Restructure the deployment model: instead of bundling and owning its own
Traefik + wg-easy, the control-plane now plugs into a Traefik and wg-easy
the operator already runs (as sibling containers, or otherwise reachable).
This avoids duplicating infra self-hosters commonly already have, and
sidesteps having to run a second ACME/DNS setup entirely — the operator's
own Traefik already has a working cert resolver.
- New control_plane/wgeasy_v15.py: a wg-easy v15 REST client, verified live
against a real running container (auth model, route shapes, and the new
per-client PresharedKey all differ materially from v14's API). Threaded
the preshared key through to the box side (furtka/wireguard.py in the
furtka repo) so v15's handshake actually completes.
- GATEWAY_CERT_RESOLVER is now a real env var instead of a GATEWAY_MODE-
derived hardcoded "le", so it can mirror whatever resolver name (if any)
the operator's own Traefik already has configured.
- Root docker-compose.yaml/ops/deploy.sh: control-plane only now. deploy.sh
asks how the operator's Traefik reaches it (same docker network, joined
via `docker network connect`, or a separate host) instead of prompting
for DNS-01 providers and port remaps single-parallel mode needed.
- single-parallel mode is retired outright — nothing needs it once
furtka-gateway doesn't own a second Traefik/wg-easy to remap ports for.
- The old bundled stack (own Traefik+wg-easy, HTTP-01) moves to
examples/starter-stack/ as the documented zero-infra on-ramp.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R3jVwAqESRjPdW9P9nvfem
Swarm ignores docker-compose.yaml's build: key, so docker stack deploy
previously needed an operator to build and tag the control-plane image
by hand. CI now publishes it to this instance's container registry on
every push to master (or a tag), and docker-compose.yaml's image: key
(overridable via GATEWAY_CONTROL_PLANE_IMAGE) lets Swarm pull it directly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0142mP6EqCkQKmR1xCp7zfmZ
Two gaps found while walking through a real single-tenant deployment
scenario, both fixed and verified against the actual running stack (not
just unit tests):
- wg-easy moves to network_mode: host so Traefik can actually route to
WireGuard peer addresses (wg-quick's route only existed inside wg-easy's
own isolated network namespace before). This forced host-level sysctls
(ops/host-sysctls.sh — Docker/runc rejects namespaced sysctls under host
networking) and a host firewall rule (ops/firewall.sh) to keep wg-easy's
now-unisolated admin API off the public interface.
- control-plane gets a real Traefik route (traefik/dynamic/control-plane.yml.example)
instead of being reachable only at 127.0.0.1:8090 — box registration
carries a WireGuard private key and needs TLS, not bare loopback HTTP.
Along the way, running the actual pinned wg-easy image
(ghcr.io/wg-easy/wg-easy:14) surfaced that it's a different, older,
Express-based codebase than what wg-easy's current docs/master branch
describe: auth is a plain Authorization header per request (no session
cookie), routes live under /api/wireguard/client, and its default
WG_ALLOWED_IPS is full-tunnel (0.0.0.0/0) rather than the split-tunnel this
design assumed. wgeasy.py and docker-compose.yaml are corrected
accordingly, verified end-to-end against the real container.
Also fixed: Docker Compose silently truncates a bcrypt hash's `$`
characters when read from .env — documented the required $$ escaping.
wg-easy + Traefik docker-compose stack (Phase 1) plus a stdlib-only
control-plane API for box registration, WireGuard peer provisioning via
wg-easy, and per-box route publish/unpublish backed by Traefik's file
provider (Phase 2, single-tenant mode). SQLite holds accounts/boxes/routes
so a later multi-tenant shared instance is the same schema with more rows,
not a reshape.
wg-easy's actual REST API was verified against its source rather than
assumed: it has no bearer-token auth (session-cookie login via
POST /api/auth/password) and no way to accept an externally-generated
public key (it always mints the keypair itself, private key included) —
both corrected from the original plan during implementation.
