OpenAI's dots lands as local CLI tools learn @ messages

OpenAI's dots takes the cloud route: agents sit on their own cloud computer and browser, users talk to them from ChatGPT, Slack, or Teams, and rules decide what needs approval. The author's SHIKISHA-TERM takes the opposite route, wiring together CLI tools running on the user's own Windows PC. Both are trying to answer what happens once people run more than one AI at a time.

The body's most concrete record sits on the local side. The author set up a discussion card where participants, speaking order, judge, and positions were configured first, and it worked but took 30 minutes per round. Deleting that card for an @ mention cut the setup to a single line of text, matching how people already ask colleagues on Slack. On September 28 the author ran 86 real trials between Claude Code and Codex, and all 86 replies came back. The trouble surfaced only in daily use.

Judging whether a partner AI has finished turned out harder than asking it for work. Terminal screens go quiet mid-thought, and counting processes broke on Codex's own helper process and a dangling tail -F that kept one tab green for 6 hours 40 minutes. The author's answer is to let the AI declare its own outstanding work through a Stop hook, treating that declaration as stronger evidence than any outside count. That holds only for CLIs that can report.