Reconstruct prior AG-UI turns as structured LlamaIndex ChatMessage objects.
Fed to AgentWorkflow.run(chat_history=...) so the model sees real tool calls and
tool results instead of the flattened {chat_history} text. Tool calls are carried
in additional_kwargs["tool_calls"] (OpenAI wire shape) rather than a
ToolCallBlock: the LiteLLM message adapter renders additional_kwargs verbatim
but does not serialize ToolCallBlock. Reasoning, when carried, is plain text in
content.
ag_ui_history_to_chat_messages(history: list[NormalizedHistoryMessage]) -> list[ChatMessage]
Convert normalized AG-UI history (from BaseAgent.history_messages) to
LlamaIndex ChatMessage objects, preserving tool calls and tool results.
Source code in datarobot_genai/llama_index/history.py
| def ag_ui_history_to_chat_messages(
history: list[NormalizedHistoryMessage],
) -> list[ChatMessage]:
"""Convert normalized AG-UI history (from ``BaseAgent.history_messages``) to
LlamaIndex ``ChatMessage`` objects, preserving tool calls and tool results.
"""
messages: list[ChatMessage] = []
for msg in history:
role = msg["role"]
content = msg.get("content") or ""
if role == "user":
messages.append(ChatMessage(role="user", content=content))
elif role == "assistant":
tool_calls = _to_openai_tool_calls(msg.get("tool_calls"))
additional_kwargs = {"tool_calls": tool_calls} if tool_calls else {}
messages.append(
ChatMessage(role="assistant", content=content, additional_kwargs=additional_kwargs)
)
elif role == "tool":
messages.append(
ChatMessage(
role="tool",
content=content,
additional_kwargs={"tool_call_id": str(msg.get("tool_call_id") or "")},
)
)
elif role == "system":
messages.append(ChatMessage(role="system", content=content))
return messages
|