test(callbacks): unwind the callbacks global the policy engine and realtime tests scaffold around

[yuneng-berri]

TLDR

Problem this solves:

• Two files set litellm.callbacks by hand, 41 times
• Some wrap the body in try/finally, most do not
• 59 TQ005 violations and the scaffolding around them

How it solves it:

• One monkeypatch.setattr per test, no try/finally
• The writes stop sitting outside the block restoring them

User Flow

No end-user behavior changes. A proxy admin running a guardrail policy through
https://litellm-domain/v1/chat/completions, or a realtime session through
https://litellm-domain/v1/realtime, gets the same steps executed in the same
order; these are the test files around that code, not the code

Relevant issues

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Shared setup: run the file, read the global back in the same interpreter, then
count what the quality rules see in it.

uv run python -c '
import litellm, pytest
before = list(litellm.callbacks)
pytest.main(["tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py", "-q", "-p", "no:randomly"])
print("litellm.callbacks after the run:", "restored" if list(litellm.callbacks) == before else "LEAKED")
'
for f in $FILES; do uv run python scripts/check_test_quality.py $f; done | grep -oE "TQ00[0-9]" | sort | uniq -c

Before (ff02d5c)

1. The first command

16 passed, 2 warnings in 2.10s
litellm.callbacks after the run: restored

2. The second command

   1 TQ003
  59 TQ005

3. uv run python scripts/test_quality_gate.py

OK: every TQ rule is within its test-suite ceiling (base origin/litellm_internal_staging)

After (5fc0a0f)

1. The first command

16 passed, 2 warnings in 1.90s
litellm.callbacks after the run: restored

1b. uv run pytest $FILES -q -p no:randomly

212 passed, 10 warnings in 64.15s (0:01:04)

2. The second command

3. uv run python scripts/test_quality_gate.py

OK: every TQ rule is within its test-suite ceiling (base origin/litellm_internal_staging)

Type

🧹 Refactoring

✅ Test

Caveats (if any)

• Both sides read "restored", so no live leak is fixed here
• litellm.callbacks is already in the conftest snapshot
• The point is these tests stop depending on it, which WS1 wants
• Line rewrapping in the diff is ruff on blocks that lost an indent
• $FILES is the two test files this PR touches
• Touches test-quality-budget.json, as do test: drop the cwd-relative sys.path.insert calls from the test suite #37802, test: unwind environment writes in tests/test_litellm with monkeypatch #37806, test(cost-calc): stop 182 global writes leaking out of the cost-calc suites #37815 and test(mcp): retire the last file of the dead tests/litellm mirror #37824

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

The PR refactors policy-engine and realtime-streaming tests to isolate litellm.callbacks through pytest’s monkeypatch fixture instead of manual assignment and cleanup.

• Removes repetitive callback snapshot and try/finally scaffolding.
• Removes an unnecessary test-local Python path modification.
• Ratchets the TQ003 and TQ005 quality budgets to reflect the reduced violations.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
test-quality-budget.json	Lowers TQ003 and TQ005 ceilings to match the test-quality improvements reported and verified in the PR evidence.
tests/test_litellm/litellm_core_utils/test_realtime_streaming.py	Replaces direct callback writes and manual cleanup with function-scoped monkeypatch restoration without changing test assertions.
tests/test_litellm/proxy/policy_engine/test_pipeline_executor.py	Removes callback snapshot and try/finally scaffolding while preserving both pipeline scenarios and their behavioral assertions.

_{Reviews (3): Last reviewed commit: "style(realtime): wrap the one signature ..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yuneng-berri]

@greptileai

[yuneng-berri]

@greptileai