Self-improving AI Employees for everyone
Python ai-agentsai-employeeai-employeesassistantautonomous-agentsco-employeescompanygrok-botinstinctloop-agentsmuseown-your-datapersonalself-hosted
"""SDK transport failures use the same provider-neutral retry boundary."""
from types import SimpleNamespace
from unittest.mock import AsyncMock, MagicMock
import httpx
import pytest
from services.llm.protocol import LLMError, LLMErrorCategory, LLMResponse, Message
from services.llm.registry import ProviderSpec
from services.llm.unifier import ChatUnifier
from services.plugin import NodeUserError
@pytest.fixture
def dispatch(monkeypatch):
from services.llm import registry
client = SimpleNamespace(chat=AsyncMock(return_value=LLMResponse(content="done")),
fetch_models=AsyncMock(return_value=["test"]), aclose=AsyncMock())
monkeypatch.setitem(registry._REGISTRY, "transport-test", ProviderSpec(
name="transport-test", factory=lambda **kwargs: client,
sdk_exception_refs=("google.genai.errors:APIError",),
))
auth = MagicMock()
auth.resolve_api_key = AsyncMock(return_value=None)
return ChatUnifier(defaults={"providers": {}}, auth_service=auth, client_cache_size=0), client
@pytest.mark.asyncio
@pytest.mark.parametrize(("error_type", "category"), [
(httpx.ReadTimeout, LLMErrorCategory.TIMEOUT),
(httpx.ConnectError, LLMErrorCategory.CONNECTION),
(httpx.RemoteProtocolError, LLMErrorCategory.CONNECTION),
# An SDK's own deadline (1b.25): an OSError on Python 3.11+, never a bug.
(TimeoutError, LLMErrorCategory.TIMEOUT),
])
async def test_sdk_transport_failure_is_retryable_and_releases_client(dispatch, error_type, category):
unifier, client = dispatch
client.chat.side_effect = error_type("private endpoint details")
with pytest.raises(LLMError) as raised:
await unifier.chat(provider="transport-test", api_key="test", model="test",
messages=[Message(role="user", content="go")], translate_errors=False)
assert raised.value.category == category
assert raised.value.retryable
assert "private endpoint" not in raised.value.user_message
client.aclose.assert_awaited_once()
@pytest.mark.asyncio
async def test_model_listing_throttle_is_actionable_without_raw_http_error(dispatch):
unifier, client = dispatch
request = httpx.Request("GET", "https://private.example/models")
response = httpx.Response(429, headers={"retry-after": "30"}, request=request)
client.fetch_models.side_effect = httpx.HTTPStatusError("private endpoint details", request=request, response=response)
with pytest.raises(NodeUserError) as raised:
await unifier.fetch_models(provider="transport-test", api_key="test")
assert "rate-limiting" in str(raised.value)
assert "private endpoint" not in str(raised.value)
client.aclose.assert_awaited_once()
@pytest.mark.asyncio
async def test_programming_error_is_not_retried_as_a_transport_failure(dispatch):
unifier, client = dispatch
bug = ValueError("provider implementation bug")
client.chat.side_effect = bug
with pytest.raises(ValueError) as raised:
await unifier.chat(provider="transport-test", api_key="test", model="test", messages=[Message(role="user", content="go")])
assert raised.value is bug