fix(router): skip non-dict tool entries in the prompt caching affinity key instead of raising (#47565)

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/router_utils/prompt_caching_cache.py

@@ -94,8 +94,10 @@ def _is_ephemeral_mark(mark: object) -> bool:
     return mark_map.get("type") == "ephemeral"
 
 
-def _tool_carries_ephemeral_mark(tool: ChatCompletionToolParam) -> bool:
-    tool_map: Final = cast(  # cast-ok: TypedDict.get rejects keys the type does not declare
+def _tool_carries_ephemeral_mark(tool: object) -> bool:
+    if not isinstance(tool, dict):
+        return False
+    tool_map: Final = cast(  # cast-ok: tools are plain JSON dicts
         Mapping[str, object], tool
     )
     if _is_ephemeral_mark(tool_map.get("cache_control")):

tests/unit/router_utils/test_prompt_caching_cache.py

@@ -608,6 +608,23 @@ def test_unmarked_tools_change_the_key_marked_tools_do_not_share():
     assert key_a != key_b
 
 
+def test_a_non_dict_tool_entry_still_derives_a_key_instead_of_raising():
+    """
+    A client can send `tools: ["not-a-tool"]`. The provider rejects that with its own 400, so the
+    routing check has to key on what it was given and let the request through, never raise a 500.
+    """
+    messages = cast(
+        list[AllMessageValues],
+        [{"role": "user", "content": [{"type": "text", "text": "hi", "cache_control": {"type": "ephemeral"}}]}],
+    )
+    tools: Final = cast(list[ChatCompletionToolParam], ["not-a-tool", _marked_tool("read", {"type": "ephemeral"})])
+
+    positions = PromptCachingCache.prefix_positions(messages, tools)
+
+    assert [position.position for position in positions] == [0, 1]
+    assert PromptCachingCache.get_prompt_caching_cache_key(messages, tools) == positions[-1].cache_key
+
+
 def test_empty_and_missing_tools_seed_the_same_key():
     """`tools=[]` and no tools at all are the same request to the provider, so they must key alike."""
     messages = cast(