Paste a stack trace into a chat and you will get a cause, a fix, and a confident order of events. Sometimes the cause is right. Often it is the most familiar bug that fits the words in the trace. The danger is editing the code to match the story and then discovering the original failure is still there, plus a new one.
Start from a reproduction you can run. One command, one input, one observable result that is wrong. If you do not have that, finding it is the task, and a proposed patch is premature. Ask the model where to look, if you want, but do not apply a patch against a bug you cannot see.
When you have the reproduction, ask for a hypothesis that predicts something new. “If this is a stale cache, the failure disappears after this key is removed.” Run that check. A hypothesis that does not predict a check is an essay. You can read it later.
Apply one change, then run the same reproduction. If it still fails, revert. Do not stack a second guess on top of an unproven first guess. The log of what you tried is more useful than a longer explanation, and it belongs in the commit message or the issue, not only in the chat.