fix(custom_httpx): don't resend request on connection error when retries are disabled

AsyncHTTPHandler.post/put/patch/delete retried once on a fresh client
whenever the first send raised ConnectError/RemoteProtocolError. That
resend sits beneath the user-facing retry budget, so max_retries=0 still
sent the request twice, which can double-bill non-idempotent requests.

Gate the new-session resend on the configured retry budget: when retries
are disabled (litellm.num_retries == 0) the original error propagates and
the request is sent at most once. Default behavior is unchanged.

Fixes #47824

litellm/llms/custom_httpx/http_handler.py

@@ -278,6 +278,10 @@ def _prepare_request_data_and_content(
     return request_data, request_content
 
 
+def _connection_error_resend_allowed() -> bool:
+    return litellm.num_retries is None or litellm.num_retries > 0
+
+
 # Cache for SSL contexts to avoid creating duplicate contexts with the same configuration
 # Key: tuple of (cafile, ssl_security_level, ssl_ecdh_curve)
 # Value: ssl.SSLContext
@@ -853,6 +857,8 @@ class AsyncHTTPHandler:
             response.raise_for_status()
             return response
         except httpx.ConnectError:
+            if not _connection_error_resend_allowed():
+                raise
             # Retry the request with a new session if there is a connection error
             new_client: Final = self.create_client(timeout=timeout, event_hooks=self.event_hooks)
             try:
@@ -923,6 +929,8 @@ class AsyncHTTPHandler:
             response.raise_for_status()
             return response
         except (httpx.RemoteProtocolError, httpx.ConnectError):
+            if not _connection_error_resend_allowed():
+                raise
             # Retry the request with a new session if there is a connection error
             new_client: Final = self.create_client(timeout=timeout, event_hooks=self.event_hooks)
             try:
@@ -995,6 +1003,8 @@ class AsyncHTTPHandler:
             response.raise_for_status()
             return response
         except httpx.ConnectError:
+            if not _connection_error_resend_allowed():
+                raise
             # Retry the request with a new session if there is a connection error
             new_client: Final = self.create_client(timeout=timeout, event_hooks=self.event_hooks)
             try:
@@ -1063,6 +1073,8 @@ class AsyncHTTPHandler:
             response.raise_for_status()
             return response
         except (httpx.RemoteProtocolError, httpx.ConnectError):
+            if not _connection_error_resend_allowed():
+                raise
             # Retry the request with a new session if there is a connection error
             new_client: Final = self.create_client(timeout=timeout, event_hooks=self.event_hooks)
             try:

tests/unit/llms/custom_httpx/test_http_handler.py

@@ -1546,6 +1546,34 @@ async def test_idempotent_disconnect_is_replayed(method: str) -> None:
         await handler.close()
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["post", "put", "patch", "delete"])
+async def test_connection_error_not_resent_when_retries_disabled(method: str, monkeypatch) -> None:
+    first_attempts: Final[list[httpx.Request]] = []  # mutable-ok: async closure capture buffer
+    resent: Final[list[httpx.Request]] = []  # mutable-ok: async closure capture buffer
+
+    async def raise_connection_error(request: httpx.Request) -> httpx.Response:
+        first_attempts.append(request)
+        raise httpx.ConnectError("connection reset after the request was written", request=request)
+
+    async def capture_resend(request: httpx.Request) -> httpx.Response:
+        resent.append(request)
+        return httpx.Response(200, request=request)
+
+    first: Final = httpx.AsyncClient(transport=httpx.MockTransport(raise_connection_error))
+    retry: Final = httpx.AsyncClient(transport=httpx.MockTransport(capture_resend))
+    monkeypatch.setattr(litellm, "num_retries", 0)
+    async with first, retry:
+        handler: Final = _RetryClientHandler(first=first, retry=retry)
+
+        with pytest.raises(httpx.ConnectError):
+            await getattr(handler, method)("https://api.example.com/runs/batch", content=b'{"a":1}')
+
+        assert len(first_attempts) == 1
+        assert resent == []
+        await handler.close()
+
+
 @pytest.fixture
 def forward_proxy_server():
     """Plain HTTP forward proxy that records the absolute URIs it is asked to fetch."""