# Release v2.19.0 — the gateway keeps its own traffic to itself

Released 2026-11-02 from `refactor` → `master`, tagged `v2.19.0`, on top of
v2.18.2 (`docs/release-v2.18.2.md`). Most of it was found on the development
gateway (a Raspberry Pi 4 on a home line, Splitting at its Maximum level)
and on the managed servers it runs.

## Changes in behaviour on upgrade — read first

The operator's rule for this release: privacy first, and whatever trades
privacy for a function becomes an option. Each of these used to show the
node's address to a third party; each is now off, with its switch in
**Settings → Privacy — the gateway's own traffic** unless said otherwise.

- **No public fallback resolver.** `/etc/resolv.conf` holds `127.0.0.1`
  alone. The built-in 1.1.1.1 / 8.8.8.8 answered the box's own lookups
  whenever gatewayd's resolver was down (a restart, an update) — straight
  out of the uplink. Settings → DNS → "Fallback resolvers of the gateway
  itself" still takes up to two; empty now means none.
- **The WAN health probe pings the uplink's first hop.** Nothing goes past
  the router. The TCP connect to 1.1.1.1 / 8.8.8.8 every 5 s is the option
  "Check the internet through public resolvers". **A dual-WAN node that
  must fail over on a dead provider behind a live router** turns it on, or
  better names servers of its own as the uplink's probe targets on the WAN
  page (they always win).
- **The gateway's own downloads go through a tunnel**: threat feeds, the
  release site (Check now, downloads) and Tailscale's relay map. "Own
  downloads go through" picks the tunnel; auto is the first hop of the
  highest-priority enabled VPN routing policy that is up. With no tunnel
  up they wait. **A node without a VPN routing policy** (a public server)
  turns on "Download straight over the WAN while that tunnel is down", or
  its threat feeds and the update check wait. A release mirror on the LAN,
  the mesh or loopback needs no tunnel.
- **The release site sees a common browser's User-Agent**, not `gatewayd`.

## The gateway's own traffic

- Threat feeds keep their last copy next to the database (`feeds/`, 0600):
  a start puts it in force at once and downloads again only once it is a
  day old; an hourly loop refreshes day-old copies. Before, every start
  downloaded every feed, and a failed download left that feed's domains
  unblocked until the next start.
- The URLhaus feed moved to malware-filter's urlhaus-filter list (abuse.ch
  ended the free downloads; every old URL answered 410). A node that has
  the old feed gets the successor at start, enabled as it was.
- Ghost mode contacts nothing external: no relay-map download, no probe of
  Tailscale's relays (the relay fallback waits), no captive-portal check,
  no feed load from a mode's actions.
- **A telemetry preset for Tailscale Inc.'s endpoints** (Security →
  Telemetry Block, "Tailscale Inc. (self-hosted tailnet)", off by
  default): tailscaled reached log.tailscale.io and, while the WAN
  stalled, plain-HTTP captive checks to controlplane/login.tailscale.com,
  from a tailnet on its own headscale. It breaks devices on
  Tailscale-hosted tailnets that resolve through the gateway, hence off.
  The installer now runs tailscaled with `--no-logs-no-support`.
- A randomized MAC goes back as soon as a driver reset writes the factory
  address back (USB adapters reset in place; it used to wait for the next
  start).
- A Splitting endpoint's exit is checked only on a click: the Tunnels page
  probed every tunnel's exit within the same seconds, which tied them
  together at the probe service.
- A tunnel proxy's lookups go to the tunnel's own resolver first; the
  parallel AAAA query of each lookup went to the fallback resolver.
- No owner data in shipped code: the map reads country and city keywords
  instead of one deployment's provider table; a reissued certificate keeps
  its name.
- Routing tables carry the uplink's own network, not a fixed
  192.168.0.0/16; DNS source classes come only from the deployment's own
  subnets.

## Splitting

- **"Rotate new connections only"** (under Privacy Level, off by default):
  a rotation reloads only the slot order, open connections stay on their
  tunnel, no full apply. A full apply every minute at the Maximum level
  took a third of a Pi 4's time and stalled open connections on moved
  slots. `PUT /api/v1/splitting/rotation {"new_connections_only": true}`.
- The engine's endpoint tables are rebuilt at each apply: after a boot
  they were missing and the split fell back to the policy's own hops.
- **A MASQUE relay redials its proxy** when its session ends (a WAN
  reset): it used to close its socket silently, WireGuard roamed back to
  the peer's direct address and the tunnel ran unwrapped while the relay
  list showed it wrapped. A tunnel that roamed off its relay is put back
  on it every 10 s; both events are logged.
- MASQUE leaves a chained WireGuard tunnel (one with an outbound) alone.
- A server switch restarts only the transport that changed, the new one
  before the old: the tunnels no longer went direct for that window.
- A split policy's status reads its split band first ("Installed, not
  routing" while it split traffic).

## Mesh

- A coordination-server front (`PUT /api/v1/mesh/control-front`, admin
  only, **off by default**): a node with a public address serves another
  gateway's headscale over HTTPS through the federation tunnel, with a
  short-lived Let's Encrypt IP certificate, only the paths a client uses.
  Devices away from home can log in again without the home address being
  exposed — but it opens a public entry to that headscale; leave it off
  unless that is wanted.

## Servers and releases

- A release downloaded from the site also takes the server files (the
  agent and the role binaries for the servers' architectures), each
  checked against the signed SHA256SUMS; a cached release is topped up.
- Check: **never** — no contact with the release site at all (manual
  policy only).
- An update, rollback or reconfigure runs to its end: the artifact review
  is noted and asked for afterwards; a serving role is bound before the
  review (a bridge stayed pinned to its old certificate for 11 minutes).
- Uploads and site downloads ride over a link that drops for seconds; the
  SSH keepalive no longer closes a slow upload; the journal cap fits /run
  on small servers.
- The release build's `requires_installer` compares the previous release
  byte for byte (every release since v2.18.0 said so wrongly).

## Fixes

- One lost ping no longer judges a tunnel dead (three requests, any reply
  passes); a bridge reads down after three failed ticks, not one.
- bridge-no-traffic counts what came between ticks, warns only while
  clients asked for the bridge's domains, and leaves the gateway's own
  lookups out; node-down and mesh-link-down wait two minutes.
- Ephemeral ports stay above the gateway's listeners (32768-60999): a
  forwarder could not bind its port after a restart.
- A backup that fails for room frees the oldest snapshot and runs again.
- The proxy forwarders refuse the clients the forward path drops
  (device-acl, untrusted trust classes).
- A policy's dead hop and a conditional forwarder's failures log once,
  not on every pass.
- The deployment mode applies with the save (it said "restart required"
  and changed the firewall at the next start).
- `log.max_size_mb` caps `/var/log/gatewayd.log` and its `.old` together;
  the packages' own log lines (MASQUE, obfs4, Splitting's recoveries)
  reach the file — they went to stderr.
- The pre-start salvage of a broken database keeps the managed users and
  API keys (hashes only).
- The installer's optional downloads (Ollama, Gitea, Go) are refused
  unless their sha256 is the pinned one.

## Upgrading

- From v2.18.2: one helper script changed (`gateway-pre-start.sh`'s
  salvage), so BUILDINFO says `requires_installer` (rightly this time) and
  the update asks for the installer's refresh first: System → Updates →
  "Refresh them from v2.19.0" (the cached release's own
  `install-gateway.sh --refresh`, checked by hash), then prepare.
- Schema 77: migration 077 adds the Tailscale preset's rows, so
  `downgrade_safe_from` is v2.19.0 — going back to v2.18.2 is the update's
  own rollback inside its window (it restores the database snapshot), not
  a plain swap of the binary.
- The privacy defaults above apply at the first start. A node that wants
  any of them back turns it on in Settings → Privacy (or Settings → DNS
  for the fallback resolvers). The v2.18.2 node fetches this release as
  before; once on v2.19.0, a node with no VPN routing policy fetches the
  next ones only with "Download straight over the WAN while that tunnel is
  down" on — or from an uploaded archive.
- From v2.18.1 or older: first as for v2.18.2 (`docs/release-v2.18.2.md`).
