Fix/update controller is_stuck()

[enyst]

Fix the stuck check in the controller:

• increase to 4 historical tuples, but prevent 4 x (any action, any observation)
• add detection for a pattern where every other tuple is the same (e.g. -1, -3, -5 are the same, -2, -4, -6 are the same) - source https://opendevin.slack.com/archives/C06U8UTKSAD/p1715691234611019
• add filtering on history, to ignore MessageAction with source=USER, and still detect the rest.

Please note:

• I think CmdOutputObservation and maybe the KillAction should ignore the command_id for the purpose of this check. I made an override of eq for the first, but I reverted from this PR to give it some thought. Any feedback is appreciated!

[xingyaoww]

LGTM!

I think CmdOutputObservation and maybe the KillAction should ignore the command_id for the purpose of this check. I made an override of eq for the first, but I reverted from this PR to give it some thought. Any feedback is appreciated!

I agree on this point -- we are only interested in comparing the "content" of the observation, but not something else. Happy to hear what other people think

[li-boxuan]

I think it would be interesting to run this PR against ./evaluation/swe_bench/scripts/run_infer.sh [llm-config-group] CodeActAgent 1. This command will run the 1st task in swe-bench-lite. This first task exhausted all 50 turns (and cost me ~$6) but seems OpenDevin was just repeating itself and wasting money...

FYI log is here:
instance_django__django-15202.log
lol damn this log contains my API key... we should definitely fix that. Anyways, I have revoked my key so don't worry about that 🤣

[rbren]

I was about to say "it'd be great to get some unit tests in here" and then I scrolled down a bit 😄

Thanks for this!