| name | openfdd-stack-ghcr |
| description | Use when pulling, rebuilding, or verifying Open-FDD GHCR stack images. Test channel is nightly. Triggers on: nightly, openfdd-central, openfdd-web, GHCR, openfdd_stack_up, OPENFDD_IMAGE_TAG. |
Stack GHCR (nightly channel → immutable verify)
Full protocol: CONTAINER_AGENT.md.
nightly is the channel selector. Qualification pulls sha-<commit> for
central/web/fieldbus/mqtt, asserts digests match :nightly, then starts the
stack with OPENFDD_IMAGE_TAG=sha-<commit>.
Newest means OCI created, not the word nightly. Run
./scripts/ghcr_newest_by_created.py openfdd-central before claiming tip.
Unmerged frontend/web → local web only. openfdd_stack_up.sh refuses
GHCR web when the tree drifted (override OPENFDD_ALLOW_STALE_GHCR_WEB=1).
Never paste a Caddy login until ./scripts/openfdd_demo_gate.sh exits 0.
Product central image is Rust/debian only (no Python).
Low-RAM hosts: never local docker build for stack images. Prune unused
images before pull; wait for GH Actions publish; then pull +
openfdd_stack_up.sh … --no-pull. See CONTAINER_AGENT.md.
Nightly publish (ghcr-openfdd-stack.yml) stamps OPENFDD_GIT_SHA into central
and writes web version.json { version, git, image_tag, service: openfdd-web }.
Optional extra image tag :3.3.2-n<run_number> sits beside :nightly /
:sha-*. UI revision comes from central /api/health first.
For Railway, use the same GHCR openfdd-web image and set
OPENFDD_CENTRAL_UPSTREAM=openfdd-central.railway.internal:8080 (assuming the
central service uses that name). Tip images resolve upstream DNS lazily
(OPENFDD_NGINX_RESOLVER=auto) so nginx does not die when .railway.internal
is not ready at process start. Deploy central healthy first, then mqtt (for
cloud MQTTS hubs), then web. Live OT hubs should include openfdd-mqtt by
default — MQTTS is the cloud transport; fieldbus stays on-prem.
CLI re-pin (bensbench): openfdd-railway-cli
— railway login verified; project gleaming-cooperation / production; use live
names (openfdd-central-cQ-F, …) with
railway service source connect --image ghcr.io/bbartling/<img>:sha-<7> in
central→mqtt→web order. Optional token file: ~/.config/railway/bensbench.env.
Railway CLI ≠ openfdd-mcp FDD tools.
Set OPENFDD_AGENT_PASSWORD on central for FDD AI / MCP (username agent →
operator JWT); do not put OPENFDD_ADMIN_PASSWORD into Cursor MCP config.
Admins may POST /api/auth/agent-token for short-lived operator JWTs.
Edge kits: operator/admin POST /api/mqtt/edge-kits or Operations MQTT UI
(ZIP never includes CA private key).
Local Compose keeps the default central:8080. Do not claim a Railway-ready
nightly until stack + MCP GHCR publish jobs are green and the target sha-*
digest resolves.
Hourly CSV append is API POST /api/csv/import/package/append on central (GHA image), not a local compile.
After GHCR publish, poll with ./scripts/ghcr_watch_central.py (legacy shim: wattlab_parity_watch_ghcr.py), then maint pull + synthetic_59_* chain — not vibe19 dump-parity.
BUILDING_50 stream sim: scripts/csv_flood_afdd_routine_sim.py — see docs/agent/CSV_FLOOD_AFDD_ROUTINE.md.
Ops patch cycle: gate FAILs → BUG_REPORT_OT_MODBUS_HAYSTACK.md Phase 7 table before fix; re-pin after publish per openfdd-railway-cli.
Workflow: ghcr-openfdd-stack.yml (retargets nightly on master).
MCP: separate rust-ghcr-mcp.yml.
