fix(controller): "agent task already running"

[xingyaoww]

• Fixes [Bug]: The Agent stuck when error occurred in loop. #1492.
• Also fixes: [Bug]: The Agent stuck when error occurred in loop. #1492 (comment) Will fix in a separate PR

[amanape]

Pausing or restarting the agent does not work. (responds with a "Current task state not recognized." message)

[xingyaoww]

@amanape should be fixed in the new commit!

[amanape]

Model: gpt-3.5-turbo
Agent: CodeActAgent

I can't seem to replicate this behavior every time. But the buttons don't seem react either now. I can't seem to pause, or restart. UI does not change. Anybody else can confirm?

[xingyaoww]

@amanape, this seems to be a much more complex issue that involves the front end. How about i revert the attempted fix for "Current task state not recognized", and starts a separate PR for this? I can reproduce some of the behavior (e.g., reset button doesn't work, pause state propagate to the backend but UI does not change, etc), but I guess this will take more time to figure out.

New PR for this: #1517

[rbren]

🤔 this seems like a legit error to me--in what case do we get a START without getting a cancel first?

[amanape]

@rbren There seem to be more issues with the agent actions than appear on the surface. I'm not sure if it was due to recent merges or they were dormant for a while.

For example, I managed to achieve a ValueError: Task (state: TaskState.PAUSED) is not in a state to add user message error than resulted in an unpaused state in the FE and paused state in the BE. This also somehow ended up rendering the <LoadPreviousSessionModal> every message after.

I'll take a peek into the FE later today and see if I can find anything. Worth setting up tests for these too

[xingyaoww]

@rbren This PR basically reverts changes I made at #1290. The reason I add that assert statement is we had a start_task_or_chat function, which we kinda disabled the ability to reset task.

Later your proposed PR broke this function down to start_task and some add_user_message function, but we forget to change the assert back :(

This is necessary in the sense that you need to kill existing task to be able to start a new one.