chore(infra): support langsmith gateway in CI (#39651)

.github/workflows/_release.yml

@@ -497,9 +497,10 @@ jobs:
         # Uses the Makefile's `integration_tests` target for the specified package
         if: ${{ startsWith(env.EFFECTIVE_WORKING_DIR, 'libs/partners/') }}
         env:
+          LANGSMITH_GATEWAY: ${{ secrets.LANGSMITH_GATEWAY }}
+          LANGSMITH_GATEWAY_API_KEY: ${{ secrets.LANGSMITH_GATEWAY_API_KEY }}
           AI21_API_KEY: ${{ secrets.AI21_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
-          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           MISTRAL_API_KEY: ${{ secrets.MISTRAL_API_KEY }}
           TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
@@ -600,7 +601,8 @@ jobs:
         partner: [ anthropic, openai ]
       fail-fast: false # Continue testing other partners if one fails
     env:
-      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+      LANGSMITH_GATEWAY: ${{ secrets.LANGSMITH_GATEWAY }}
+      LANGSMITH_GATEWAY_API_KEY: ${{ secrets.LANGSMITH_GATEWAY_API_KEY }}
       ANTHROPIC_FILES_API_IMAGE_ID: ${{ secrets.ANTHROPIC_FILES_API_IMAGE_ID }}
       ANTHROPIC_FILES_API_PDF_ID: ${{ secrets.ANTHROPIC_FILES_API_PDF_ID }}
       OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

.github/workflows/integration_tests.yml

@@ -288,8 +288,9 @@ jobs:
           LANGSMITH_PROJECT: ${{ vars.LANGSMITH_PROJECT || 'scheduled-testing-py' }}
           LANGSMITH_TAGS: "github-actions,${{ matrix.working-directory }},python-${{ matrix.python-version }},sha-${{ github.sha }}"
 
+          LANGSMITH_GATEWAY: ${{ secrets.LANGSMITH_GATEWAY }}
+          LANGSMITH_GATEWAY_API_KEY: ${{ secrets.LANGSMITH_GATEWAY_API_KEY }}
           AI21_API_KEY: ${{ secrets.AI21_API_KEY }}
-          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           ANTHROPIC_FILES_API_IMAGE_ID: ${{ secrets.ANTHROPIC_FILES_API_IMAGE_ID }}
           ANTHROPIC_FILES_API_PDF_ID: ${{ secrets.ANTHROPIC_FILES_API_PDF_ID }}
           ASTRA_DB_API_ENDPOINT: ${{ secrets.ASTRA_DB_API_ENDPOINT }}

libs/partners/anthropic/tests/conftest.py

@@ -8,6 +8,10 @@ from vcr import VCR  # type: ignore[import-untyped]
 def remove_request_headers(request: Any) -> Any:
     for k in request.headers:
         request.headers[k] = "**REDACTED**"
+    # Redact the URI so cassettes stay host-agnostic (e.g. direct API vs. the
+    # LangSmith gateway). Matching drops the `uri` matcher below, so this only
+    # affects what newly recorded cassettes persist.
+    request.uri = "**REDACTED**"
     return request
 
 
@@ -21,6 +25,9 @@ def remove_response_headers(response: dict) -> dict:
 def vcr_config() -> dict:
     """Extend the default configuration coming from langchain_tests."""
     config = base_vcr_config()
+    # Match on method + body only. Dropping `uri` keeps cassettes host-agnostic
+    # so they replay whether requests go to the API directly or via the gateway.
+    config["match_on"] = [m for m in config.get("match_on", []) if m != "uri"]
     config["before_record_request"] = remove_request_headers
     config["before_record_response"] = remove_response_headers
     config["serializer"] = "yaml.gz"

libs/partners/anthropic/tests/integration_tests/test_chat_models.py

@@ -1268,7 +1268,9 @@ def test_structured_output_thinking_force_tool_use() -> None:
     # Structured output currently relies on forced tool use, which is not supported
     # when `thinking` is enabled. When this test fails, it means that the feature
     # is supported and the workarounds in `with_structured_output` should be removed.
-    client = anthropic.Anthropic()
+    # Use the client resolved off `ChatAnthropic` so requests honor any
+    # configured base URL / credentials (e.g. the LangSmith gateway).
+    client = ChatAnthropic(model="claude-sonnet-4-5-20250929")._client  # type: ignore[call-arg]
     with pytest.raises(anthropic.BadRequestError):
         _ = client.messages.create(
             model="claude-sonnet-4-5-20250929",

libs/partners/anthropic/tests/integration_tests/test_llms.py

@@ -1,5 +1,6 @@
 """Test Anthropic API wrapper."""
 
+import os
 from collections.abc import Generator
 
 import pytest
@@ -11,6 +12,21 @@ from tests.unit_tests._utils import FakeCallbackHandler
 
 MODEL = "claude-sonnet-4-5-20250929"
 
+# The deprecated `AnthropicLLM` class has no LangSmith gateway support, so it
+# cannot authenticate when requests are routed through the gateway. Skip the
+# network-calling tests in that case; they still run against a direct
+# `ANTHROPIC_API_KEY`. Mirrors the gateway truthiness in `langchain_core`.
+_GATEWAY_ENABLED = (os.getenv("LANGSMITH_GATEWAY") or "").lower() not in (
+    "",
+    "false",
+    "0",
+    "no",
+)
+_skip_under_gateway = pytest.mark.skipif(
+    _GATEWAY_ENABLED,
+    reason="AnthropicLLM is deprecated and not compatible with the LangSmith gateway",
+)
+
 
 @pytest.mark.requires("anthropic")
 def test_anthropic_model_name_param() -> None:
@@ -24,6 +40,7 @@ def test_anthropic_model_param() -> None:
     assert llm.model == "foo"
 
 
+@_skip_under_gateway
 def test_anthropic_call() -> None:
     """Test valid call to anthropic."""
     llm = AnthropicLLM(model=MODEL)  # type: ignore[call-arg]
@@ -31,6 +48,7 @@ def test_anthropic_call() -> None:
     assert isinstance(output, str)
 
 
+@_skip_under_gateway
 def test_anthropic_streaming() -> None:
     """Test streaming tokens from anthropic."""
     llm = AnthropicLLM(model=MODEL)  # type: ignore[call-arg]
@@ -42,6 +60,7 @@ def test_anthropic_streaming() -> None:
         assert isinstance(token, str)
 
 
+@_skip_under_gateway
 def test_anthropic_streaming_callback() -> None:
     """Test that streaming correctly invokes on_llm_new_token callback."""
     callback_handler = FakeCallbackHandler()
@@ -56,6 +75,7 @@ def test_anthropic_streaming_callback() -> None:
     assert callback_handler.llm_streams > 1
 
 
+@_skip_under_gateway
 async def test_anthropic_async_generate() -> None:
     """Test async generate."""
     llm = AnthropicLLM(model=MODEL)  # type: ignore[call-arg]
@@ -63,6 +83,7 @@ async def test_anthropic_async_generate() -> None:
     assert isinstance(output, LLMResult)
 
 
+@_skip_under_gateway
 async def test_anthropic_async_streaming_callback() -> None:
     """Test that streaming correctly invokes on_llm_new_token callback."""
     callback_handler = FakeCallbackHandler()

libs/partners/openai/tests/integration_tests/chat_models/test_base.py

@@ -34,6 +34,26 @@ if TYPE_CHECKING:
 
 MAX_TOKEN_COUNT = 100
 
+# Whether requests route through the LangSmith gateway. Mirrors the truthiness
+# used by `langchain_core`'s gateway resolution.
+_GATEWAY_ENABLED = (os.environ.get("LANGSMITH_GATEWAY") or "").lower() not in (
+    "",
+    "false",
+    "0",
+    "no",
+)
+
+
+def _gateway_or_provider_key() -> str:
+    """Return an API key valid for the endpoint the base URL resolves to.
+
+    When the LangSmith gateway is enabled, requests route through it and must
+    authenticate with the gateway key rather than the provider key.
+    """
+    if _GATEWAY_ENABLED:
+        return os.environ["LANGSMITH_GATEWAY_API_KEY"]
+    return os.environ["OPENAI_API_KEY"]
+
 
 @pytest.mark.scheduled
 def test_chat_openai() -> None:
@@ -66,7 +86,7 @@ def test_chat_openai_model() -> None:
 
 
 def test_callable_api_key(monkeypatch: pytest.MonkeyPatch) -> None:
-    original_key = os.environ["OPENAI_API_KEY"]
+    original_key = _gateway_or_provider_key()
 
     calls = {"sync": 0}
 
@@ -83,7 +103,7 @@ def test_callable_api_key(monkeypatch: pytest.MonkeyPatch) -> None:
 
 
 async def test_callable_api_key_async(monkeypatch: pytest.MonkeyPatch) -> None:
-    original_key = os.environ["OPENAI_API_KEY"]
+    original_key = _gateway_or_provider_key()
 
     calls = {"sync": 0, "async": 0}
 
@@ -366,7 +386,10 @@ async def test_astream() -> None:
             assert full.usage_metadata["input_tokens"] > 0
             assert full.usage_metadata["output_tokens"] > 0
             assert full.usage_metadata["total_tokens"] > 0
-        else:
+        # The LangSmith gateway always emits a usage chunk regardless of
+        # `stream_options.include_usage`, so the opt-out assertions below only
+        # hold when not routing through it.
+        elif not _GATEWAY_ENABLED:
             assert chunks_with_token_counts == 0
             assert full.usage_metadata is None