From b8f38829af90901d294f48950a555e6bd8bdf06c Mon Sep 17 00:00:00 2001
From: Charles Cheng <charlescheng@rezona.ai>
Date: Fri, 19 Jun 2026 18:10:58 +0800
Subject: [PATCH] refactor: address Greptile P2 review items for cache_control
 /v1/messages fix

- Clarify apply_to_anthropic_messages_request() Returns: docstring: the
  no-injection early-return path returns original references (no deep copy),
  while the injection path returns deep-copied values; the previous wording
  overstated the no-copy path's mutation safety.
- Add explanatory note in _supports_native_anthropic_messages() docstring
  explaining why get_llm_provider() is called unconditionally even when
  custom_llm_provider is set: routing prefixes like converse/ make
  naive prefix-stripping unsafe (wrong native-support answer for Bedrock
  Converse models), so the optimisation is deferred to a follow-up.
- NON_CACHEABLE_TOOL_TYPES in anthropic_cache_control_hook is already
  imported from litellm.constants.ANTHROPIC_NON_CACHEABLE_TOOL_TYPES
  (shared with transformation.py), so no divergence risk.
---
 .../anthropic_cache_control_hook.py             | 17 ++++++++---------
 .../messages/handler.py                         |  7 +++++++
 2 files changed, 15 insertions(+), 9 deletions(-)

diff --git a/litellm/integrations/anthropic_cache_control_hook.py b/litellm/integrations/anthropic_cache_control_hook.py
index 3424eb456ff..e3b833b4be4 100644
--- a/litellm/integrations/anthropic_cache_control_hook.py
+++ b/litellm/integrations/anthropic_cache_control_hook.py
@@ -10,6 +10,7 @@ import copy
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union, cast
 
 from litellm._logging import verbose_logger
+from litellm.constants import ANTHROPIC_NON_CACHEABLE_TOOL_TYPES
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.integrations.custom_prompt_management import CustomPromptManagement
 from litellm.integrations.prompt_management_base import PromptManagementClient
@@ -31,13 +32,9 @@ else:
 # breakpoints: "A maximum of 4 blocks with cache_control may be provided."
 MAX_CACHE_CONTROL_BLOCKS = 4
 
-# Tool types Anthropic rejects `cache_control` on. Mirrors the exclusion in
-# litellm/llms/anthropic/chat/transformation.py so the native /v1/messages path
-# never marks a tool the API would 400 on.
-NON_CACHEABLE_TOOL_TYPES = (
-    "tool_search_tool_regex_20251119",
-    "tool_search_tool_bm25_20251119",
-)
+# Re-export for any existing internal callers while the canonical definition
+# lives in litellm.constants (shared with transformation.py).
+NON_CACHEABLE_TOOL_TYPES = ANTHROPIC_NON_CACHEABLE_TOOL_TYPES
 
 
 class AnthropicCacheControlHook(CustomPromptManagement):
@@ -287,8 +284,10 @@ class AnthropicCacheControlHook(CustomPromptManagement):
                 downstream provider transforms still receive them.
 
         Returns:
-            A tuple of (messages, system, tools) with cache control applied. The
-            inputs are deep-copied, so the caller's originals are never mutated.
+            A tuple of (messages, system, tools). When there are no injection
+            points, the original references are returned as-is (no copy is
+            made). When injection occurs, the inputs are deep-copied before
+            modification so the caller's originals are not mutated.
         """
         injection_points: list[CacheControlInjectionPoint] = non_default_params.pop(
             "cache_control_injection_points", []
diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/handler.py b/litellm/llms/anthropic/experimental_pass_through/messages/handler.py
index 86e2c66df6a..6b13be94268 100644
--- a/litellm/llms/anthropic/experimental_pass_through/messages/handler.py
+++ b/litellm/llms/anthropic/experimental_pass_through/messages/handler.py
@@ -197,6 +197,13 @@ def _supports_native_anthropic_messages(
     provider selected purely by endpoint (e.g. an ``/anthropic`` ``api_base``)
     resolves the same way it will in the handler, instead of being misread as a
     fallback and having its cache injection skipped.
+
+    Note: ``get_llm_provider`` is called unconditionally here even when
+    ``custom_llm_provider`` is already set. Skipping it is non-trivial because
+    model strings may carry routing prefixes (e.g. ``converse/``) that affect
+    both the resolved model name and whether ``get_provider_anthropic_messages_config``
+    returns a native config — plain prefix-stripping produces wrong answers for
+    those cases. Left as a potential optimisation for a follow-up.
     """
     from litellm.types.utils import LlmProviders