Threads for completed agent runs can be resumed

[TanGentleman]

When executing a graph in tangents, we have two cases for agent behavior:

1. Runs until completion. Guarantee that state['is_done']==True.
2. Runs until interrupted. Guarantee that thread can be resumed with a new user input.

Applications invoking this graph should not attempt to resume a graph that has already been completed (Case 1). Here is an example where this causes erroneous behavior:

INFO:__main__:Continuing session: 765d52ad-880d-4f7e-9713-0c421f810a66
INFO:root:Executing action: ActionType.GENERATE
Stream callback: True
INFO:openai._base_client:Retrying request to /chat/completions in 0.450344 seconds
INFO:openai._base_client:Retrying request to /chat/completions in 0.982073 seconds
INFO:root:{'success': False, 'data': None, 'error': 'Connection error.'}
INFO:root:Saved action: ActionType.GENERATE
ERROR:root:Action failed, marking task as failed
INFO:root:Saving task chat_task to convex
INFO:root:Saved failed task
WARNING:root:Saved failed tasks: ['chat_task']
INFO:root:No tasks remaining!
INFO:__main__:Workflow completed.
INFO:__main__:Workflow completed.
WARNING:__main__:Empty response received.
INFO:__main__:Removed incomplete exchange.
INFO:__main__:Continuing session: 765d52ad-880d-4f7e-9713-0c421f810a66
WARNING:__main__:Empty response received.
INFO:__main__:Removed incomplete exchange.

User input that leads to an invalid response is popped off. These cases should be avoided.

Proposed solution:

• Remove session_id from the gradio state when a graph is run to completion.

[TanGentleman]

Solving this one isn't too hard! The important flow is this:

1. UI is rendered based on chat_history referenced by gradio components.
2. User submits input
3. UI re-rendered with user message, empty assistant message
4. Graph is executed, with our custom callback streaming to UI as assistant message (mutating chat_history directly)

Issue: The handle submit function (generate_assistant_response) requires the status of the run.
Solution: We now propagate a workflow_completed bool so the session state is updated alongside the other gradio components as outputs to this function.

Effect: After a run is completed (backend), existing chat history (UI) is passed to new graph as expected with in a new thread/session ID.