Merge remote-tracking branch 'origin/nate_messages_native_during_call_mask' into nate_messages_integration_local

# Conflicts:
#	litellm/messages/dispatch.py

litellm/messages/dispatch.py

@@ -1,3 +1,4 @@
+import asyncio
 import inspect
 from collections.abc import AsyncIterator, Awaitable, Callable, Coroutine, Iterator, Mapping
 from types import MappingProxyType
@@ -199,6 +200,9 @@ async def _native_amessages(
     args: tuple[object, ...],
     kwargs: Mapping[str, object],
 ) -> MessagesResult:
+    # Python's handler leaves the event loop (run_in_executor) before it builds the provider body, so
+    # tasks scheduled next to the call, such as the proxy's during_call guardrails, edit the request first
+    await asyncio.sleep(0)
     # Python's async @client wrapper stamps these on every failure; the proxy reads `timeout` for
     # the x-litellm-timeout header and the router reads `num_retries`.
     try:

tests/test_litellm_rust/messages/test_requests.py

@@ -1,3 +1,4 @@
+import asyncio
 from typing import Final
 
 import pytest
@@ -146,3 +147,20 @@ async def test_optional_positional_body_parameters_reach_the_wire(server: Record
     assert response["id"] == MESSAGES_RESPONSE["id"]
     body, _ = sent(server)
     assert (body["metadata"], body["system"], body["temperature"]) == (args[3], args[6], args[7])
+
+
+@pytest.mark.asyncio
+async def test_in_place_edits_from_a_task_started_before_the_call_reach_the_wire(server: RecordingServer) -> None:
+    messages: Final = [{"role": "user", "content": "during about secret"}]
+
+    async def mask() -> None:
+        messages[0]["content"] = "during about [REDACTED]"
+
+    async def during_call() -> None:
+        await asyncio.gather(mask())
+
+    guardrail: Final = asyncio.create_task(during_call())
+    await asyncio.gather(guardrail, call(server, messages=messages))
+
+    body, _ = sent(server)
+    assert body["messages"] == [{"role": "user", "content": "during about [REDACTED]"}]