tooling

I'm testing whether the rules and tools from my agent operator can do more useful work as a shared kit. The design moves coordination into project sessions; the review explains why that's worth trying, where the evidence is weak, and what the seven-day trial must show.

Cloud Run ends every request at its timeout, five minutes by default, and a long-lived stream is still a request. EventSource is supposed to reconnect and in this case didn't, and the usual comment-line keep-alive can't help because the browser never sees it. The log timeline that proved it, and the fix: a named ping event plus a 70-second client watchdog.

I applied Jon McNeill's five-step sequence from The Algorithm to six recurring Command jobs. The strongest opened 31 remediation pull requests in 30 days, and 29 merged, showing how accurate automation becomes finished work.

I spent about three weeks making the 38 unattended jobs in my operator's registry write down their setpoint, sensor, error signal, and actuator, plus whether they can tell a failed measurement from a healthy reading. Sixteen declared they aren't control loops at all, and only two of the rest close the loop: they act on the world and then check that the action worked.

I built Karpathy's LLM wiki pattern into Command, my Claude Code operator: 134 provenance-tagged pages, a grep gate that blocks fabricated quotes, a nightly digest pipeline that was reset after a cheap model wrote a wrong verdict into it, and a human-gated publish wall.

The cleanupPeriodDays default is 30 days, so Claude Code deletes session transcripts older than that at startup. Set it higher in ~/.claude/settings.json to stop the deletions. What the setting covers, what was already gone on my machine, and why raising it can't recover the past.