v0.5.0

Added

• Topology clients labelled by hostname/IP, not MAC (#35). The topology view
  showed each associated client as a raw MAC, which is poor UX. GET /topology
  now enriches clients with their DHCP-assigned ip and hostname, resolved on
  the controller from dnsmasq's lease file (/tmp/dhcp.leases) — a member node
  is a bridged dumb AP and holds no leases, so resolution happens where the merged
  graph is served. The view labels a client by hostname, falling back to IP, then
  the MAC when neither resolved; the MAC and IP are kept on the node's data. A
  client with no lease (static, self-addressed, or transient) renders by MAC as
  before.
• Any ethernet jack just works on a non-controller node (#42). A satellite's
  wan jack was a dead routed wan: only its lan jack(s) carried clients or extended
  the network, so plugging a computer — or another node — into the wan jack did
  nothing. On a non-controller node the network posture now folds every
  ethernet jack (incl. the wan jack, network.wan.device) into br-lan as a plain
  bridge relay port and stands the routed wan down, so any device works on any
  port with zero configuration. A controller keeps its wan jack as the routed
  internet uplink.
• Dual-band / multi-band client AP (#36). A claimed home previously
  broadcast its client AP on a single radio (typically 5 GHz), so 2.4 GHz-only
  devices had no AP to join. ApplyProfile now authors the client AP across
  multiple bands: the primary omm_ap stays on the radio/band-resolved
  radio, and each band in the new ap_bands profile field ("2g"/"5g"/"6g")
  resolves to that node's matching radio, authored as omm_ap_<band> (e.g.
  omm_ap_2g, coexisting with omm_mesh on the same 2.4 GHz radio/channel).
  Empty ap_bands defaults to also broadcasting on 2.4 GHz, so every home gets a
  dual-band AP with no configuration; set a single band (e.g. ["5g"]) to opt
  out. Bands absent on a node are skipped (not fatal), and narrowing ap_bands
  prunes the now-stale omm_ap_<band> sections on re-apply.
• Xiaomi AX3600 as a build/deploy target. build-devices.sh gains an
  ax3600 label and deploy.sh recognises it. The board is a Qualcomm IPQ8071A
  (qualcommax/ipq807x, Cortex-A53) — the same aarch64_cortex-a53 ISA group as
  the ZB8103AX, so the release feed's arm64 package already covered it; this just
  makes the local dev tooling first-class for it. Because two boards now share
  one ISA, deploy.sh disambiguates by board_name (xiaomi,ax3600) before
  falling back to uname -m. No profile changes were needed: the mesh radio is
  auto-selected by band, which picks the AX3600's 2.4 GHz radio (radio2).

Changed

• Backhaul model: wired is primary, the mesh is a carrier-loss backup. The
  short-lived "batman-always-on, per-port enslavement" model (which routed a wired
  node over the wireless mesh and broke ethernet clients in a mixed
  wired+wireless bridge) is replaced by the failover model: every ethernet port is
  a plain br-lan relay, and the 802.11s mesh + batman-adv is authored as a
  standby that the daemon's carrier-toggle failover brings up only when the
  wired uplink loses carrier (and tears back down when the wire returns) — so the
  fast wire is always preferred and wired + mesh never bridge-loop. The uplink port
  to watch is auto-detected (the bridge port through which the node reaches its
  gateway, so any jack works) or set via uplink_port. Wired ports are enslaved to
  batman only via an explicit batman_ports; batman still forwards the mesh
  loop-free multi-hop when it activates. Restores internal/backhaul (the switch
  loop) and removes the per-port beacon classifier (PortScan / SniffOMMBeacon /
  Classify).
• Network posture management now defaults ON (manage_network, opt-out).
  Making a wan jack usable as a client/backhaul port physically requires standing
  down its routed role first — a netdev cannot be both network.wan's device and
  a br-lan member, and a satellite that keeps its own DHCP/gateway becomes a
  rogue DHCP server on the home segment. That bridged dumb-AP posture used to sit
  behind manage_network=0, so #42 could not be delivered without it. The flag
  now defaults to 1: a non-controller stands down its routed wan and every
  ethernet jack joins the home L2, while a controller keeps the stock routed-wan
  gateway (never disrupted). Set manage_network=0 to opt out entirely on a
  device you hand-wire and do not want meshd to reconfigure.
• Mesh-node network posture now bridges into the home (single gateway). A
  claimed satellite (manage_network=1) previously only stood down its
  authoritative DHCP, leaving its own routed/NAT'd wan up — so its bridged
  802.11s mesh was an island and mesh traffic could not reach the home WAN, which
  egresses only through the controller's gateway over the mesh. The Mesh-node
  posture now authors the same bridged shape as Guest: it folds every ethernet
  jack (incl. the wan jack) into br-lan as a plain relay, lan becomes a DHCP
  client, the routed wan/wan6 are disabled, and authoritative DHCP is stood
  down, so the node is a pure L2 bridge into the home and its default route points
  at the controller. The 802.11s mesh stays a carrier-loss backhaul standby (see
  the backhaul-model change above), so wired + mesh never bridge-loop. Verified
  end-to-end on hardware: reset → wired auto-onboard → mesh-node, pulling the real
  home profile.

v0.4.0

What's Changed

• Per-device cross-build wrapper + 802.11s mesh / wired multi-AP docs by @and-elf in #15
• 802.11s auto-fallback to multi-AP + per-node backhaul in LuCI by @and-elf in #16
• Lifecycle-managed network posture (Guest dumb-AP so discovery works) by @and-elf in #17
• Discovery/enrollment fixes + zero-config wired onboarding by @and-elf in #18
• docs(changelog): record unreleased onboarding/network/discovery changes by @and-elf in #19
• Zero-touch defaults + mesh-capable wpad provisioning by @and-elf in #20
• fix(config): default identity dir to an absolute path by @and-elf in #22
• feat(meshd): auto-propagate controller profile edits to joined nodes by @and-elf in #21
• feat(profiles): host the 802.11s mesh on a dedicated backhaul radio by @and-elf in #23
• fix: a configured node actually joins its controller on boot; failover reconciles mesh state by @and-elf in #25
• feat(backhaul): prioritize ethernet, fail over to the wireless mesh by @and-elf in #24
• feat(backhaul): batman-adv multi-hop routing (#26) by @and-elf in #30
• fix(topology): connect real nodes with typed links in the topology view by @and-elf in #31
• feat(topology): surface onboarded-but-silent nodes in the graph (#29) by @and-elf in #32
• Fix topology view (wired links, clients, LuCI) and add node removal by @and-elf in #33
• feat(web): adopt the LuCI host theme when embedded in an iframe by @and-elf in #34

Full Changelog: v0.3.0...v0.4.0

v0.3.0

Added

• Backhaul connection type in the topology graph. Each node now reports how
  it reaches the mesh — over a wired ethernet uplink or the wireless mesh —
  derived from a configured uplink interface under /sys/class/net
  (carrier/operstate). The value rides on the node's self vertex, is
  preserved by the controller-side aggregator when that node is merged into the
  mesh-wide graph, and surfaces as a backhaul field on a topology node; the
  Topology view marks wired nodes with a solid border and wireless with a dashed
  one. Set the uplink interface with MESHD_BACKHAUL_IFACE (UCI backhaul_iface);
  empty => unknown. New internal/topology/backhaul.go.
• Status LED reflecting onboarding state. meshd now drives a node's status
  LED from its onboarding state so an installer can read it off the device
  without a companion app: blinking while unclaimed, a heartbeat while it joins,
  solid once a home is active. The LED is the kernel sysfs LED named by
  MESHD_LED_NAME (UCI led_name, default green:status); a board lacking that
  LED is a graceful no-op, so the same build runs unchanged across hardware. Set
  MESHD_LED=0 (UCI led_enabled '0') to leave the LED alone. New
  internal/deviceled package.
• Wired auto-onboard. An unclaimed node that is on the wire (ethernet
  backhaul) can now enroll into a discovered controller unattended, with no setup
  wizard: when it is still unclaimed, its ethernet uplink is up, and a controller
  other than its own Home has been discovered, it joins (the lowest discovered
  home_id, chosen deterministically), applies the returned profile, marks setup
  complete and tears down the first-boot setup AP. Opt-in via
  MESHD_AUTO_ONBOARD_WIRED (UCI auto_onboard_wired, default off); requires
  MESHD_BACKHAUL_IFACE to be set, runs only when no explicit MESHD_JOIN
  controllers are configured, and completes unattended only when the controller
  auto-adopts. New internal/onboard package.

v0.2.0

Added

• First-boot setup access point. While a device is unclaimed (setup not
  complete) meshd now brings up a known, label-printable WiFi AP
  (OMM-Setup-<last4-of-node-id>) on a small static network
  (192.168.254.1/24) serving its open management API, so a companion app can
  reach an out-of-the-box node before it has joined any network. The AP is torn
  down automatically once onboarding completes. Open by default; set
  MESHD_SETUP_AP_KEY for WPA2, MESHD_SETUP_AP_RADIO to choose the radio, or
  MESHD_SETUP_AP=0 to disable (e.g. a radio-less wired controller). New
  internal/setupap package; uci.Client gained SetSection/Delete. Covered
  by a real-OpenWrt-container e2e (TestSetupAPLifecycleE2E) asserting the uci
  sections appear on boot and are removed once onboarding completes.
• Companion onboarding app. The Vue frontend now also builds as a
  cross-platform companion app (Capacitor: Android/iOS/desktop), whose first goal
  is to make adding a node a few-tap flow. It discovers controllers on the LAN
  (native mDNS _mesh._tcp, falling back to the daemon's /scan), can target a
  specific device over the LAN (an unclaimed node's setup-AP address or a
  controller's announced URL), and guides adding a node end to end (/onboard):
  read the device's setup label (QR — OMM-JSON, the WIFI: standard, or a bare
  SSID), join its setup AP, reach it, request enrollment (/enroll/join), and
  confirm adoption — including signing in to a split-mode controller (LuCI
  session.login over /ubus) when its management API is localhost-bound. The
  native capabilities (mDNS / WiFi-join / QR scan) sit behind a bridge that is a
  no-op in the browser, so the existing PWA is unchanged. See the design spec in
  doc/companion-app.md.
• Guided onboarding wizard + wireless-only enrollment. The onboarding flow is
  now a three-page wizard — choose Home → choose device → confirm — that
  auto-progresses between steps and adopts the node in the background once the
  app holds a controller client, with no manual approve step. A node without an
  Ethernet uplink can now be enrolled over WiFi: meshd gains
  POST /setup/uplink (unclaimed devices only) which brings up a station
  wifi-iface + DHCP-client network from supplied home-WiFi credentials so the
  node can reach its controller, torn down with the setup AP once onboarding
  completes (internal/setupap EnableUplink). On Android the wizard offers a
  setup-AP picker (OMM-Setup-*); iOS/web keep the QR/manual path. Covered by a
  real-OpenWrt-container e2e (TestSetupUplinkE2E). See §13–§14 of
  doc/companion-app.md.
• Companion-app local dev workflow. scripts/run-dev-stack.sh plus a Vite
  dev-server proxy for all meshd REST endpoints make it possible to drive the
  companion app against a local meshd for manual enrollment testing.
  MESHD_DEV_CORS (development only) lets a cross-origin app call the management
  API directly, logging a warning when enabled. The onboarding wizard can also
  target an explicit node URL (e.g. a wired node) instead of the setup-AP default.
• Companion app packaging. The frontend now builds as native Android/iOS apps
  (Capacitor) with the QR (@capacitor-mlkit/barcode-scanning) and WiFi-join
  (@falconeta/capacitor-wifi-connect) plugins wired in; desktop ships as the
  installable PWA. Native platform projects are generated on a dev machine
  (gitignored) via npm run cap:*. See the build steps, required permissions and
  the on-device verification matrix in
  doc/companion-app-packaging.md. (Native mDNS
  awaits a Capacitor-8 plugin; until then discovery falls back to the daemon's
  /scan.)
• Release publishes the Android companion app. A v* tag now also builds
  and (with an ANDROID_KEYSTORE_BASE64 secret) signs the Android APK and
  attaches it to the GitHub Release. The job is decoupled from the meshd package
  jobs, so it never blocks the OpenWrt release; iOS (App Store/TestFlight) and
  desktop (the installable PWA) are out of scope. Required secrets are listed in
  doc/companion-app-packaging.md.
• End-to-end test for the LuCI integration. TestLuCIWorkflowE2E boots a
  real OpenWrt userland with the built meshd + luci-app-meshd packages and
  the full LuCI stack (ubusd + rpcd + uhttpd), then drives the operator
  workflows over the authenticated /ubus endpoint exactly as the PWA does:
  the ACL gate, node enrollment + adopt, the Home/profile lifecycle, and a
  wireless client device surfacing through the topology read. Runs in the e2e
  CI job; see doc/luci-integration-testing.md.

Fixed

• Creating a Home through the setup wizard failed with ubus error 5.
  Selecting a freshly created Home applied its (non-existent) profile, and the
  API treated the missing profile as a fatal 500 — even though meshd's own
  auto-select already treats it as non-fatal. Selecting a Home with no profile
  yet now succeeds; only real apply failures error.
• Opaque ubus error 5 from the LuCI Mesh Manager. The rpcd plugin used
  curl -f, which discarded meshd's JSON error body on any HTTP error, leaving
  rpcd to report a bare NO_DATA (5). The plugin now passes meshd's
  {"error": …} body back through so the PWA shows the real reason.
• Local meshd builds now run on OpenWrt. scripts/build.sh defaults to
  CGO_ENABLED=0, producing a statically-linked binary; the previous dynamic
  (glibc) build failed on OpenWrt's musl userland with
  can't execute '/usr/bin/meshd': No such file or directory.
• Onboarding wizard's default node client. A function-typed prop default was
  treated as a factory, so the default client was a function rather than a
  client; the wizard now reaches the node without an explicit createClient.

v0.1.1

Fixed

• LuCI Mesh Manager "Access denied". The LuCI host view handed the embedded
  PWA L.env.token (LuCI's CSRF token) as the /ubus session token; rpcd does
  not recognise it as a session, so every meshd ubus call the PWA made was
  rejected. The view now passes L.env.sessionid (the rpcd ubus session id),
  so the Mesh Manager works inside LuCI.

v0.1.0

First release with the full secure OpenWrt integration: a mutual-TLS mesh
control plane, a native LuCI app, and signed opkg and apk package feeds.

Added

• Mesh mutual TLS. The mesh control plane is served over mutual TLS on both
  the server and client side, using a CN-not-hostname identity model with
  trust-on-first-use.
• Per-Home PKI. Each Home is its own certificate authority; nodes are issued
  a Home-CA node certificate automatically at adoption.
• Separate control planes. Management and mesh control planes run on
  separate listeners, isolating the device-facing API from the inter-node mesh.
• LuCI app (luci-app-meshd). rpcd plugin, ACLs, and an embedded LuCI view;
  shipped in releases (architecture-independent). The PWA runs inside LuCI over
  an authenticated ubus session, auto-selecting the ubus transport when embedded.
• opkg feed. Every release publishes an opkg feed index
  (Packages/Packages.gz), usign-signed (Packages.sig) when a key is
  configured; the signing public key (omm-feed.pub) is published as an asset.
• apk packages & feed. Real, signed apk-v3 packages (apk mkpkg) plus a
  signed apk repository index (packages.adb) for OpenWrt's apk userland; the
  signing public key (omm-apk.pub) is published as an asset.
• Profile application. Home profiles are applied when the active Home
  changes, with a netifd reload so the changes take effect immediately.

Changed

• On install, uci-defaults flips meshd to its secure posture automatically.
• Discovery derives the announced API host from the packet source, removing the
  need to configure it manually.

Removed

• The Home/node-delete and factory-reset API endpoints.

Fixed

• Release version strings are now valid for apk (stricter than opkg): non-tag
  builds use 0.0.0_git instead of the apk-invalid 0.0.0-dev.
• The release publish step is idempotent — re-running a release no longer fails
  if the release already exists.
• Profiles are re-applied when the active Home changes.
• Removed duplicate meshd init/config files left at old paths.

Security

• Inter-node mesh traffic is authenticated and encrypted end-to-end (per-Home CA
  • mutual TLS).
• opkg and apk package feeds are cryptographically signed and verified on the
  device against published public keys.

v0.0.2: fix(ci): make the release publish step idempotent

`gh release create` hard-fails if a release for the tag already exists, so
re-running the workflow or re-pushing a tag broke with "release already
exists". Create the release only when missing; otherwise upload assets to
the existing one with --clobber.

v0.0.1

first release to test the pipeline