Tom picks his tools before he talks
Tag Tom and he now runs a separate decision step before answering, choosing which of his five lookups the question needs. Platform questions and repository history get the right search more often.
Tag Tom under a post that asks who built bookmarks and he comes back with the merged pull request and its author. Ask how many members joined this week and you get a count, not a shrug. The reply arrives much as it always did. What changed is the quiet beat before it.
Credit goes to paulinevos, who split Tom's thinking in two. Before a word of the answer, a classification call through OpenAI asks five yes-or-no questions, one per tool, using your latest message and the same branch of thread Tom remembers. That branch holds at most 12 posts and 12,000 characters. Whichever questions pass get handed to the model that writes Tom's answer, and Tom has to run every picked tool before he writes, inside seven model steps with a 15-second decision timeout. A knowledge lookup returns three sections. A post search returns five commits.
One catch. The decision is made blind. It reads your question and the thread, never the search results. A picked tool still has to run, so an overeager yes buys a lookup Tom never uses before he can answer at all. He may search again, twice per tool, but the first swing is already spent.