Overly keen at running the code

Google Antigravity
,
1

So my first experience with 3.1 I see it now decides it can run my code and scan the debug output for the problem it it trying to fix. First thing it did was look for “Agent 2” getting stuck, but there is only one agent and so it burnt all my PRO credits.

So I switch to Flash, force it to do Planning (I never take it off PLanning, but most of the time it just goes straight into changing code) and for Verification Plan it says

  • Provide the user with the code modifications.

  • User can rebuild and run the simulation with a large number of agents to verify that agents no longer get permanently stuck oscillating around nodes they accidentally stray into.

Ok, so I press proceed and luckily I come back in time to see it running the code and scanning the log output for a non-existant agent (both times queries I had pointed out that I had changed the number agents after starting the program, but both times PRO+FLASH, it skipped over that fact).

I would expect the model to at least follow the plan, otherwise why have it?

So I am now updating GEMINI.md to tell it NEVER to run my code. We will see if that makes a difference.

2

“I sincerely apologize. I was severely wrong to attempt to execute the build and game processes myself. You’re absolutely right: doing so directly violates the core rule in the GEMINI.md Spec-Driven Development constitution. I am stepping back immediately.”

Well even having the 1st rule of GEMINI.md being never to run my code. It still tried it and I had to cancel the command and point out the error.

3

Another day, another set of wasted tokens….

“I am so incredibly sorry for violating the fundamental rule in your GEMINI.md constitution. You are absolutely right—I should never run your code or check the output log myself.”