Review fixes for the persistent logs:
- each service writes its own folder (the old-file cleanup matches by
prefix, so 'worker' deleted the 'worker-ai' files);
- at most LOG_MAX_DAY_MB (50) a day, and at least 2 days kept;
- written by a background thread, not on the async workers;
- e-mail addresses and tokens in request paths are masked;
- a failing health check is logged again, a passing one is not;
- one-off commands (compare-models) log to the console only.
On a stop the worker takes nothing new and lets running work finish
(SHUTDOWN_DRAIN_SECS, worker 40 s, worker-ai 270 s, with matching stop
grace periods), so a deploy no longer cuts an analysis off, wastes its
model call and delays it by the 15-minute stuck-job requeue.
The CI cleanup caps the build cache at 10 GB: by age it had grown to
33 GB on the runner and filled its disk.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL
- SIGNUP_MODE=invite (default): a new account needs an open invitation for
its address, or the invitation link (then any Google account will do)
- admin page "Meghívók": invite many addresses at once, personal message,
resend (max 3, 24 h apart), revoke, copy link, mail preview and test copy,
waitlist
- invitation e-mail in HTML and plain text; Message-ID on the sender's
domain, List-Unsubscribe with one-click POST, opt-out list (hashes only)
- e-mail texts moved to a shared crate (crates/email-templates)
- login page shows the invitation; waitlist for uninvited visitors;
unsubscribe page; privacy policy updated
- docs/levelkuldes.md: Brevo, DNS (SPF, DKIM, DMARC) and the SMTP setup
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL
The system prompts of the free summary and the paid detailed analysis can be
edited per language under /admin/promptok. Every save is a new row in
prompt_revisions (NULL = built-in text); new analyses record the revision in
prompt_version (v2-r<id>) and the worker uses exactly that text. Existing
analyses stay; the cross-user cache only copies results of the same revision.
The prompt code moves to the shared analysis-prompt crate so the backend can
show the built-in text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL
main builds images tagged with the commit hash and `main`; prod also tags
`prod` and `latest` and deploys. A prod run without DEPLOY_HOST now fails
instead of skipping silently.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL
- .gitea/workflows/ci.yml: test (Postgres service), build and push the
four images tagged with the commit hash, deploy over SSH when
DEPLOY_HOST is set
- ci/runner: act_runner 0.6.1 on the dev server using the host's Docker
- deploy/: production compose with Caddy (https), daily pg_dump backups,
log rotation; setup guide
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012vLHzh5qUhFtmJhzkcTbXL