fix: repair for run from custom provider

Author: leonardmq

libs/core/kiln_ai/adapters/adapter_registry.py

@@ -90,16 +90,13 @@ def litellm_core_provider_config(
         run_config_properties.model_provider_name == ModelProviderName.openai_compatible
         and not run_config_properties.model_name.startswith("user_model::")
     ):
+        # Legacy openai_compatible model ids are "provider_name::model_id". We pull the
+        # provider name out for base_url lookup, but keep the full slug on the run config
+        # so it is faithfully persisted to disk and can be rehydrated later (e.g. on repair).
         model_id = run_config_properties.model_name
-        try:
-            openai_compatible_provider_name, model_id = model_id.split("::")
-        except Exception:
+        if "::" not in model_id:
             raise ValueError(f"Invalid openai compatible model ID: {model_id}")
-
-        # Update a copy of the run config properties to use the openai compatible provider
-        updated_run_config_properties = run_config_properties.model_copy(deep=True)
-        updated_run_config_properties.model_name = model_id
-        run_config_properties = updated_run_config_properties
+        openai_compatible_provider_name = model_id.split("::", 1)[0]
 
     config = lite_llm_core_config_for_provider(
         core_provider_name, openai_compatible_provider_name

libs/core/kiln_ai/adapters/provider_tools.py

@@ -245,6 +245,11 @@ def kiln_model_provider_from(
 def lite_llm_provider_model(
     model_id: str,
 ) -> KilnModelProvider:
+    # Legacy openai_compatible model ids are "{provider_name}::{model_id}".
+    # litellm only needs the bare model id (it builds "openai/<model_id>"),
+    # so strip the provider prefix here. The full slug stays on run_config.model_name.
+    if "::" in model_id:
+        model_id = model_id.split("::", 1)[1]
     return KilnModelProvider(
         name=ModelProviderName.openai_compatible,
         model_id=model_id,

libs/core/kiln_ai/adapters/test_adapter_registry.py

@@ -289,17 +289,72 @@ def test_openai_compatible_adapter(basic_task):
         assert isinstance(adapter, LiteLlmAdapter)
         assert adapter.config.additional_body_options == {"api_key": "test-key"}
         assert adapter.config.base_url == "https://test.com/v1"
-        assert adapter.config.run_config_properties.model_name == "test-model"
+        # The full slug is preserved on the run config so it can be persisted and rehydrated.
+        # The bare model id is only stripped at the litellm boundary (see lite_llm_provider_model).
+        assert (
+            adapter.config.run_config_properties.model_name
+            == "some-provider::test-model"
+        )
         assert (
             adapter.config.run_config_properties.model_provider_name
             == "openai_compatible"
         )
+        assert adapter.model_provider().model_id == "test-model"
         assert adapter.config.run_config_properties.prompt_id == "simple_prompt_builder"
         assert (
             adapter.config.run_config_properties.structured_output_mode == "json_schema"
         )
 
 
+def test_openai_compatible_adapter_preserves_model_name_for_rehydration(basic_task):
+    """Regression: building an adapter must not strip the legacy "{provider}::{model_id}"
+    prefix off run_config.model_name. _properties_for_task_output writes that name to
+    disk; if it's stripped, repair (and any other rehydration path) re-reads a name
+    without "::" and crashes splitting it again. See: openai/gpt-oss-safeguard-20b repair bug.
+    """
+    with patch("kiln_ai.adapters.provider_tools.Config.shared") as mock_config_shared:
+        mock_config_shared.return_value.openai_compatible_providers = [
+            {
+                "name": "vllm local",
+                "base_url": "http://localhost:8000/v1",
+                "api_key": "",
+            }
+        ]
+
+        original_run_config = KilnAgentRunConfigProperties(
+            model_name="vllm local::openai/gpt-oss-safeguard-20b",
+            model_provider_name=ModelProviderName.openai_compatible,
+            prompt_id="simple_prompt_builder",
+            structured_output_mode="json_schema",
+        )
+
+        adapter = adapter_for_task(
+            kiln_task=basic_task,
+            run_config_properties=original_run_config,
+        )
+
+        assert isinstance(adapter, LiteLlmAdapter)
+        # The slug round-trips intact on the run config that gets persisted.
+        assert (
+            adapter.config.run_config_properties.model_name
+            == "vllm local::openai/gpt-oss-safeguard-20b"
+        )
+        # Litellm sees the bare model id (no "::").
+        assert adapter.model_provider().model_id == "openai/gpt-oss-safeguard-20b"
+
+        # Simulate the repair flow: rebuild the adapter from the (now-correctly persisted)
+        # run config. This used to raise "Invalid openai compatible model ID: ..." because
+        # the first call had stripped the prefix.
+        rehydrated = adapter_for_task(
+            kiln_task=basic_task,
+            run_config_properties=adapter.config.run_config_properties.model_copy(
+                deep=True
+            ),
+        )
+        assert isinstance(rehydrated, LiteLlmAdapter)
+        assert rehydrated.config.base_url == "http://localhost:8000/v1"
+
+
 def test_custom_openai_compatible_provider(mock_config, basic_task):
     adapter = adapter_for_task(
         kiln_task=basic_task,

libs/core/kiln_ai/adapters/test_provider_tools.py

@@ -662,19 +662,24 @@ def test_openai_compatible_provider_config(mock_shared_config):
         config.run_config_properties.model_provider_name
         == ModelProviderName.openai_compatible
     )
-    assert config.run_config_properties.model_name == "gpt-4"
+    # Full slug is preserved on the run config; the bare model id is only used at the litellm boundary.
+    assert config.run_config_properties.model_name == "test_provider::gpt-4"
     assert config.additional_body_options == {"api_key": "test-key"}
     assert config.base_url == "https://api.test.com"
 
 
 def test_litellm_provider_model_success(mock_shared_config):
-    """Test successful creation of an OpenAI compatible provider"""
+    """Test successful creation of an OpenAI compatible provider.
+
+    The provider's model_id is the bare model id (without the legacy "provider::" prefix),
+    since that's what litellm needs to call the upstream API.
+    """
     model_id = "test_provider::gpt-4"
 
     provider = lite_llm_provider_model(model_id)
 
     assert provider.name == ModelProviderName.openai_compatible
-    assert provider.model_id == model_id
+    assert provider.model_id == "gpt-4"
     assert provider.supports_structured_output is False
     assert provider.supports_data_gen is False
     assert provider.untested_model is True
@@ -697,7 +702,7 @@ def test_lite_llm_config_no_api_key(mock_shared_config):
         config.run_config_properties.model_provider_name
         == ModelProviderName.openai_compatible
     )
-    assert config.run_config_properties.model_name == "gpt-4"
+    assert config.run_config_properties.model_name == "no_key_provider::gpt-4"
     assert config.additional_body_options == {"api_key": "NA"}
     assert config.base_url == "https://api.nokey.com"