fix(bedrock): raise_for_status in WrappedBotoClient so gateway errors surface (#95)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: vldcmp-uipath

packages/uipath_langchain_client/CHANGELOG.md

@@ -2,6 +2,11 @@
 
 All notable changes to `uipath_langchain_client` will be documented in this file.
 
+## [1.14.1] - 2026-06-23
+
+### Fixed
+- `WrappedBotoClient` (the httpx-backed Bedrock shim) now calls `raise_for_status()` in `converse`, `invoke_model`, and the streaming generator before reading the response. Previously a non-2xx gateway response (e.g. 403 License-not-available) was parsed as a normal result and handed to `langchain_aws`, which raised a misleading `ValueError("No 'output' key found in the response from the Bedrock Converse API ... misconfiguration of endpoint or region")` — the real status code and `detail` were lost. Gateway HTTP errors now surface as the patched `UiPathAPIError` subclass (e.g. `UiPathPermissionDeniedError`), matching the OpenAI and Vertex paths. For streaming responses the error body is read first so the typed exception retains its `detail`.
+
 ## [1.14.0] - 2026-06-15
 
 ### Added

packages/uipath_langchain_client/src/uipath_langchain_client/__version__.py

@@ -1,3 +1,3 @@
 __title__ = "UiPath LangChain Client"
 __description__ = "A Python client for interacting with UiPath's LLM services via LangChain."
-__version__ = "1.14.0"
+__version__ = "1.14.1"

packages/uipath_langchain_client/src/uipath_langchain_client/clients/bedrock/utils.py

@@ -57,6 +57,12 @@ def _stream_generator(
         if self.httpx_client is None:
             raise ValueError("httpx_client is not set")
         with self.httpx_client.stream("POST", "/", json=_serialize_bytes(request_body)) as response:
+            if response.is_error:
+                # The gateway returns a non-streamed JSON error body; read it so
+                # the patched raise_for_status surfaces it (with detail) instead
+                # of the EventStreamBuffer choking on a non-event payload.
+                response.read()
+            response.raise_for_status()
             buffer = EventStreamBuffer()
             for chunk in response.iter_bytes():
                 buffer.add_data(chunk)
@@ -71,12 +77,12 @@ def _stream_generator(
     def invoke_model(self, **kwargs: Any) -> Any:
         if self.httpx_client is None:
             raise ValueError("httpx_client is not set")
-        return {
-            "body": self.httpx_client.post(
-                "/",
-                json=json.loads(kwargs.get("body", "{}")),
-            )
-        }
+        response = self.httpx_client.post(
+            "/",
+            json=json.loads(kwargs.get("body", "{}")),
+        )
+        response.raise_for_status()
+        return {"body": response}
 
     def invoke_model_with_response_stream(self, **kwargs: Any) -> Any:
         return {"body": self._stream_generator(json.loads(kwargs.get("body", "{}")))}
@@ -90,7 +96,7 @@ def converse(
     ) -> Any:
         if self.httpx_client is None:
             raise ValueError("httpx_client is not set")
-        return self.httpx_client.post(
+        response = self.httpx_client.post(
             "/",
             json=_serialize_bytes(
                 {
@@ -99,7 +105,9 @@ def converse(
                     **params,
                 }
             ),
-        ).json()
+        )
+        response.raise_for_status()
+        return response.json()
 
     def converse_stream(
         self,

tests/langchain/clients/bedrock/test_wrapped_boto_client.py

@@ -0,0 +1,60 @@
+"""Unit tests for ``WrappedBotoClient`` HTTP error surfacing.
+
+The shim talks to the LLM Gateway over httpx instead of AWS. It must call
+``raise_for_status()`` so gateway HTTP errors (e.g. 403 License-not-available)
+propagate as exceptions, rather than being parsed as a normal result and then
+mis-reported downstream (langchain_aws raises a misleading "No 'output' key"
+``ValueError`` when the response lacks the expected fields).
+"""
+
+import json
+
+import httpx
+import pytest
+from uipath_langchain_client.clients.bedrock.utils import WrappedBotoClient
+
+_ERROR_BODY = {
+    "title": "License not available",
+    "status": 403,
+    "detail": "License not available for LLM usage.",
+}
+
+
+def _wrapped(handler: object) -> WrappedBotoClient:
+    transport = httpx.MockTransport(handler)  # type: ignore[arg-type]
+    return WrappedBotoClient(
+        httpx_client=httpx.Client(transport=transport, base_url="http://gateway")
+    )
+
+
+def test_converse_raises_on_http_error() -> None:
+    client = _wrapped(lambda request: httpx.Response(403, json=_ERROR_BODY))
+    with pytest.raises(httpx.HTTPStatusError):
+        client.converse(messages=[{"role": "user", "content": [{"text": "hi"}]}])
+
+
+def test_converse_returns_body_on_success() -> None:
+    payload = {"output": {"message": {"role": "assistant", "content": [{"text": "ok"}]}}}
+    client = _wrapped(lambda request: httpx.Response(200, json=payload))
+    assert client.converse(messages=[]) == payload
+
+
+def test_invoke_model_raises_on_http_error() -> None:
+    client = _wrapped(lambda request: httpx.Response(403, json=_ERROR_BODY))
+    with pytest.raises(httpx.HTTPStatusError):
+        client.invoke_model(body=json.dumps({"prompt": "hi"}))
+
+
+def test_converse_stream_raises_on_http_error() -> None:
+    # The generator defers work until iterated, so the error surfaces on consume.
+    client = _wrapped(lambda request: httpx.Response(403, json=_ERROR_BODY))
+    stream = client.converse_stream(messages=[])["stream"]
+    with pytest.raises(httpx.HTTPStatusError):
+        list(stream)
+
+
+def test_invoke_model_with_response_stream_raises_on_http_error() -> None:
+    client = _wrapped(lambda request: httpx.Response(403, json=_ERROR_BODY))
+    stream = client.invoke_model_with_response_stream(body=json.dumps({"prompt": "hi"}))["body"]
+    with pytest.raises(httpx.HTTPStatusError):
+        list(stream)