_separate_tool_calls corrupts conversation history when multiple MCP tool calls require approval #5293
Description
System Info
NA
Information
- The official example scripts
- My own modified scripts
🐛 Describe the bug
In StreamingResponseOrchestrator._separate_tool_calls(), when the model returns a single choice containing multiple MCP tool calls that all require approval (pending or denied), the method calls next_turn_messages.pop() once per tool call inside the inner loop. The first pop() correctly removes the assistant message appended, but every subsequent pop() deletes an unrelated message from the conversation history.
Error logs
NA
Expected behavior
Regardless of how many tool calls in a single choice require approval, only the one assistant message appended for that choice should be removed.