tooling

A log check flagged 16 of 31 Claude Code turns that ended on a failing test. None ended on a false claim; it missed the one that did. Why, and 3 rules to try.

Four tests I ran on Graph Starz, a map of pictures: replaying uploads, a bulk import, scoring 16 picture models against museum records, and blind judging.

I'm testing whether my agent operator's rules and tools work better as a shared kit in project sessions: why it's worth trying and what the trial must show.

Cloud Run ends every request at its timeout, streams included, and EventSource didn't reconnect. My fix: a named ping event and a 70-second client watchdog.

I applied Jon McNeill's five-step Algorithm to six recurring Command jobs. The strongest opened 31 remediation pull requests in 30 days, and 29 merged.

I had the 38 unattended jobs in my operator's registry declare a setpoint, sensor, error signal and actuator. Sixteen aren't loops; only two close the loop.

I built Karpathy's LLM wiki pattern into Command: 134 provenance-tagged pages, a gate that blocks fabricated quotes, and a human-gated publish step.

Claude Code deletes session transcripts older than 30 days by default. Raise cleanupPeriodDays in ~/.claude/settings.json to stop it; lost logs stay lost.