scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-06 02:48:13 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 103

docs(e2e): state what the replay lane does and does not prove about provider credentials

Browse source 
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
mateo 2026-09-22 01:11:07 +00:00
parent 09f01a8445
commit d00a39c79b
1 changed files with 1 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
tests/e2e/CONTRIBUTING.md
View file

@ -146,7 +146,7 @@ Tests marked `@pytest.mark.e2e` hard-fail when no proxy answers `/health/livelin
E2E_MATRIX_PROVIDERS=openai,anthropic E2E_MATRIX_AUTH_MODES=env_ref uv run pytest tests/e2e/llm_translation/test_endpoint_matrix_e2e.py
```
Providers with a generic edge mount (OpenAI and Anthropic today) are edge-wired and marked `replayable`, so their cells join the scheduled record/replay lane and run with zero provider egress on weekdays; the rest hit their provider live in every fixture mode. The module registers the deployments its selected cells need as one `register_models` batch in a module fixture, so a run pays the data-plane reload and propagation budget once rather than once per cell, which is what keeps a 60-cell replay in the tens of seconds
Providers with a generic edge mount (OpenAI and Anthropic today) are edge-wired and marked `replayable`, so their cells join the scheduled record/replay lane and run with zero provider egress on weekdays; the rest hit their provider live in every fixture mode. The two lanes prove different things: replay proves routing and response shape, since the edge keys fixtures on the request body and never on the provider auth header, so only the live and record lanes prove the proxy sent the deployment's own secret, and even there an `inline` or `stored_credential` cell whose value equals the runner's `os.environ/` value cannot tell the two resolutions apart. The module registers the deployments its selected cells need as one `register_models` batch in a module fixture, so a run pays the data-plane reload and propagation budget once rather than once per cell, which is what keeps a 60-cell replay in the tens of seconds
## What a complete test looks like