Pushing straight to GitHub from KZ is a lottery — TCP to github.com
times out often enough that git push becomes a flake. Fix: Forgejo runs
on the stage server (sqlite, single container), all pushes go there
first (local network, always reliable), a systemd timer mirrors the
whole repo into GitHub every 10 minutes so GitHub stays up-to-date as
a backup + CI source.
What's committed here is the infra-as-code side:
- deploy/forgejo/docker-compose.yml — Forgejo 7 on :3000 (HTTP) and :2222 (SSH)
- deploy/forgejo/food-market-forgejo.service — systemd unit that drives compose
- deploy/forgejo/mirror-to-github.sh + mirror timer/service — push to GH every 10 min
- deploy/forgejo/nginx.conf — vhost for git.zat.kz (certbot to be run once DNS is set)
- docs/forgejo.md — how to clone/push, operations, what's left for the user (DNS + certbot)
GitHub Actions CI is untouched: commits land on GitHub via the mirror
and the self-hosted runner picks them up as before.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Telegram bridge lets me drive the local Claude Code tmux session from my
phone — inbound messages are typed into the 'claude' session, pane diffs
are streamed back as plain Telegram messages (TUI noise, tool-call
blocks, echoed user input and already-sent lines are filtered so only
the assistant's actual reply reaches the chat). Deployed as
food-market-telegram-bridge.service, reads creds from
/etc/food-market/telegram.env (not committed).
Also committing the local docker-registry unit for reproducibility —
registry:2 on 127.0.0.1:5001, data persisted in
/opt/food-market-data/docker-registry.
Setup docs in docs/telegram-bridge.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stage's external pulls from ghcr.io flap on KZ network — the self-hosted
runner pushes images into a local registry:2 (systemd-managed,
/opt/food-market-data/docker-registry) and docker-compose now pulls from
localhost:5001 via \$REGISTRY. ghcr.io is still tagged and pushed as
off-site backup, but ghcr push failure no longer fails the build.
Setup done on the host (not in workflow):
- systemd unit food-market-registry.service (enabled, restart on failure)
- /etc/docker/daemon.json: \"insecure-registries\": [\"127.0.0.1:5001\"]
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
.github/workflows/ci.yml — on push/PR:
- backend job: dotnet restore/build/test with a live postgres service
- web job: pnpm install + vite build + tsc, uploads dist artifact
- pos job: windows-latest, dotnet publish self-contained win-x64
single-file exe as artifact
.github/workflows/docker.yml — on push to main (if src changed) or manual:
- api image → ghcr.io/nurdotnet/food-market-api:{latest,sha}
- web image → ghcr.io/nurdotnet/food-market-web:{latest,sha}
- uses buildx + GHA cache
deploy/Dockerfile.api — multi-stage (.NET 8 sdk → aspnet runtime),
healthcheck on /health, App_Data + logs volumes mounted.
deploy/Dockerfile.web — node20 build → nginx 1.27 runtime; ships the
Vite dist + nginx.conf that proxies /api, /connect, /health to api
service and serves the SPA with fallback to index.html.
deploy/nginx.conf — SPA + API reverse-proxy configuration.
deploy/docker-compose.yml — production-shape stack: postgres 16 +
api (from ghcr image) + web (from ghcr image), named volumes, env-
driven tags so stage/prod can pin specific SHAs.
deploy/backup.sh — pg_dump wrapper with 3 modes: local (brew
postgres), --docker (compose container), --remote HOST:PORT. Writes
gzipped dumps to ~/food-market-backups, 30-day retention.
docs/24x7.md — explains where Claude/CI/stage live, which pieces
depend on the Mac, and the exact steps to hand off secrets via
~/.food-market-secrets/ so I can push them into GitHub Secrets.
Next, once user supplies Proxmox + FTP + Telegram creds: stage deploy
workflow, notification workflow, and (optional) claude-runner VM so
I no longer depend on the Mac being awake.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
