Ops automation

Hourly feedback triage

scripts/feedback-triage.mjs runs hourly from the host crontab (installed

2026-07-07, guarded by flock):

0 * * * * flock -n /tmp/paidup-feedback-triage.lock /usr/bin/node /srv/shared/Projects/paidup/scripts/feedback-triage.mjs >> /srv/paidup/logs/feedback-triage.log 2>&1

What it does per run:

• Reads up to TRIAGE_MAX_ROWS (default 3) status='new' rows from the prod

feedback table through docker exec paidup-postgres psql (the DB has no

published port; DATABASE_URL is read from /srv/paidup/.env).

• Classifies bug-likeness (scripts/lib/feedback-triage-helpers.mjs,

unit-tested in scripts/feedback-triage.test.mjs): kind=bug always,

otherwise defect-language heuristics.

• Bug-like rows (max TRIAGE_MAX_CODEX_ROWS=1 codex run per hour) get an

isolated git worktree under /srv/paidup/triage-worktrees on branch

fix/feedback-<id> and a codex exec --full-auto investigation with a

20-minute budget. Commits present → branch kept for human review; no

commits / codex failure → worktree and branch are dropped.

• Every processed row becomes status='triaged' with the verdict merged

into meta.triage (bug_like, outcome, branch, detail,

triaged_at).

• Telegram notification per row via TELEGRAM_BOT_TOKEN/TELEGRAM_CHAT_ID

(from env or /srv/paidup/.env); logged and skipped when unset.

Review fix branches with git branch --list 'fix/feedback-*'; merge or drop.