🐛 Bug: inconsistent behavior when interrupting message streams between agent mode (using LangGraph Python) and direct LLM mode

♻️ Reproduction Steps

I've noticed inconsistent behavior when interrupting message streams between agent mode (using LangGraph Python) and direct LLM mode:

• Agent Mode (LangGraph Python):
  When I interrupt a streaming response and immediately send a new message, there's no response
  After waiting a while and sending another message, both the previously cancelled reply and the new reply appear together
  The interrupted message doesn't seem to be properly cancelled
  This can be reproduced using the Research Canvas demo: https://www.copilotkit.ai/examples/canvas-research

• Direct to LLM Mode:
  When I interrupt a streaming response and immediately send a new message, it responds immediately with the latest message
  Works as expected

✅ Expected Behavior

How can I make agent mode behave consistently with direct LLM mode when handling message interruptions? Is there a configuration or implementation pattern I should follow to ensure interrupted agent requests are properly cancelled and new messages are processed immediately?

❌ Actual Behavior

https://www.copilotkit.ai/examples/canvas-research

https://github.com/user-attachments/assets/987fabc4-fff3-4e5a-98b7-b41847cbe2b4

𝌚 CopilotKit Version

demo used https://www.copilotkit.ai/examples/canvas-research

pnpm list --depth=0 | findstr /i "@copilotkit"
@copilotkit/runtime 1.10.2
@copilotkit/runtime-client-gql 1.10.2
@copilotkit/shared 1.10.2

📄 Logs (Optional)

Start by reproducing the interruption sequence in the Research Canvas demo linked in the issue, using CopilotKit 1.10.2 and inspecting the @copilotkit/runtime, @copilotkit/runtime-client-gql, and @copilotkit/shared packages. Compare agent-mode streaming with direct-to-LLM mode and trace what happens after cancellation. Done means an interrupted agent response is cancelled and a new message is processed immediately without the old reply appearing later.