mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-12 03:48:56 +00:00
d5c9d17d47
Motivation: _probe_upstream_auth predicts whether a true_passthrough MCP server will accept the caller's request before a stateful session opens, so LiteLLM can return a 401 up front instead of masking the rejection as an empty tool list. The probe only ever sent Accept and Authorization upstream. When an upstream identifies the caller through a different header the admin already configured in server.extra_headers (a gateway-issued platform header, for example), the probe always looked anonymous to it, so it answered 401 for every caller including ones with a valid session, and every connect got the pre-session challenge regardless of whether the caller was actually authorized. Approach: Added _extra_headers_for_probe, which takes the caller's raw request headers and the server's configured extra_headers list and returns just the values for those header names (case-insensitively matched), always excluding authorization since the probe already sets that separately. _probe_upstream_auth gained an extra_headers parameter that merges into the outgoing probe request. The is_true_passthrough branch in _raise_preemptive_401_for_unauthenticated_servers now passes the caller's configured extra headers into the probe, using the raw_headers parameter that function already receives. The is_oauth_passthrough / _check_passthrough_upstream_auth path is untouched, matching the issue's stated scope. Validation: PYTHONPATH=. pytest tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py -v passes all 240 tests (235 existing + 5 new). Reverting only the server.py change (git stash) makes the 5 new/changed tests fail, including one that raises the exact HTTPException: 401 the issue describes, confirming they are a real regression guard rather than a structural check. ruff check, ruff format --check, and basedpyright were run on both changed files: no new basedpyright errors were introduced in server.py (236 before and after), and the small increase in the test file (reportPrivateUsage / reportAny noise from calling internal helpers directly, matching this test file's existing whitebox pattern) stays well under the repo's basedpyright-code-budget.json limits for those rules. Disclosure: dependencies were installed into a local virtualenv rather than through the maturin/editable install this repo normally uses, because of disk-space limits in this environment; litellm was imported directly from source with PYTHONPATH, and the Rust bridge already no-ops gracefully when its compiled extension is absent. No live MCP server was reachable to reproduce the issue end to end; the fix and its tests are demonstrated at the unit level described above. Report: https://github.com/BerriAI/litellm/issues/40099 Signed-off-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com> Assisted-by: claude-sonnet-5 (via Claude Code) |
||
|---|---|---|
| .. | ||
| mcp_server | ||
| __init__.py | ||