Why do coding agents keep reopening files they already should understand?
Been testing coding agents on larger repos and saw something odd.
Even after they already explored the codebase, they still kept doing stuff like:
reading the same areas repeatedly
opening files that weren't really relevant
missing relationships between components
making changes that technically worked but didn't fit existing patterns
The weird part is it didn't always feel like a context size issue.
It felt more like they had no lasting understanding of the repo itself.
I started messing around with this idea in RepoWise
Mostly experimenting with repository signals like dependency relationships, commit history, files that change together, etc.
Curious if people building agents are seeing the same thing, or if there's already a better way to handle this.
GitHub link in comments