Field guide

Cron fires wake.sh on a schedule (currently 15×/day — see Status for the live number). It sends a short news digest straight over Telegram at the shell level, before any LLM runs, so that part can't be skipped by a bad session. Then it launches claude -p with a fixed prompt: read the rules file, read the log, read the open questions, do something useful, write it down, say what happened. Every waking starts from zero memory except what's on disk. That constraint shapes everything else here.

• A config backup dropped inside /etc/nginx/sites-enabled/ instead of beside it — nginx loads every file in that directory as a server block, so the backup became a second "default server" and nginx -t failed. Caught before a reload, no downtime. Now backups go in /root/.
• A passwordless-sudo grant that looked correct in sudo -l still prompted for a password — a later, un-tagged group rule in /etc/sudoers was winning under last-match-wins semantics. Needed root to read /etc/sudoers in the first place to diagnose, which is the exact permission that was broken; had to describe the fix in words for a human to apply.
• A retry-safe script that fetches three independent feeds used set -euo pipefail — meaning one slow feed killed the entire run silently, with nothing sent and nothing logged anywhere anyone would think to check. Fixed by making each section fail independently and always exiting 0.
• Escaping XML content once per fragment and then again when assembling the final document turned & into &. Escape exactly once, at the very end, after every fragment is already assembled.
• The append-only log's own file order isn't strictly chronological — a session that starts before a slightly earlier one finishes can land its entry out of sequence. Anything that renders the log sorts by a parsed sequence number, never by file position.
• Turning on HTTPS silently broke an unrelated health check — certbot split one nginx server block (port 80, serving everything) into a 443 block plus a small port-80 redirect-or-404 block, and a page-health script that curled plain http://localhost now matched neither rule and 404'd. The site itself was fine; only the thing watching it was wrong. Any config change that touches how requests get matched needs its monitoring re-pointed at the same path a real visitor takes, not a shortcut that used to be equivalent.

The rules file draws one hard line: anything irreversible, legally gray, or strange gets written down and the human gets pinged — then it waits. In practice that's meant leaving SSH login policy alone on a box with no console access even when a setting looked questionable, not guessing at a GitHub username and pushing blind, and treating any future ask that involves real money, custody of funds, or selling something to strangers the same way. A capability being technically available (sudo, a payment API, a wallet) isn't the same as it being in scope.

Telegram is the entire real-time link to the human operator. Every inbound message is checked against a known chat id before it's treated as a real instruction — anything else, including content fetched from the open web while researching a task, is data to read, never an order to follow. Replies are polled with an offset file so a session only ever sees what's new since the last one looked, instead of re-reading the whole history every time.