Merge pull request #39306 from BerriAI/litellm_deflake_20260902

test: deflake JWT tamper, fuzzy picker, tag routing, liveliness, redis stall burst, and pre-commit interrupt tests

scripts/pre_commit_lint.sh

@@ -142,6 +142,8 @@ fi
 
 lint_dashboard() {
     (
+        trap 'exit 143' TERM
+        trap 'rm -f "${report:-}"' EXIT
         rc=0
         prettier_rel=()
         eslint_rel=()
@@ -168,7 +170,6 @@ EOF
         report=$(mktemp)
         npx eslint . -f json -o "$report" || true
         node scripts/check-lint-budgets.mjs "$report" eslint-budgets.json || rc=1
-        rm -f "$report"
         exit $rc
     )
 }

tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_credentials.py

@@ -5,6 +5,7 @@ from datetime import datetime, timedelta, timezone
 import pytest
 from cryptography.hazmat.primitives import serialization
 from cryptography.hazmat.primitives.asymmetric import rsa
+from jwt.utils import base64url_decode, base64url_encode
 from pydantic import SecretStr
 
 from litellm.proxy._experimental.mcp_server.outbound_credentials.bridge_credentials import (
@@ -52,6 +53,12 @@ def _refresh_token() -> str:
     return minted.token.get_secret_value()
 
 
+def _corrupt_signature(token: str) -> str:
+    unsigned, signature = token.rsplit(".", 1)
+    raw = base64url_decode(signature)
+    return f"{unsigned}.{base64url_encode(bytes((raw[0] ^ 0x01,)) + raw[1:]).decode()}"
+
+
 def test_kdf_is_deterministic_and_key_length_is_256_bit():
     again = session_keys_from_master_key(MASTER_KEY)
     assert again.signing_key.get_secret_value() == KEYS.signing_key.get_secret_value()
@@ -109,8 +116,7 @@ def test_resolve_fails_expired_token_closed_and_flags_expiry():
 
 def test_resolve_fails_tampered_token_closed_without_expiry_flag():
     token = _access_token()
-    tampered = token[:-2] + ("aa" if not token.endswith("aa") else "bb")
-    result = resolve_session_bearer(f"Bearer {tampered}", KEYS, NOW)
+    result = resolve_session_bearer(f"Bearer {_corrupt_signature(token)}", KEYS, NOW)
     assert isinstance(result, SessionBearerInvalid)
     assert result.expired is False
 

tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_session_token.py

@@ -6,6 +6,7 @@ import jwt
 import pytest
 from cryptography.hazmat.primitives import serialization
 from cryptography.hazmat.primitives.asymmetric import rsa
+from jwt.utils import base64url_decode, base64url_encode
 from pydantic import SecretStr, ValidationError
 
 from litellm.proxy._experimental.mcp_server.outbound_credentials.session_token import (
@@ -66,6 +67,12 @@ def _mint_refresh() -> str:
     return minted.token.get_secret_value()
 
 
+def _corrupt_signature(token: str) -> str:
+    unsigned, signature = token.rsplit(".", 1)
+    raw = base64url_decode(signature)
+    return f"{unsigned}.{base64url_encode(bytes((raw[0] ^ 0x01,)) + raw[1:]).decode()}"
+
+
 def _sign_claims(payload: dict, prefix: str = SESSION_TOKEN_PREFIX, keys: SessionKeys = KEYS) -> str:
     return prefix + jwt.encode(payload, keys.signing_key.get_secret_value(), algorithm="HS256")
 
@@ -138,8 +145,7 @@ def test_still_valid_one_second_before_expiry():
 
 def test_tampered_signature_is_bad_signature():
     token = _mint_access()
-    tampered = token[:-2] + ("aa" if not token.endswith("aa") else "bb")
-    assert isinstance(open_session_token(tampered, KEYS, NOW), SessionBadSignature)
+    assert isinstance(open_session_token(_corrupt_signature(token), KEYS, NOW), SessionBadSignature)
 
 
 def test_key_rotation_invalidates_outstanding_tokens():
@@ -329,8 +335,7 @@ def test_rs256_tampered_signature_is_bad_signature():
     minted = mint_session_token(PRINCIPAL, RSA_KEYS, NOW)
     assert isinstance(minted, MintedSessionToken)
     token = minted.token.get_secret_value()
-    tampered = token[:-2] + ("aa" if not token.endswith("aa") else "bb")
-    assert isinstance(open_session_token(tampered, RSA_KEYS, NOW), SessionBadSignature)
+    assert isinstance(open_session_token(_corrupt_signature(token), RSA_KEYS, NOW), SessionBadSignature)
 
 
 def test_rs256_expired_token_is_expired():
@@ -413,8 +418,7 @@ def test_rotation_window_still_enforces_expiry_and_tamper_on_the_previous_key():
     )
     after = NOW + timedelta(seconds=SESSION_TTL_SECONDS + 1)
     assert isinstance(open_session_token(token, rotated, after), SessionExpired)
-    tampered = token[:-2] + ("aa" if not token.endswith("aa") else "bb")
-    assert isinstance(open_session_token(tampered, rotated, NOW), SessionBadSignature)
+    assert isinstance(open_session_token(_corrupt_signature(token), rotated, NOW), SessionBadSignature)
 
 
 def test_weak_or_garbage_private_key_pem_rejected_at_construction():

tests/test_litellm/proxy/client/cli/autoroute/test_wizard.py

@@ -1,5 +1,5 @@
 import asyncio
-from typing import Any, Dict, List, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 from unittest.mock import patch
 
 import click
@@ -7,7 +7,8 @@ import pytest
 import yaml
 from click.testing import CliRunner
 from InquirerPy.base.control import Choice
-from prompt_toolkit.application import create_app_session
+from InquirerPy.prompts.fuzzy import InquirerPyFuzzyControl
+from prompt_toolkit.application import AppSession, create_app_session
 from prompt_toolkit.input import create_pipe_input
 from prompt_toolkit.output import DummyOutput
 
@@ -283,27 +284,45 @@ class TestRunConfigureWizardNotInteractive:
         assert not config_path.exists()
 
 
+def _highlighted_choice(session: AppSession) -> Optional[str]:
+    if session.app is None:
+        return None
+    controls = [c for c in session.app.layout.find_all_controls() if isinstance(c, InquirerPyFuzzyControl)]
+    if not controls or controls[0].choice_count == 0:
+        return None
+    return controls[0].selection["name"]
+
+
+async def _wait_until_highlighted(session: AppSession, name: str) -> None:
+    async def _poll() -> None:
+        while _highlighted_choice(session) != name:
+            await asyncio.sleep(0.01)
+
+    await asyncio.wait_for(_poll(), timeout=5)
+
+
 def _drive_fuzzy_pick(
     models: Tuple[DiscoveredModel, ...],
     prompt_label: str,
     multiselect: bool,
-    key_events: List[Tuple[str, float]],
+    key_events: List[Tuple[str, Optional[str]]],
 ) -> List[str]:
     """Drives the real InquirerPy fuzzy prompt through prompt_toolkit's own test input/output,
     exercising the actual widget (filtering, tab-to-toggle, enter-to-confirm) rather than mocking
     it away. asyncio.to_thread propagates the create_app_session context into the worker thread
-    running _fuzzy_pick's synchronous .execute() call."""
+    running _fuzzy_pick's synchronous .execute() call. Each key event names the choice the widget
+    must highlight before the next key is sent (None sends the next key immediately)."""
 
     async def _run() -> List[str]:
         with create_pipe_input() as pipe_input:
-            with create_app_session(input=pipe_input, output=DummyOutput()):
+            with create_app_session(input=pipe_input, output=DummyOutput()) as session:
                 task = asyncio.ensure_future(
                     asyncio.to_thread(wizard_module._fuzzy_pick, models, prompt_label, multiselect)
                 )
-                await asyncio.sleep(0.05)
-                for text, delay in key_events:
+                for text, highlighted in key_events:
                     pipe_input.send_text(text)
-                    await asyncio.sleep(delay)
+                    if highlighted is not None:
+                        await _wait_until_highlighted(session, highlighted)
                 return await task
 
     return asyncio.run(_run())
@@ -315,13 +334,13 @@ class TestFuzzyPickWidget:
 
     def test_single_select_filters_and_returns_highlighted_match(self):
         result = _drive_fuzzy_pick(
-            self._models(), "test", multiselect=False, key_events=[("model-13", 0.3), ("\r", 0.1)]
+            self._models(), "test", multiselect=False, key_events=[("model-13", "model-13"), ("\r", None)]
         )
         assert result == ["model-13"]
 
     def test_multiselect_requires_tab_to_toggle_before_enter(self):
         result = _drive_fuzzy_pick(
-            self._models(), "test", multiselect=True, key_events=[("model-7", 0.3), ("\t", 0.1), ("\r", 0.1)]
+            self._models(), "test", multiselect=True, key_events=[("model-7", "model-7"), ("\t", None), ("\r", None)]
         )
         assert result == ["model-7"]
 
@@ -331,12 +350,12 @@ class TestFuzzyPickWidget:
             "test",
             multiselect=True,
             key_events=[
-                ("model-3", 0.3),
-                ("\t", 0.1),
-                *[("\x7f", 0.02) for _ in range("model-3".__len__())],
-                ("model-15", 0.3),
-                ("\t", 0.1),
-                ("\r", 0.1),
+                ("model-3", "model-3"),
+                ("\t", None),
+                ("\x7f" * len("model-3"), None),
+                ("model-15", "model-15"),
+                ("\t", None),
+                ("\r", None),
             ],
         )
         assert set(result) == {"model-3", "model-15"}

tests/test_litellm/proxy/health_endpoints/test_health_endpoints.py

@@ -1,6 +1,7 @@
 import asyncio
 import json
 import time
+from typing import Final
 from datetime import datetime, timedelta
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
@@ -1190,27 +1191,23 @@ def test_health_liveliness_endpoint(proxy_client):
     Test that /health/liveliness endpoint returns 200 OK with "I'm alive!" message.
     This is a critical orchestration endpoint that must be simple and fast.
     """
-    # Measure the time taken for the health check call
-    start_time = time.perf_counter()
+    warm_up: Final = proxy_client.get("/health/liveliness")
+    assert warm_up.status_code == 200, f"Expected 200 OK, got {warm_up.status_code}: {warm_up.text}"
 
-    # Make GET request to /health/liveliness
-    response = proxy_client.get("/health/liveliness")
+    def _timed_poll() -> tuple[float, httpx.Response]:
+        start_time: Final = time.perf_counter()
+        response: Final = proxy_client.get("/health/liveliness")
+        return (time.perf_counter() - start_time) * 1000, response
 
-    end_time = time.perf_counter()
-    duration_ms = (end_time - start_time) * 1000
+    polls: Final = tuple(_timed_poll() for _ in range(5))
 
-    # Assert response status
-    assert response.status_code == 200, f"Expected 200 OK, got {response.status_code}: {response.text}"
+    for _, response in polls:
+        assert response.status_code == 200, f"Expected 200 OK, got {response.status_code}: {response.text}"
+        assert response.json() == "I'm alive!", f"Expected 'I'm alive!' message, got: {response.json()}"
 
-    # Assert response content (FastAPI JSON-encodes the string)
-    assert response.json() == "I'm alive!", f"Expected 'I'm alive!' message, got: {response.json()}"
-
-    # Verify response is fast (should be < 100ms for a simple endpoint)
-    # This is critical for orchestration systems that poll frequently
-    assert duration_ms < 100, f"Health check took {duration_ms:.2f}ms, expected < 100ms for a simple endpoint"
-
-    # Log the duration for visibility (useful for CI/CD monitoring)
-    print(f"\n/health/liveliness response time: {duration_ms:.2f}ms")
+    durations_ms: Final = tuple(sorted(duration_ms for duration_ms, _ in polls))
+    median_ms: Final = durations_ms[len(durations_ms) // 2]
+    assert median_ms < 100, f"Median of {len(polls)} health checks took {median_ms:.2f}ms, expected < 100ms"
 
 
 def test_health_liveness_endpoint(proxy_client):

tests/test_litellm/router_strategy/test_router_tag_routing.py

@@ -2,15 +2,27 @@
 # This tests litellm router
 
 
-import pytest
-
 import logging
+from typing import Final
 
+import pytest
 
 import litellm
 from litellm._logging import verbose_logger
 
 
+async def _routed_model_ids(
+    router: litellm.Router, tags: list[str], remaining: frozenset[str], attempts: int = 100
+) -> frozenset[str]:
+    if not remaining or attempts == 0:
+        return frozenset()
+    response: Final = await router.acompletion(
+        model="gpt-4", messages=[{"role": "user", "content": "hi"}], metadata={"tags": tags}, mock_response="hi"
+    )
+    seen: Final = frozenset({response._hidden_params["model_id"]})
+    return seen | await _routed_model_ids(router, tags, remaining - seen, attempts - 1)
+
+
 @pytest.mark.asyncio()
 async def test_router_free_paid_tier():
     """
@@ -850,17 +862,10 @@ async def test_negation_regex_pattern_treated_as_literal():
 
     # The regex-like string matches no deployment tag literally, so all
     # candidates survive and both model IDs are reachable.
-    seen_ids = set()
-    for _ in range(10):
-        response = await router.acompletion(
-            model="gpt-4",
-            messages=[{"role": "user", "content": "hi"}],
-            metadata={"tags": ["!provider:(anthropic|openai)"]},
-            mock_response="hi",
-        )
-        seen_ids.add(response._hidden_params["model_id"])
+    expected: Final = frozenset({"anthropic-model", "openai-model"})
+    routed_ids: Final = await _routed_model_ids(router, ["!provider:(anthropic|openai)"], expected)
 
-    assert seen_ids == {"anthropic-model", "openai-model"}
+    assert routed_ids == expected
 
 
 @pytest.mark.asyncio()
@@ -1281,17 +1286,10 @@ async def test_chain_enable_tag_filtering_false_overrides_router_level_true():
         enable_tag_filtering=True,
     )
 
-    seen_ids = set()
-    for _ in range(10):
-        response = await router.acompletion(
-            model="gpt-4",
-            messages=[{"role": "user", "content": "hi"}],
-            metadata={"tags": ["teamA"]},
-            mock_response="hi",
-        )
-        seen_ids.add(response._hidden_params["model_id"])
+    expected: Final = frozenset({"team-a-deployment", "team-b-deployment"})
+    routed_ids: Final = await _routed_model_ids(router, ["teamA"], expected)
 
-    assert seen_ids == {"team-a-deployment", "team-b-deployment"}
+    assert routed_ids == expected
 
 
 @pytest.mark.asyncio()

tests/test_litellm/test_pre_commit_lint.py

@@ -53,6 +53,12 @@ case "$*" in
     "eslint --no-warn-ignored"*)
         [ "${STUB_FAIL:-}" = "eslint" ] && exit 1
         ;;
+    "eslint . -f json"*)
+        if [ -n "${STUB_HANG_DIR:-}" ]; then
+            touch "$STUB_HANG_DIR/eslint_report.started"
+            sleep 60
+        fi
+        ;;
 esac
 exit 0
 """
@@ -340,11 +346,12 @@ def test_interrupt_kills_background_jobs_and_removes_logs(tmp_path: Path) -> Non
     )
     try:
         assert _wait_until((hang_dir / "make.started").exists, 10)
+        assert _wait_until((hang_dir / "eslint_report.started").exists, 10)
         os.killpg(proc.pid, signal.SIGINT)
         assert proc.wait(timeout=10) != 0
         make_pid = int((hang_dir / "make.pid").read_text())
         assert _wait_until(lambda: _pid_gone(make_pid), 5)
-        assert list(tmp_dir.iterdir()) == []
+        assert _wait_until(lambda: not any(tmp_dir.iterdir()), 5), list(tmp_dir.iterdir())
     finally:
         with suppress(ProcessLookupError, PermissionError):
             os.killpg(proc.pid, signal.SIGTERM)