database agent improvements

[wraith stump]

a few things to try

• add the question is: $question to the end of the prompt
• add a module description to your sql module
• this won't affect reliability, but it's kind of strange to call LastReply and discard the actual reply instead of returning it. Was there a specific reason for that?

[mortal snow]

also worth trying GPT 4.1, it's supposed to be better at following instructions

[primal star]

I was using last reply and returning it, but that resulted in double outputs in the terminal. For some strange reason, I though I needed to assign the call to something in order to execute. That might be something from a really old version? I remember trying it on the build pipeline I was working on and it wouldn't actually do anything (maybe bad memory/testing)

I'll add the question back into the prompt, I thought we wouldn't need that anymore based on the env but will definitely try it.

Tried 4.1 with the same results, its odd that sometimes it works - like all day with Gemini - and then suddenly stopped working.

Thank you for the feedback, will try those now

[wraith stump]

You're probably seeing the trace output from the agent running and then the output itself. one will have a 🤖 next to it

[primal star]

Follow up:

1. Removed the last reply and its does not perform any work: https://dagger.cloud/mccallister-dev/traces/a96f5466b458f0f0337641a306c97d66
2. Added the last reply back and it worked, but now I'm still not assigning a variable which feels odd that I'm not handling the error - which probably won't occur but still
3. Added "the question is" and it seems to be working

[primal star]

yeah - it was confusing to look at on the screen

[wraith stump]

sure, there should be a visual difference between the two. You still want to return it though I think. If you ever wanted to call that function from another function, it has to return it's response