[jeffdupont]

Problem

Two tests in packages/client/tests/test_lifecycle.py fail locally when the whole suite runs, and pass on their own and in CI:

• TestInitClientSDKKeyPath::test_uses_options_sdk_key_over_env_var
• TestInitClientSDKKeyPath::test_retries_after_a_failed_init_instead_of_replaying_it

Both assert the exact Config(...) call, and got staging base_uri / stream_uri / events_uri keyword arguments they did not expect.

Cause

litellm/__init__.py calls load_dotenv() on import when LITELLM_MODE is DEV, its default. pytest imports the litellm package tests during collection, before any test runs. find_dotenv searches upward from the calling file, which is litellm's own install directory, so a .venv inside a checkout finds that checkout's .env. A .env that sets LD_BASE_URI, LD_STREAM_URI or LD_EVENTS_URI therefore puts them in os.environ for the rest of the run.

Neither failing test clears those variables. test_retries_after_a_failed_init_instead_of_replaying_it copies os.environ minus only LD_SDK_KEY, and test_uses_options_sdk_key_over_env_var patches os.environ without clearing it. init_client passes all three variables to Config.

CI has no .env, so it never sees this.

Checked by importing each provider SDK in a clean process: only litellm sets LD_BASE_URI.

Fix

An autouse clear_endpoint_env fixture in test_lifecycle.py deletes the three variables around each test with monkeypatch, the same way clear_endpoint_env in test_skills_fdv2.py already does for its two.

Verification

• Full suite with a parent .env setting the three variables: 2 failed before, 3714 passed, 18 skipped after (on the current main).
• make lint format-check: clean.

Not in this PR

litellm loading a .env on import also affects customers of launchdarkly-ai-litellm-*, not only tests: a project whose virtualenv sits inside it loads that project's .env, whatever the working directory. The SDK itself never loads one (test_never_loads_a_dotenv_file), but the dependency does. That is a separate question.

🤖 Generated with Claude Code