fix: restore original cache control as it can be used complementary to new auto strategy

Author: philipph-askui

src/askui/models/shared/tools.py

@@ -20,6 +20,7 @@
 
 from askui.models.shared.agent_message_param import (
     Base64ImageSourceParam,
+    CacheControlEphemeralParam,
     ContentBlockParam,
     ImageBlockParam,
     TextBlockParam,
@@ -394,7 +395,12 @@ def to_params(self) -> list[ToolParam]:
             for tool_name, tool in tool_map.items()
             if self._include is None or tool_name in self._include
         }
-        return list(filtered_tool_map.values())
+        result = list(filtered_tool_map.values())
+        if result:
+            result[-1]["cache_control"] = CacheControlEphemeralParam(
+                type="ephemeral",
+            )
+        return result
 
     def _get_mcp_tool_params(self) -> dict[str, ToolParam]:
         if not self._mcp_client:

src/askui/models/shared/truncation_strategies.py

@@ -5,7 +5,9 @@
 from typing_extensions import override
 
 from askui.models.shared.agent_message_param import (
+    CacheControlEphemeralParam,
     MessageParam,
+    TextBlockParam,
     ToolParam,
 )
 from askui.models.shared.prompts import ActSystemPrompt
@@ -173,8 +175,37 @@ def _should_truncate(self) -> bool:
     @property
     @override
     def messages(self) -> list[MessageParam]:
+        self._move_cache_control_to_last_non_tool_result_user_message()
         return self._messages
 
+    def _move_cache_control_to_last_non_tool_result_user_message(self) -> None:
+        found_last = False
+        for message in reversed(self._messages):
+            if message.role == "user" and not _is_tool_result_user_message(message):
+                if not found_last:
+                    found_last = True
+                    if isinstance(message.content, str):
+                        message.content = [
+                            TextBlockParam(
+                                text=message.content,
+                                cache_control=CacheControlEphemeralParam(
+                                    type="ephemeral",
+                                ),
+                            )
+                        ]
+                    elif len(message.content) > 0:
+                        last_content = message.content[-1]
+                        if hasattr(last_content, "cache_control"):
+                            last_content.cache_control = CacheControlEphemeralParam(
+                                type="ephemeral",
+                            )
+                else:
+                    if isinstance(message.content, list) and message.content:
+                        last_content = message.content[-1]
+                        if hasattr(last_content, "cache_control"):
+                            last_content.cache_control = None
+                    break
+
     def _truncate(self) -> None:  # noqa: C901
         messages_to_remove_min = min(
             len(self._messages) - self._max_messages_after_truncation, 0