[Bug]: `Message.to_chat_dict` emits `cache_control` when caching is disabled

Is there an existing issue for the same bug?

• I have searched existing issues and this is not a duplicate.

Bug Description

In openhands-sdk/openhands/sdk/llm/message.py, Message.to_chat_dict can emit a cache_control marker even when called with cache_enabled=False.

When list serialization is selected for another reason, such as function_calling_enabled=True, _list_serializer processes existing cache_prompt=True content without receiving or checking cache_enabled.

For a tool message, the marker is promoted to the top-level message dictionary despite caching being disabled.

Expected Behavior

Given a tool message containing TextContent(cache_prompt=True), calling to_chat_dict with cache_enabled=False and function_calling_enabled=True should return the serialized tool message without any cache marker:

{
    "content": [{"type": "text", "text": "Tool response"}],
    "role": "tool",
    "tool_call_id": "call_123",
    "name": "test_tool",
}

cache_enabled=False should suppress both message-level and content-level cache_control fields, even when another feature requires list serialization.

Actual Behavior

Message.to_chat_dict returns:

{
    "content": [{"type": "text", "text": "Tool response"}],
    "role": "tool",
    "cache_control": {"type": "ephemeral"},
    "tool_call_id": "call_123",
    "name": "test_tool",
}

The returned message contains a top-level cache marker even though cache_enabled is False.

Steps to Reproduce

1. Check out Software Agent SDK main at commit 007721b3d2bfccd1469f8008a514af13b7ae1b71.

2. Add the following test to tests/sdk/llm/test_message.py:

def test_message_tool_role_omits_cache_prompt_when_cache_disabled():
    from openhands.sdk.llm.message import Message, TextContent

    message = Message(
        role="tool",
        content=[TextContent(text="Tool response", cache_prompt=True)],
        tool_call_id="call_123",
        name="test_tool",
    )

    result = message.to_chat_dict(
        **{
            **DEFAULT_SERIALIZATION_OPTS,
            "cache_enabled": False,
            "function_calling_enabled": True,
        }
    )

    assert "cache_control" not in result
    assert "cache_control" not in result["content"][0]

3. Run:

uv run pytest tests/sdk/llm/test_message.py -q

4. Observe that the newly added assertion fails while the 30 existing tests pass.

Acceptance Criteria

• cache_enabled=False prevents list serialization from emitting message-level or content-level cache_control fields.
• Disabling caching remains effective when list serialization is selected by function calling or vision support.
• Existing cache_enabled=True behavior for tool and non-tool messages remains unchanged.
• Regression coverage includes a tool message with cache_prompt=True, cache_enabled=False, and function_calling_enabled=True.

Installation Method

Source checkout; focused pytest reproduction after the documented development setup

If you selected "Other", please specify

No response

SDK Version

main@007721b3d2bfccd1469f8008a514af13b7ae1b71 (openhands-sdk 1.42.1)

Version Confirmation

• I have confirmed this bug exists on the LATEST version of OpenHands SDK

Python Version

3.13.2

Model Name (if applicable)

Not applicable; reproduced by directly invoking deterministic message-serialization code

Operating System

MacOS

Logs and Error Messages

tests/sdk/llm/test_message.py .....F.........................            [100%]

FAILED tests/sdk/llm/test_message.py::test_message_tool_role_omits_cache_prompt_when_cache_disabled
AssertionError: assert 'cache_control' not in {'cache_control': {'type': 'ephemeral'}, 'content':
[{'text': 'Tool response', 'type': 'text'}], 'name': 'test_tool', 'role': 'tool', ...}

1 failed, 30 passed in 0.14s

Minimal Code Sample

from openhands.sdk.llm.message import Message, TextContent

message = Message(
    role="tool",
    content=[TextContent(text="Tool response", cache_prompt=True)],
    tool_call_id="call_123",
    name="test_tool",
)

result = message.to_chat_dict(
    cache_enabled=False,
    vision_enabled=False,
    function_calling_enabled=True,
    force_string_serializer=False,
    send_reasoning_content=False,
)

print(result)
# {
#     'content': [{'type': 'text', 'text': 'Tool response'}],
#     'role': 'tool',
#     'cache_control': {'type': 'ephemeral'},
#     'tool_call_id': 'call_123',
#     'name': 'test_tool',
# }

Screenshots and Additional Context

At the tested commit, to_chat_dict selects list serialization when any of caching, vision, or function calling is enabled:

if not force_string_serializer and (
    cache_enabled or vision_enabled or function_calling_enabled
):
    message_dict = self._list_serializer(vision_enabled=vision_enabled)

However, _list_serializer does not receive cache_enabled. It calls item.to_llm_dict(), which emits cache_control whenever the content object's cache_prompt field is true:

item_dicts = item.to_llm_dict()

For tool messages, it then unconditionally records that caching is requested and promotes the marker to the message level:

if self.role == "tool" and item.cache_prompt:
    role_tool_with_prompt_caching = True
    for d in item_dicts:
        d.pop("cache_control", None)

The public LLM formatting path passes the current cache setting into to_chat_dict. APIBasedCritic also calls this method with cache_enabled=False and function_calling_enabled=True. Consequently, a message already carrying cache_prompt=True can emit a cache marker through these serialization paths even though caching is explicitly disabled.

One possible fix direction is to pass cache_enabled into _list_serializer and use it to gate both content-level cache markers and the tool-message promotion logic. Existing behavior when cache_enabled=True should remain unchanged.

Start in openhands-sdk/openhands/sdk/llm/message.py by tracing Message.to_chat_dict into _list_serializer, then run the focused reproduction in tests/sdk/llm/test_message.py with pytest. Done means the tool-message case omits both cache_control fields when caching is disabled, while existing enabled-cache behavior and the surrounding tests remain passing.