test(proxy): pin proxy_server.py non-route surface behavior (PR1)

[yuneng-berri]

Summary

PR1 of the proxy_server.py behavior-pinning project (Notion plan). Fills the seven PR1 placeholder files under tests/test_litellm/proxy/proxy_server/ (created by the harness PR #28827) with behavior pins for the non-route surface of litellm/proxy/proxy_server.py.

Pinned domains (101 symbols total, from the PR1 pin list):

• Lifecycle (8): proxy_startup_event, proxy_shutdown_event, _initialize_shared_aiohttp_session, cleanup_router_config_variables, save_worker_config, initialize, load_from_azure_key_vault, cost_tracking
• Misc helpers (5): check_request_disconnection, _resolve_typed_dict_type, _resolve_pydantic_type, get_litellm_model_info, run_ollama_serve
• OpenAPI customization (8): _generate_stable_operation_id, _strip_operation_id_method_suffix, ensure_unique_openapi_operation_ids, _inject_websocket_stubs_into_openapi_schema, get_openapi_schema, custom_openapi, mount_swagger_ui, _get_cors_config
• Exception handlers (4): openai_exception_handler, _close_dangling_otel_server_span, otel_request_validation_exception_handler, otel_unhandled_exception_handler
• Spend counters (15): get_current_spend, increment_spend_counters, update_cache, and 12 supporting helpers
• Background health (8): _get_process_rss_mb, _rss_mb_for_log, _run_direct_health_check_with_instrumentation, _schedule_background_health_check_db_save, _get_endpoint_exception_status, _write_health_state_to_router_cache, _adaptive_router_flusher_loop, _run_background_health_check
• Config scrubbers (3): _is_remote_module_url, _scrub_guardrail_inner, _scrub_db_overlay_remote_module_loads
• Streaming helpers (10): data_generator, async_data_generator, async_assistants_data_generator, select_data_generator, and 6 supporting helpers
• ProxyConfig class methods (40): __init__, is_yaml, _load_yaml_file, _get_config_from_file, _process_includes, save_config, _check_for_os_environ_vars, _get_team_config, load_team_config, _init_cache, switch_on_llm_response_caching, get_config, update_config_state, get_config_state, load_credential_list, parse_search_tools, _load_environment_variables, load_config, _init_non_llm_configs, _init_policy_engine, _load_alerting_settings, initialize_secret_manager, get_model_info_with_id, _delete_deployment, _add_deployment, decrypt_model_list_from_db, _update_llm_router, _add_callback_from_db_to_in_memory_litellm_callbacks, _add_callbacks_from_db_config, _encrypt_env_variables, _decrypt_and_set_db_env_variables, _decrypt_db_variables, _encrypt_env_variables_for_db, _parse_router_settings_value, _get_hierarchical_router_settings, _add_router_settings_from_db_config, _add_general_settings_from_db_config, _reschedule_spend_log_cleanup_job, _update_general_settings, _update_config_fields

Each pin has ≥1 happy-path test with a strong assertion (normalize(...) == {...}, .model_validate(...), or dict-eq with ≥3 keys) and ≥1 error-path test (raises, error-named, or 4xx/5xx).

Numbers

Gate	Target	This PR
Pin coverage (_pin_check.py)	all 101 pins	PASS (101 pins, 214 happy+error pairs)
Line coverage on proxy_server.py (new-tests-only)	≥25%	32.80%
Branch coverage (new-tests-only)	≥18%	20.91%
Runtime, -n 4	<3 min	~22 s (233 tests)

Test plan

• uv run pytest tests/test_litellm/proxy/proxy_server/ -n 4 — 233 passed, 0 failed
• python tests/test_litellm/proxy/proxy_server/_coverage_check.py --pr-target 1 — PASS (line 32.80% / 25%, branch 20.91% / 18%)
• python tests/test_litellm/proxy/proxy_server/_pin_check.py --list .pin_list.txt — PASS
• uv run black tests/test_litellm/proxy/proxy_server/ — clean

Out of scope

• Forwarding routes (PR2) and control-plane routes (PR3) — separate PRs branched off this one's merge per the plan's sequencing.
• Cleanup of pre-existing scattered tests — explicitly deferred (Decision return usage for all providers - As OpenAI does #11 in the plan).

---

Generated by Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fills seven placeholder test files under tests/test_litellm/proxy/proxy_server/ with 233 behavior-pin tests covering 101 symbols from the non-route surface of litellm/proxy/proxy_server.py. All changes are additive test code; no production source is modified.

• 101 symbols pinned across lifecycle, spend counters, OpenAPI customization, exception handlers, background health, config scrubbers, streaming helpers, and all 40+ ProxyConfig methods — each with at least one happy-path and one error-path assertion.
• Coverage gates met: 32.80% line coverage and 20.91% branch coverage on proxy_server.py from the new tests alone (targets were 25% / 18%); .gitignore gains entries for the two local artifact files produced by the pin-check and coverage scripts.

Confidence Score: 4/5

Safe to merge — all changes are additive test code with no production source touched. The two findings are minor test-quality issues with no runtime impact.

The PR adds 233 new tests with no modifications to existing tests or production code. One assertion in test_lifecycle.py is tautological (both sides of a param_count comparison use the same runtime expression), so it will never detect a signature regression on that axis. A second test in the same file spins for a real 1.2-second wall-clock window because asyncio.sleep is not patched, which could add latency or flakiness under CI load. Both are confined to a single test file and do not affect other tests or the production path.

tests/test_litellm/proxy/proxy_server/test_lifecycle.py — tautological param_count assertion and unmocked sleep timeout

Important Files Changed

Filename	Overview
tests/test_litellm/proxy/proxy_server/test_lifecycle.py	Replaces placeholder with 26 behavior-pin tests for lifecycle, helpers, and small utilities. Two minor issues: a tautological param_count assertion in the initialize signature test, and a real 1.2-second asyncio.wait_for timeout in the disconnection test.
tests/test_litellm/proxy/proxy_server/test_spend_counters.py	Replaces placeholder with 30 behavior-pin tests covering all 15 spend-counter symbols. Good use of helper factories and consistent assertion patterns; no issues found.
tests/test_litellm/proxy/proxy_server/test_proxy_config.py	Replaces placeholder with 80+ behavior-pin tests covering all 40+ ProxyConfig methods plus 3 module-level scrubbers. All external dependencies properly mocked; no issues found.
tests/test_litellm/proxy/proxy_server/test_background_health.py	Replaces placeholder with behavior-pin tests for all 8 background health-check helpers. asyncio.sleep is correctly patched in every loop test; no issues found.
tests/test_litellm/proxy/proxy_server/test_exception_handlers.py	Replaces placeholder with behavior-pin tests for the 4 exception handlers. Well-structured; tests both happy path and edge cases such as empty error lists and missing OTEL spans.
tests/test_litellm/proxy/proxy_server/test_openapi_customization.py	Replaces placeholder with behavior-pin tests for all 8 OpenAPI/CORS helpers. Fresh FastAPI instances are used in place of the global app, keeping tests fully isolated.
tests/test_litellm/proxy/proxy_server/test_streaming_helpers.py	Replaces placeholder with behavior-pin tests for all 10 streaming helper symbols. Real async generators are used correctly; all mocks are scoped to the function under test.
.gitignore	Adds .pin_list.txt and .cov_new.xml to gitignore — local artifacts from the pin-check and coverage scripts.

_{Reviews (1): Last reviewed commit: "test(proxy): pin proxy_server.py non-rou..." | Re-trigger Greptile}

[greptile-apps]

The param_count entry in both the observed dict and the expected dict uses len(sig.parameters) — the same expression — so the assertion param_count == param_count is always true regardless of the actual parameter count. If initialize gains or loses parameters this test will not catch it. Replace the right-hand side with a hard-coded integer that reflects the current signature.

Suggested change

	observed = {
	"is_async": inspect.iscoroutinefunction(initialize),
	"param_count": len(sig.parameters),
	"has_model": "model" in sig.parameters,
	"has_config": "config" in sig.parameters,
	}
	assert normalize(observed) == {
	"is_async": True,
	"param_count": len(sig.parameters),
	"has_model": True,
	"has_config": True,
	}
	expected_param_count = len(sig.parameters) # pin the exact count here
	observed = {
	"is_async": inspect.iscoroutinefunction(initialize),
	"param_count": len(sig.parameters),
	"has_model": "model" in sig.parameters,
	"has_config": "config" in sig.parameters,
	}
	assert normalize(observed) == {
	"is_async": True,
	"param_count": expected_param_count,
	"has_model": True,
	"has_config": True,
	}

[greptile-apps]

Unmocked 1.2-second wall-clock wait per run

asyncio.sleep is not patched here, so check_request_disconnection spins through real sleeps until asyncio.wait_for cancels it after 1.2 s. This adds real elapsed time to every suite run and can be flaky under CPU pressure on CI. Consider patching asyncio.sleep to AsyncMock(return_value=None) at the start of the test and lowering the wait_for timeout to something like 0.05 s; the test would still verify the TimeoutError path without blocking for over a second.