fix(tests): unblock local_testing_part1 + litellm_router_testing on openai SDK 2.34+

[mateo-berri]

Summary

local_testing_part1 and litellm_router_testing have been consistently red on litellm_internal_staging since #27557 merged. That PR relaxed the openai pin from ==2.33.0 to >=2.20.0,<3.0.0, and uv.lock now resolves to openai==2.36.0. OpenAI Python SDK 2.34 introduced Admin API Key support and rewrote credential validation — the constructor now enforces credentials via an internal _enforce_credentials=True and rejects api_key="" with:

openai.OpenAIError: Missing credentials. Please pass an api_key, workload_identity, admin_api_key, or set the OPENAI_API_KEY or OPENAI_ADMIN_KEY environment variable.

(The workload_identity / admin_api_key / OPENAI_ADMIN_KEY names are the giveaway — none existed in 2.33.0.)

Three tests broke as a result; this PR fixes them in the tests, not in litellm itself, because each test was constructing a throwaway client whose api_key value never mattered.

What broke

• tests/local_testing/test_exceptions.py::test_exception_with_headers — all 14 parametrizations (openai/azure × sync/async × call_type). The test builds an openai.{Async,}OpenAI / openai.{Async,}AzureOpenAI with api_key="" purely as a placeholder; the actual .create is patched out with a 429 mock.
• tests/local_testing/test_router.py::test_router_dynamic_cooldown_correct_retry_after_time — same pattern, single test.
• tests/local_testing/test_exceptions.py::test_completion_perplexity_exception_on_openai_client — different: it deletes PERPLEXITYAI_API_KEY + OPENAI_API_KEY and expects openai.AuthenticationError with specific wording. With 2.34+ the SDK raises OpenAIError at construction (wrapped by LiteLLM as InternalServerError), and the message text changed.

What changed

• api_key="" → api_key="sk-test" in the five constructor sites. The upstream .create is mocked, so the value is ignored.
• test_completion_perplexity_exception_on_openai_client: catch any exception and assert the env-var name PERPLEXITY_API_KEY is in the message — that substring is stable across both old and new SDKs and is what the user-visible error contract is really about. Wrap the env-var restore in try/finally so the originals are always restored.

Verification

All 16 previously-failing tests now pass locally with AZURE_API_BASE + PERPLEXITYAI_API_KEY + OPENAI_API_KEY set (matching CI env):

============================== 16 passed in 6.29s ==============================

Test plan

• CI local_testing_part1 is green
• CI litellm_router_testing is green

---

Note

Low Risk
Low risk: changes are confined to test code to accommodate stricter OpenAI SDK credential validation; no production logic is modified.

Overview
Fixes failing local tests on OpenAI Python SDK 2.34+ by avoiding api_key="" placeholder clients and by loosening Perplexity missing-credentials assertions.

Tests now construct OpenAI/Azure clients with a dummy non-empty key ("sk-test"), and test_completion_perplexity_exception_on_openai_client uses pytest.raises(openai.OpenAIError) + try/finally env restoration, asserting the stable PERPLEXITY_API_KEY substring instead of exact error type/message.

^{Reviewed by Cursor Bugbot for commit c6aefe1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Greptile Summary

Test-only fixes to unblock two CI suites that started failing after the openai dependency was relaxed and the lockfile resolved to a newer version whose 2.34 release introduced stricter credential validation at client construction time.

• Replaces empty string api keys with a dummy placeholder in five OpenAI/AzureOpenAI client constructors; the actual network calls are patched with mock responses so the placeholder value is never used.
• Rewrites test_completion_perplexity_exception_on_openai_client to catch openai.OpenAIError (covering both the old AuthenticationError path and the new construction-time error wrapped by LiteLLM as InternalServerError), asserts a stable env-var name substring in the message, and moves env-var restoration to a try/finally block that was previously absent in some exception branches.

Confidence Score: 5/5

All changes are test-only; no production code is touched, and the dummy api key values are never sent over the wire since the underlying client calls are mocked.

Changes are narrowly scoped to three tests: swapping empty api keys for a dummy placeholder (mechanical, safe given mocking), and rewriting one perplexity test to broaden the caught exception type and fix env-var cleanup. No production logic is affected.

No files require special attention.

Important Files Changed

Filename	Overview
tests/local_testing/test_exceptions.py	Two fixes: replaces api_key="" with "sk-test" in five OpenAI/AzureOpenAI client constructors (calls are mocked); rewrites test_completion_perplexity_exception_on_openai_client to broaden exception to openai.OpenAIError, check stable PERPLEXITY_API_KEY substring, and restore env vars via try/finally.
tests/local_testing/test_router.py	Single-line fix: api_key="" replaced with "sk-test" for a throwaway OpenAI client whose .create call is fully mocked with a 429 response.

_{Reviews (2): Last reviewed commit: "fix(tests): narrow perplexity exception ..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai