Tags

v0.7.0

v0.6.2

### Added

- SHA256SUMS published alongside .deb / .rpm on tag releases.
  Enables direct consumption by Ansible's get_url checksum: and
  hand-verification via sha256sum -c. SHA-only verification is
  partial — neither the artifacts nor SHA256SUMS are signed yet;
  signing tracked separately and will land in a future release.

### Fixed

- Discovery configs no longer republish twice on initial broker
  connect. Each entity registers once per fresh-session connect,
  halving discovery traffic and journal "Registering topic" lines
  per startup.

v0.6.1

This release ships no binary changes. It's a patch-version bump to
validate the CI release flow that failed on v0.6.0 (release-cli
image deprecation surfaced as a bash-not-available failure during
release-page creation; fixed by migration to `gitlab-org/cli` and
POSIX-rewriting the changelog extraction script). Operators on
v0.6.0 do not need to upgrade; the binary is byte-identical.

v0.6.0

This release is a **breaking** modernisation of the MQTT discovery shape
to align with Home Assistant's recommended payload conventions. Operator
impact is mostly upside (per-host device card, unlocked entity-management
UI), but one entity goes away (`version`) and every entity's discovery
JSON gains new fields. Read the `### Changed` and `### Removed` notes
below before upgrading existing hosts.

- Per-cycle health-summary log line. After each successful publish cycle the
  binary writes `publish cycle complete: N topics published` at info level.
  Visible in `journalctl` / `systemctl status` as a one-line confirmation that
  the publish loop is alive, and used by the local smoke harness as a
  greppable marker for cycle-count assertions.
- **Publish-task heartbeat (Piece B) and systemd watchdog integration.**
  A second supervisor branch now watches the publish task itself: if no
  successful publish cycle completes within 180 seconds the binary
  log-and-bails (`publish-task: no progress (...)`) and systemd restarts
  it. This catches the silent-deaf class where the eventloop is healthy
  (PingResp keeps the broker connection lively) but the publish task is
  wedged — channel saturation, blocked await, the v0.5.x field signature
  where the binary went silent for thirteen minutes after a
  keepalive-timeout reconnect. The unit file gains `Type=notify`,
  `NotifyAccess=main`, and `WatchdogSec=360`; the binary signals
  `READY=1` after registering with HA and `WATCHDOG=1` after each
  successful publish cycle. WatchdogSec sits as the outermost layer for
  hangs not caught by either supervisor branch (hard signals,
  pathological scheduler stalls). Operator effect: the
  `availability offline` LWT is now a reliable disconnect signal —
  the binary either publishes correctly or is killed and restarted
  cleanly, no third state where it's running but silent.
- New `version` subcommand: `system-mqtt version` prints the running binary's
  version to stdout. Useful for confirming which build is installed on a host
  without parsing systemd unit files or package manifests.
- Startup-time legacy-topic sweep. On every MQTT connect the binary publishes
  empty-retained to a list of deprecated discovery topic paths, clearing
  ghost entities left on the broker by prior versions. Currently sweeps the
  pre-v0.4.0 `available` topic (LWT-rename casualty) and the post-v0.5.1
  `version` topic (removed in this release). The list is a `const` in
  `src/main.rs`; future entity removals append a template.

- systemd unit file gained `StartLimitIntervalSec=600` and
  `StartLimitBurst=10`. Previously a sustained network or broker outage
  would produce an indefinite restart loop (`RestartSec=10` made the
  default 10s/5-burst limit mathematically unreachable). After 10
  restarts in 600s the unit now enters failed state; operator-visible
  via `systemctl status system-mqtt`, and Home Assistant sees the
  LWT-driven `availability offline` and stays there durably rather than
  flickering. Recovery once the underlying issue is resolved:
  `systemctl reset-failed system-mqtt && systemctl start system-mqtt`.
- Supervisor now detects eventloop task death and stuck `poll()` calls via a
  timer-driven no-progress check and a `JoinHandle` health check, in addition
  to the existing event-driven detection. Defensive coverage; certain wedge
  classes that previously required manual intervention are now caught and
  trigger systemd restart within ~120s. Supervisor log-and-bail messages
  follow a new `<source>: <description> (<stale signal>)` format with
  measured stale-signal parentheticals — easier to grep and post-mortem in
  `journalctl`.
- **Discovery payloads now include `unique_id` and `device.identifiers`
  on every entity.** The standalone `availability` sensor additionally
  carries the rich device-metadata block (`name`, `sw_version`) per
  Home Assistant's shared-device pattern. Operator impact: entities now
  group under a per-host device card in HA (Settings → Devices &
  Services → MQTT) instead of appearing ungrouped, and the cog-menu UI
  controls (entity_id, friendly_name, icon, area) unlock for every
  entity. `sw_version` on the device card is sourced from the same
  `CARGO_PKG_VERSION` constant used by the `version` subcommand.

- `ExecReload=` directive from the systemd unit. The binary doesn't
  handle SIGHUP, so the prior `ExecReload=/bin/kill -HUP $MAINPID`
  setting made `systemctl reload system-mqtt` kill the process and
  let `Restart=on-failure` revive it — restart-as-reload semantics
  that surprised more than they helped. With the directive removed
  `systemctl reload` now correctly fails (`Job type reload is not
  applicable...`); use `systemctl restart system-mqtt` to pick up
  config changes.
- **Standalone `version` MQTT topic and its discovery registration.**
  The topic was a placeholder until `sw_version` had a canonical home;
  that home now exists on the per-host device card via the
  availability sensor's rich device-metadata block (see `### Changed`).
  Operator impact: the standalone "version" sensor disappears from
  Home Assistant; any operator automations referencing the
  `system-mqtt/<HOST>/version` MQTT topic will need to migrate to
  reading `sw_version` off the device card instead. Existing brokers
  with a retained `version/config` discovery topic from a prior
  release are cleared automatically by the new legacy-topic sweep
  (see `### Added`). The `system-mqtt version` CLI subcommand is
  unchanged and still prints the running binary's version on demand.

- **Discovery configs now republish on broker fresh-session reconnect.**
  When the broker accepts a reconnect with `session_present=false` —
  which it does after a broker restart, retained-state eviction, or
  any other broker-side state loss — the binary re-publishes the full
  discovery config set, the legacy-topic empty-retained sweep, and the
  `availability online` retained payload. Previously the binary
  registered with Home Assistant once at process startup and never
  re-emitted, so a broker that lost retained state across a restart
  would leave HA seeing entities marked unavailable or vanished until
  the binary itself was restarted. State-topic publishes were
  unaffected, so the loss was easy to miss in journal output but
  obvious in HA's UI. Now operator-recovery is automatic on the next
  successful reconnect; no manual `systemctl restart system-mqtt`
  needed.
- Swap percentage now correctly computed as `used_swap / total_swap` rather
  than `used_swap / free_swap`. Published values were previously the wrong
  ratio (and clamped to 0–100% by accident on most hosts); after upgrade the
  swap graph in Home Assistant will show a step-change to the correct
  utilisation figure.
- Discovery payloads for the `availability`, `uptime`, and `battery_state`
  sensors no longer include an empty-string `state_class` key. The field
  is now omitted entirely on these sensors (it stays present with its
  valid value on `cpu` / `memory` / `swap` / `battery_level` / drive
  topics). Clears the schema-validation warnings Home Assistant logs on
  entity registration.
- Per-cycle swap publish now skipped on hosts with no swap configured
  (`total_swap() == 0`), instead of publishing the literal string `"NaN"`
  from a 0/0 division. Home Assistant on no-swap hosts will see the swap
  sensor stop receiving updates rather than receiving an unparseable
  value; the discovery registration for the swap topic is unchanged so
  the entity still appears.

v0.5.1

v0.5.1: CI validation release

No code changes since v0.5.0. Purpose is to exercise the tag
pipeline end-to-end with the four .gitlab-ci.yml changes that
landed between v0.5.0 and now: binstall for cargo-audit /
cargo-deb / cargo-generate-rpm, rust:slim base image, and
pull-only cache policy on test/clippy.

Main-branch pipeline validated on 4e0a7d3 (pipeline
release stages for the first time under the new configuration.

v0.5.0

v0.5.0: rumqttc → rumqttc-v4-next fork (temporary bridge)

Clears five RustSec advisories (RUSTSEC-2026-0099 / 0049 / 0098 /
0104, RUSTSEC-2025-0134) against rumqttc 0.25.1's rustls-webpki /
rustls-pemfile pins. Migration framed explicitly as a temporary
bridge pending upstream rumqttc resumption (bytebeamio/rumqtt#1029);
see docs/decisions.md entry dated 2026-04-24 for the full context.

Opportunistic: enables rustls ring backend via the fork's
use-rustls-ring feature, sheds aws_lc_sys. Stripped release binary
6.52 MiB → 4.56 MiB (-30.1%).

v0.4.3

v0.4.3 — CI publishes GitLab Release with .deb + .rpm attached on
tag push. Native apt/dnf install via Package Registry parked upstream
(Debian Experiment, RPM not GA on gitlab.com). No wire or contract
changes. Patch bumped from 0.4.2 rather than repeated because the
prior 0.4.2 artifact existed only as a locally-built test package,
and install-state on the test host is unverified.