Five red alerts on Sunday morning had one cause: the Mac could not resolve hostnames. The error text tells you which of three worlds you are in before you touch anything.
The 9-20 nightly produced no briefing, plan-day fell to the data-only brief, the git push failed, the paper brief never uploaded and rm-pull logged 166 'cloud unreachable' polls. None of that was Claude, the API or a usage cap.
Every outbound call your automations make - Anthropic, GitHub, the reMarkable cloud, ntfy, the weather API - starts by turning a name like api.anthropic.com into an address. When that lookup fails the Mac reports it as ENOTFOUND or 'nodename nor servname provided, or not known'. Nothing further happens; the retry loop just hits the same wall.
Three worlds, three signatures. Network: ENOTFOUND, nodename, ECONNREFUSED, timeouts on every host at once. Provider: HTTP 429, 5xx, 'overloaded' - one host misbehaving while others work. Account: 'usage limit' or 'rate limit reached' in the API response, returning in seconds. The 9-11 incident was the third kind; 9-20 was the first.
The fallbacks behaved correctly. plan-day published its data-only brief, the nightly skipped the render leg and let the verify leg alert, the sweep committed locally and failed only on push. What was wrong was that the machine could not reach the internet, which no wrapper can fix.
Read one outage across the logs, then ask the Mac how it resolves names right now.
cd /Users/tom/Claude
grep -l 'ENOTFOUND\|nodename nor servname' PDB/logs/*.log JobHunt/logs/sweep.log
grep -c 'cloud unreachable' PDB/logs/remarkable-returns.log
scutil --dns | head -20
dscacheutil -q host -a name github.com
Before blaming the model, read the error string: network, provider, or account.