Independent Auditing of AI Agents. Run by human or the agent itself, to answer the most crucial question in the AI Agent Economy. Is the agent doing what is supposed to do? With iFixAi you can have this answer in less than 120 seconds.
Python agent-evaluationaiai-alignmentai-evaluationai-governanceai-safetyclidiagnostic-tooleu-ai-acthallucination-detectioniso-42001llm-evaluationllm-securitynist-ai-rmfowasp-llmprompt-injectionpythonresponsible-airisk-assessmentrisk-management
fix(gemini): isolate concurrent SDK credentials per request (#190)
Rudy Mizrahi Celekli0a87cd670e72
diff --git a/ifixai/providers/gemini.py b/ifixai/providers/gemini.pyindex 034f69d..db3bedf 100644--- a/ifixai/providers/gemini.py+++ b/ifixai/providers/gemini.py@@ -1,6 +1,8 @@import asyncio+import osimport google.generativeai as genai+from google.ai.generativelanguage import GenerativeServiceAsyncClientfrom google.api_core import exceptions as google_exceptionsfrom ifixai.core.types import ChatMessage, ProviderConfig@@ -31,8 +33,6 @@ class GeminiProvider(ChatProvider):messages: list[ChatMessage],config: ProviderConfig,) -> str:- genai.configure(api_key=config.api_key)-model_name = config.model or DEFAULT_MODELendpoint = config.endpoint or "https://generativelanguage.googleapis.com"@@ -67,80 +67,92 @@ class GeminiProvider(ChatProvider):attempts = config.max_retries + 1backoff = INITIAL_BACKOFF_SECONDS- for attempt in range(attempts):- try:- response = await asyncio.wait_for(- model.generate_content_async(contents),- timeout=float(config.timeout),- )+ # GenerativeModel lazily resolves a process-global client. Bind a+ # request-owned client before scheduling the generation so another SUT+ # or judge call cannot replace this request's API key.+ async with GenerativeServiceAsyncClient(+ client_options={+ "api_key": (+ config.api_key or os.environ.get("GEMINI_API_KEY")+ or os.environ.get("GOOGLE_API_KEY")+ ),+ },+ ) as native_client:+ model._async_client = native_client+ for attempt in range(attempts):+ try:+ response = await asyncio.wait_for(+ model.generate_content_async(contents),+ timeout=float(config.timeout),+ )- if not response.candidates:- raise ProviderResponseError(+ if not response.candidates:+ raise ProviderResponseError(+ provider="gemini",+ endpoint=endpoint,+ details="No candidates in response",+ )++ candidate = response.candidates[0]+ text_parts = [+ part.text+ for part in candidate.content.parts+ if hasattr(part, "text") and part.text+ ]+ if not text_parts:+ raise ProviderEmptyContentError(+ provider="gemini",+ endpoint=endpoint,+ details="No text parts in response candidate",+ )+ return "\n".join(text_parts)++ except asyncio.TimeoutError as exc:+ raise ProviderTimeoutError(provider="gemini",endpoint=endpoint,- details="No candidates in response",- )-- candidate = response.candidates[0]- text_parts = [- part.text- for part in candidate.content.parts- if hasattr(part, "text") and part.text- ]- if not text_parts:- raise ProviderEmptyContentError(+ details=f"Request timed out after {config.timeout}s",+ ) from exc+ except google_exceptions.Unauthenticated as exc:+ raise ProviderAuthError(provider="gemini",endpoint=endpoint,- details="No text parts in response candidate",- )- return "\n".join(text_parts)-- except asyncio.TimeoutError as exc:- raise ProviderTimeoutError(- provider="gemini",- endpoint=endpoint,- details=f"Request timed out after {config.timeout}s",- ) from exc- except google_exceptions.Unauthenticated as exc:- raise ProviderAuthError(- provider="gemini",- endpoint=endpoint,- details=str(exc),- ) from exc- except google_exceptions.PermissionDenied as exc:- raise ProviderAuthError(- provider="gemini",- endpoint=endpoint,- details=str(exc),- ) from exc- except google_exceptions.ResourceExhausted as exc:- if attempt < attempts - 1:- await asyncio.sleep(backoff)- backoff *= BACKOFF_MULTIPLIER- continue- raise ProviderRateLimitError(- provider="gemini",- endpoint=endpoint,- details=str(exc),- ) from exc- except google_exceptions.ServiceUnavailable as exc:- raise ProviderConnectionError(- provider="gemini",- endpoint=endpoint,- details=str(exc),- ) from exc- except google_exceptions.GoogleAPIError as exc:- raise ProviderResponseError(- provider="gemini",- endpoint=endpoint,- details=str(exc),- ) from exc-- raise ProviderRateLimitError(- provider="gemini",- endpoint=endpoint,- details="Exhausted all retry attempts",- )+ details=str(exc),+ ) from exc+ except google_exceptions.PermissionDenied as exc:+ raise ProviderAuthError(+ provider="gemini",+ endpoint=endpoint,+ details=str(exc),+ ) from exc+ except google_exceptions.ResourceExhausted as exc:+ if attempt < attempts - 1:+ await asyncio.sleep(backoff)+ backoff *= BACKOFF_MULTIPLIER+ continue+ raise ProviderRateLimitError(+ provider="gemini",+ endpoint=endpoint,+ details=str(exc),+ ) from exc+ except google_exceptions.ServiceUnavailable as exc:+ raise ProviderConnectionError(+ provider="gemini",+ endpoint=endpoint,+ details=str(exc),+ ) from exc+ except google_exceptions.GoogleAPIError as exc:+ raise ProviderResponseError(+ provider="gemini",+ endpoint=endpoint,+ details=str(exc),+ ) from exc++ raise ProviderRateLimitError(+ provider="gemini",+ endpoint=endpoint,+ details="Exhausted all retry attempts",+ )def _format_messages(diff --git a/ifixai/tests/test_gemini_client_isolation.py b/ifixai/tests/test_gemini_client_isolation.pynew file mode 100644index 0000000..ea3ad4d--- /dev/null+++ b/ifixai/tests/test_gemini_client_isolation.py@@ -0,0 +1,109 @@+"""Concurrent Gemini callers must retain their own request credentials."""++import asyncio++import pytest++from ifixai.core.types import ChatMessage, ProviderConfig+from ifixai.providers.base import ProviderConnectionError++[email protected]("keys", [("owned-key-A", "owned-key-B"), ("owned-key-A", "owned-key-A")])+async def test_concurrent_gemini_requests_keep_their_own_api_key(monkeypatch, keys):+ genai = pytest.importorskip("google.generativeai")+ glm = pytest.importorskip("google.ai.generativelanguage")+ from google.generativeai import client++ from ifixai.providers import gemini++ native_clients = []++ def request_client(**kwargs):+ # Construct the actual installed async SDK client, including API-key+ # credentials. Substitute only the final outbound generation call.+ native = glm.GenerativeServiceAsyncClient(**kwargs)+ native_clients.append(native)++ async def generate_content(request, **options):+ await asyncio.sleep(0)+ api_key = native.transport._credentials.token+ return glm.GenerateContentResponse(candidates=[glm.Candidate(+ content=glm.Content(parts=[glm.Part(text=api_key)]),+ finish_reason=glm.Candidate.FinishReason.STOP,+ )])++ monkeypatch.setattr(native, "generate_content", generate_content)+ return native++ def lazy_global_client():+ return request_client(client_options=client._client_manager.client_config["client_options"])++ monkeypatch.setattr(client, "get_default_generative_async_client", lazy_global_client)+ monkeypatch.setattr(gemini, "GenerativeServiceAsyncClient", request_client, raising=False)+ provider = gemini.GeminiProvider()+ try:+ results = await asyncio.gather(*[+ provider.send_message([ChatMessage(content="hello")], ProviderConfig(provider="gemini", api_key=key, max_retries=0))+ for key in keys+ ])+ assert results == list(keys)+ finally:+ for native in native_clients:+ await native.transport.close()+ assert genai is not None++[email protected]("exit_kind", ["success", "error", "cancel"])+async def test_request_owned_sdk_client_closes_on_every_exit(monkeypatch, exit_kind):+ pytest.importorskip("google.generativeai")+ glm = pytest.importorskip("google.ai.generativelanguage")+ grpc = pytest.importorskip("grpc")+ from google.api_core.exceptions import ServiceUnavailable+ from google.generativeai import client++ from ifixai.providers import gemini++ native_clients = []+ entered = asyncio.Event()++ def request_client(**kwargs):+ native = glm.GenerativeServiceAsyncClient(**kwargs)+ native_clients.append(native)++ async def generate_content(request, **options):+ if exit_kind == "error":+ raise ServiceUnavailable("owned transport unavailable")+ if exit_kind == "cancel":+ entered.set()+ await asyncio.Event().wait()+ return glm.GenerateContentResponse(candidates=[glm.Candidate(+ content=glm.Content(parts=[glm.Part(text="owned reply")]),+ finish_reason=glm.Candidate.FinishReason.STOP,+ )])++ monkeypatch.setattr(native, "generate_content", generate_content)+ return native++ monkeypatch.setattr(client, "get_default_generative_async_client", lambda: request_client(client_options=client._client_manager.client_config["client_options"]))+ monkeypatch.setattr(gemini, "GenerativeServiceAsyncClient", request_client, raising=False)+ config = ProviderConfig(provider="gemini", api_key="owned-key-A", max_retries=0)+ try:+ call = gemini.GeminiProvider().send_message([ChatMessage(content="hello")], config)+ if exit_kind == "cancel":+ task = asyncio.create_task(call)+ await asyncio.wait_for(entered.wait(), timeout=5)+ task.cancel()+ with pytest.raises(asyncio.CancelledError):+ await task+ elif exit_kind == "error":+ with pytest.raises(ProviderConnectionError):+ await call+ else:+ assert await call == "owned reply"+ assert len(native_clients) == 1+ assert native_clients[0].transport.grpc_channel.get_state() == grpc.ChannelConnectivity.SHUTDOWN+ finally:+ for native in native_clients:+ await native.transport.close()