add action message

[filip-michalsky]

why

This is second, reworked attempt.

Paul mentioned a vision where stagehand will be a tool we can pass to an LLM to perform an action on the web. In that sense, we DO want to return success/ failure signal and a reason for the failure so that the agent/developer can heal their pipelines (either manually or agentically).

what changed

we are returning a message from a finished act() method. This is currently not utilized in evals so we need to add evals to test it.

test plan

need to add evals for this.

[filip-michalsky]

My thinking behind this updated API design:

By return a message from act() we do not break anything in existing evals. It's a simply optional return value which developer can use to retry actions and perhaps change the prompt until the action gets done. It can also help with the prompt optimizer where an another LLM can change the action prompt if previous action failed, etc.

[filip-michalsky]

This needs to get fixed - I resolved the merge conflicts but now missing some banalyzer dependencies. Also homedepot eval is still not passing.

[filip-michalsky]

I added "smooth" to the scrolling behavior, the hypothesis is that it reduces hard reloads of the site for slow sites but its just based on the one eval (homedepot)