pin %s/%s; reviewed %s selection is %s",
                 provider,
@@ -357,11 +362,9 @@ def load_slot_config(*, github_default_model: str = "") -> list[SlotDefinition]:
             continue
         name = str(entry.get("name") or f"slot{idx}").strip() or f"slot{idx}"
         slots.append(SlotDefinition(name=name, provider=provider, model=model))
-    # A syntactically valid config can still yield no usable entries (for
-    # example, every profile was retired or every model was blocked).  That is
-    # equivalent to no usable slot configuration; retain the reviewed defaults
-    # instead of silently disabling every provider.
-    return slots or fallback_slots
+    # A present slot file is an allowlist.  If every configured slot is
+    # unusable, fail closed instead of broadening execution to default providers.
+    return slots if slot_entries else fallback_slots
 
 
 def apply_slot_env_overrides(

tests/scripts/test_api_client.py

@@ -215,7 +215,7 @@ def fake_response(*args: object, **kwargs: object) -> object:
 
     assert api_client.fetch_pull_request("owner/repo", 7, "tok") == {"title": "PR"}
     assert api_client.fetch_pull_request_diff("owner/repo", 7, "tok") == response.text
-    assert calls == [{"payload": None, "accept": "application/vnd.github.diff"}]
+    assert calls == [{"payload": None, "accept": "application/vnd.github.v3.diff"}]
 
 
 def test_fetch_pull_request_rejects_nonobject_payload(monkeypatch: pytest.MonkeyPatch) -> None:

tests/tools/test_llm_registry_selection.py

@@ -102,18 +102,18 @@ def test_registry_decision_update_changes_slot_without_slot_edit(
     assert "model" not in json.loads(slots_path.read_text())["slots"][0]
 
 
-def test_legacy_slot_pin_cannot_override_reviewed_selection(
+def test_legacy_slot_pin_is_honored_when_current_and_unblocked(
     monkeypatch: pytest.MonkeyPatch, tmp_path: Path
 ) -> None:
     registry_path = tmp_path / "registry.json"
     slots_path = tmp_path / "slots.json"
     _write_registry(registry_path)
-    _write_slots(slots_path, model="model-frontier")
+    _write_slots(slots_path, model="model-frontier", profile="")
     monkeypatch.setenv(registry.ENV_MODEL_REGISTRY_CONFIG, str(registry_path))
     monkeypatch.setenv(registry.ENV_SLOT_CONFIG, str(slots_path))
 
-    assert registry.load_slot_config()[0].model == "model-balanced"
-    assert registry.configured_model_for_provider("openai") == "model-balanced"
+    assert registry.load_slot_config()[0].model == "model-frontier"
+    assert registry.configured_model_for_provider("openai") == "model-frontier"
 
 
 def test_unknown_explicit_profile_fails_closed(
@@ -126,13 +126,12 @@ def test_unknown_explicit_profile_fails_closed(
     monkeypatch.setenv(registry.ENV_MODEL_REGISTRY_CONFIG, str(registry_path))
     monkeypatch.setenv(registry.ENV_SLOT_CONFIG, str(slots_path))
 
-    # An unresolved slot profile must not disable the provider-level reviewed
-    # selection used as the safe fallback.
-    assert registry.load_slot_config()[0].model == "model-balanced"
+    # A present slot config is an allowlist; an unresolved profile fails closed.
+    assert registry.load_slot_config() == []
     assert registry.configured_model_for_provider("openai") == "model-balanced"
 
 
-def test_all_unusable_slot_entries_fall_back_to_reviewed_defaults(
+def test_all_unusable_slot_entries_fail_closed(
     monkeypatch: pytest.MonkeyPatch, tmp_path: Path
 ) -> None:
     registry_path = tmp_path / "registry.json"
@@ -142,7 +141,7 @@ def test_all_unusable_slot_entries_fall_back_to_reviewed_defaults(
     monkeypatch.setenv(registry.ENV_MODEL_REGISTRY_CONFIG, str(registry_path))
     monkeypatch.setenv(registry.ENV_SLOT_CONFIG, str(slots_path))
 
-    assert registry.load_slot_config()[0].model == "model-balanced"
+    assert registry.load_slot_config() == []
 
 
 def test_noncurrent_selected_model_fails_closed(

tools/ci_failure_triage.py

@@ -297,20 +297,19 @@ def _get_llm_client() -> tuple[_LLMClient, str] | None:
 
     if github_token:
         model = configured_model_for_provider("github-models")
-        if not model:
-            return None
-        return (
-            cast(
-                _LLMClient,
-                ChatOpenAI(
-                    model=model,
-                    base_url=GITHUB_MODELS_BASE_URL,
-                    api_key=cast(Any, github_token),
-                    temperature=0.1,
+        if model:
+            return (
+                cast(
+                    _LLMClient,
+                    ChatOpenAI(
+                        model=model,
+                        base_url=GITHUB_MODELS_BASE_URL,
+                        api_key=cast(Any, github_token),
+                        temperature=0.1,
+                    ),
                 ),
-            ),
-            "github-models",
-        )
+                "github-models",
+            )
     if openai_token is None:
         return None
     model = configured_model_for_provider("openai")

tools/evaluate_model_benchmark.py

@@ -153,7 +153,7 @@ def evaluate_benchmark(payload: dict[str, Any], policy: dict[str, Any]) -> dict[
     noninferiority_margin = float(gates["paired_success_noninferiority_margin"])
 
     for candidate in candidates:
-        model_id = str(candidate["model_id"])
+        model_id = str(candidate["model_id"]).strip()
         metrics = metrics_by_model[model_id]
         gate_results = {
             "minimum_adjudicated_cases": metrics["sample_count"] >= minimum_cases,

tools/llm_registry.py

@@ -332,7 +332,12 @@ def load_slot_config(*, github_default_model: str = "") -> list[SlotDefinition]:
                 select_model_for_profile(provider=provider, profile=profile, registry=registry)
                 or ""
             )
-        if provider and explicit_model and explicit_model != model:
+        explicit_entry = registry_entry_for(provider, explicit_model, registry=registry)
+        if provider and explicit_model and not configured_profile and explicit_entry and not (
+            explicit_entry.blocked or explicit_entry.lifecycle != "current"
+        ):
+            model = explicit_model
+        elif provider and explicit_model and explicit_model != model:
             logger.warning(
                 "Ignoring slot model pin %s/%s; reviewed %s selection is %s",
                 provider,
@@ -357,11 +362,9 @@ def load_slot_config(*, github_default_model: str = "") -> list[SlotDefinition]:
             continue
         name = str(entry.get("name") or f"slot{idx}").strip() or f"slot{idx}"
         slots.append(SlotDefinition(name=name, provider=provider, model=model))
-    # A syntactically valid config can still yield no usable entries (for
-    # example, every profile was retired or every model was blocked).  That is
-    # equivalent to no usable slot configuration; retain the reviewed defaults
-    # instead of silently disabling every provider.
-    return slots or fallback_slots
+    # A present slot file is an allowlist.  If every configured slot is
+    # unusable, fail closed instead of broadening execution to default providers.
+    return slots if slot_entries else fallback_slots
 
 
 def apply_slot_env_overrides(