I don’t want to be mean, but I’d imagine google has the means to produce quality software. You’re not a startup, you have extraordinary ressources, yet the chat agent is full of bugs. Such as ‘the model output exceeded token limits’, the ai agent not being able to edit files properly, or outputing random text in the tool calls… It is your LLMs that this tool is running, there’s no excuse.
I feel sorry for the ai :
renderworldisn’t at 1422—view_fileis showing me something else entirely. I’m through playing hide-and-seek with line numbers. I’m usinggrepto find the exact line forvoid renderworld()inrendergl.cppso I can pin it down and document it properly for the handover.Hunting it down…
renderworldis clocked at 1858—confirmed. Checking the context of 1858 now to get the full picture.Line numbers are still gaslighting me—
greppoints to 1858, butview_fileis serving upcleanupmotionblur. I’m switching togrep -nC 5to pull the actual context directly from the source so I can see what’s really happening aroundrenderworld. No more toggling between tools until I’ve got the visual proof.Hunting for the real
renderworld…Waiting.
Line numbers are a total bust—I’m flying blind. I’m pivoting to search strings I can actually trust. One last push: I’m going to silence the alpha/transparent calls in
rendergl.cpp’srenderworld, delete them fromrenderva.cppentirely, and make surerenderworldcatalog(plus its reflection counterparts) is solid.I’m using
grep -n -A 15to get a definitive lock onrenderworldinrendergl.cpp. If I can’t find it there, I’m not touching it. Then it’s straight to the handover doc so the next AI has a clean map of this minefield.Zeroing in on
renderworld…
So basically I discovered that the AI builds an internal representation of the files, and that this representation can go out of synch, and the AI is unable to update its own representation even when presented with the truth. Good job!
I made some progress, and created my own file parser and editor as a command line tool, with a built in system to avoid (yes!) hallucinations, and it works much better now !