Copy cached chat completion results

python/packages/autogen-ext/src/autogen_ext/models/cache/_chat_completion_cache.py

@@ -274,14 +274,12 @@ async def create(
         if cached_result is not None:
             if isinstance(cached_result, CreateResult):
                 # Cache hit from previous non-streaming call
-                cached_result.cached = True
-                return cached_result
+                return cached_result.model_copy(update={"cached": True})
             elif isinstance(cached_result, list):
                 # Cache hit from previous streaming call - extract the final CreateResult
                 for item in reversed(cached_result):
                     if isinstance(item, CreateResult):
-                        item.cached = True
-                        return item
+                        return item.model_copy(update={"cached": True})
                 # If no CreateResult found in list, fall through to make actual call
 
         result = await self.client.create(
@@ -325,12 +323,13 @@ async def _generator() -> AsyncGenerator[Union[str, CreateResult], None]:
                     # Cache hit from previous streaming call
                     for result in cached_result:
                         if isinstance(result, CreateResult):
-                            result.cached = True
-                        yield result
+                            yield result.model_copy(update={"cached": True})
+                        else:
+                            yield result
                     return
                 elif isinstance(cached_result, CreateResult):
                     # Cache hit from previous non-streaming call - convert to streaming format
-                    cached_result.cached = True
+                    cached_result = cached_result.model_copy(update={"cached": True})
 
                     # If content is a non-empty string, yield it as a streaming chunk first
                     if isinstance(cached_result.content, str) and cached_result.content:

python/packages/autogen-ext/tests/models/test_chat_completion_cache.py

@@ -58,6 +58,19 @@ async def test_cache_basic_with_args() -> None:
     assert response2.content == responses[2]
 
 
+@pytest.mark.asyncio
+async def test_cache_hit_does_not_mutate_original_result() -> None:
+    _, prompts, system_prompt, _, cached_client = get_test_data(num_messages=1)
+    messages = [system_prompt, UserMessage(content=prompts[0], source="user")]
+
+    original = await cached_client.create(messages)
+    cached = await cached_client.create(messages)
+
+    assert not original.cached
+    assert cached.cached
+    assert original is not cached
+
+
 @pytest.mark.asyncio
 async def test_cache_structured_output_with_args() -> None:
     responses, prompts, system_prompt, _, cached_client = get_test_data(num_messages=4)