fix(client/async): correctly retry in all cases (#704)

src/anthropic/_base_client.py

@@ -1591,7 +1591,7 @@ async def _request(
         except Exception as err:
             log.debug("Encountered Exception", exc_info=True)
 
-            if retries_taken > 0:
+            if remaining_retries > 0:
                 return await self._retry_request(
                     input_options,
                     cast_to,

tests/test_client.py

@@ -10,6 +10,7 @@
 import tracemalloc
 from typing import Any, Union, cast
 from unittest import mock
+from typing_extensions import Literal
 
 import httpx
 import pytest
@@ -873,7 +874,14 @@ def test_retrying_status_errors_doesnt_leak(self, respx_mock: MockRouter) -> Non
     @pytest.mark.parametrize("failures_before_success", [0, 2, 4])
     @mock.patch("anthropic._base_client.BaseClient._calculate_retry_timeout", _low_retry_timeout)
     @pytest.mark.respx(base_url=base_url)
-    def test_retries_taken(self, client: Anthropic, failures_before_success: int, respx_mock: MockRouter) -> None:
+    @pytest.mark.parametrize("failure_mode", ["status", "exception"])
+    def test_retries_taken(
+        self,
+        client: Anthropic,
+        failures_before_success: int,
+        failure_mode: Literal["status", "exception"],
+        respx_mock: MockRouter,
+    ) -> None:
         client = client.with_options(max_retries=4)
 
         nb_retries = 0
@@ -882,6 +890,8 @@ def retry_handler(_request: httpx.Request) -> httpx.Response:
             nonlocal nb_retries
             if nb_retries < failures_before_success:
                 nb_retries += 1
+                if failure_mode == "exception":
+                    raise RuntimeError("oops")
                 return httpx.Response(500)
             return httpx.Response(200)
 
@@ -1840,8 +1850,13 @@ async def test_retrying_status_errors_doesnt_leak(self, respx_mock: MockRouter)
     @mock.patch("anthropic._base_client.BaseClient._calculate_retry_timeout", _low_retry_timeout)
     @pytest.mark.respx(base_url=base_url)
     @pytest.mark.asyncio
+    @pytest.mark.parametrize("failure_mode", ["status", "exception"])
     async def test_retries_taken(
-        self, async_client: AsyncAnthropic, failures_before_success: int, respx_mock: MockRouter
+        self,
+        async_client: AsyncAnthropic,
+        failures_before_success: int,
+        failure_mode: Literal["status", "exception"],
+        respx_mock: MockRouter,
     ) -> None:
         client = async_client.with_options(max_retries=4)
 
@@ -1851,6 +1866,8 @@ def retry_handler(_request: httpx.Request) -> httpx.Response:
             nonlocal nb_retries
             if nb_retries < failures_before_success:
                 nb_retries += 1
+                if failure_mode == "exception":
+                    raise RuntimeError("oops")
                 return httpx.Response(500)
             return httpx.Response(200)