Version history

Version history

The 4.x line. For 3.x and earlier, see the Python project's own version history — everything published for it stays where it is and keeps installing.

4.1.0

Responsiveness. Tapping + or − repeatedly no longer queues up behind the unit, a request a unit ignores is asked again in two seconds rather than ten, and opening the app again no longer waits for a reconnect.

Why 4.0.2 felt slower than 3.2.0, from two days of a real server's request log and a probe watching a unit's connection packet by packet:

• every control was two round-trips — read the unit, then write it — 1.47 s where 3.2.0 took one;
• controls to a busy unit ran one after another, each holding a server thread while it waited. Ten taps finished 32 s after the tapping stopped, each answered with a temperature already tapped past, and the other units and even the plain unit list (11.7 s) waited behind them;
• a request the unit ignored cost 12.5 s: a ten-second timeout, then a reconnect. Units do ignore the odd request — about one in thirty — while the connection itself stays perfectly usable.

What changed

• Controls merge. One that arrives while the unit is busy folds into the next command, which carries every change and answers all of them.
• One round-trip per control, built on what the unit reported in the last ten seconds; the unit is read first only when nothing that recent is on hand.
• Reads share. A read that waited behind another read or a control takes its answer, so opening the app — a batch read and the live stream at once — reads each unit once, not twice.
• Nothing that only needs a name waits on a unit: GET /api/units and the Nerd panel read copies kept beside each unit.
• An ignored request is resent after 2 s on the same connection. A reply that turns up late after all is discarded rather than taken as the answer to the next command, and a connection the unit has closed is noticed before a request is written into it.
• Keeping connections warm, BREEZE_KEEP_WARM: for 30 minutes after the server was last used, each quiet unit is read every 20–25 s so its connection outlives the unit's 30-second idle timer. Any duration, off, or always. See Configuration.
• The Nerd panel shows keep_warm, whether it is active right now, and for each unit a link block — connections opened, requests resent, stale replies discarded. A unit with a weak signal shows up there first.

Measured on a real unit, same server and network:

	4.0.2	4.1.0
ten taps 0.25 s apart: last reply	32 s after the tapping stopped	1.5 s after the last tap
commands sent for those ten taps	10, two round-trips each	5, one round-trip each
GET /api/units while a unit is busy	up to 11.7 s	2 ms
a read after 45 s idle	~1.8 s, reconnecting	0.73 s

Windows installer

Installing over an existing copy upgrades it in place. Before, it re-registered the service from defaults — back to the LAN address and port 8420, --behind-proxy dropped, the environment replaced (anything added with nssm edit, BREEZE_WORKERS included) and the inbound LAN firewall rule reopened on a machine set up proxy-only. Now it stops the service, replaces the files, changes only which executable the service runs, and starts it again if it was running. It also finds a previous install directory again — it had been looking in the 32-bit registry view, where it never writes — and a service left over from the Python line is still configured from scratch, as it has to be.

The Breeze app, 2.2.8, pairs with this: it shows only the newest tap's reply, reverts a failed control to what the server last reported rather than to "before this tap", and does not let a push or poll overwrite a unit while a control for it is in flight. With both, the temperature no longer walks back through values already tapped past.

4.0.2

Settings that had quietly become constants are settings again, two new package managers, and two CLI papercuts.

Configuration

Three values were compile-time constants in 4.0.0 and are read from the environment again. One default changed:

AC_CODE_TTL	60	seconds a pairing code lives
AC_TOKEN_TTL_DAYS	3650	days a device credential lives — was 90
AC_AUTH_SKEW_SECONDS	60	signed-request clock skew, each way

Ten years rather than ninety days because the failure mode of a short lifetime is a phone that stops working while its owner is away from the LAN that could re-approve it. 0 or less means never.

Fixing them had been justified as removing "a switch whose only use was a worse configuration". For the credential lifetime that was simply wrong — on a LAN deployment the alternative to a long life is re-pairing every phone in the house on a timer — and removing the setting silently shortened it for anyone who had set it.

GET /api/system reports the live values now rather than the constants, and both worker settings are clamped where they are read, so a BREEZE_WORKERS=0 is reported as the 1 the server is actually running.

New

• Void Linux and Gentoo packages. Ten .xbps — five architectures × glibc and musl — in a signed repository, and a Gentoo overlay with app-misc/breeze-core-bin. See Ports and architectures.
• BREEZE_BG_WORKERS, how many units the background state poller contacts at once. Default 1, which is the previous behaviour exactly. Raise it when a poll pass stops fitting inside AC_STREAM_TICK — at roughly a second per unit, more than about five means every tick starts later than the last. It is not a count of background threads: the timer runner, the scheduler and the poller are one thread each because each is a singleton role, and a second scheduler would fire every program twice. See Configuration.
• Init templates for runit, s6, SysV, supervisord and launchd, in deploy/init/, alongside the systemd, OpenRC, procd and BSD rc scripts the packages already install.

Fixed

• breeze-core with no arguments started a server. It prints the usage now, which is what the reference did. The old default was justified by "an init script may simply exec the binary" — nothing does; all eleven init scripts here pass serve explicitly. The cost was paid by people instead: a curious breeze-core at a shell either collided with the running server ("Address in use") or failed to read a config it has no permission for ("Permission denied"), and since neither message mentions serving, the binary looked like it needed root to print its own help.
• A store error did not say which file. Permission denied (os error 13) and nothing else — no path, no indication that a file was even involved. Every store error names its path now, and a permission failure adds that the stores belong to the service account.
• The legacy-config hint printed as one long line, with thirty-space gaps in the middle of it, from a multi-line string literal that carried its own source indentation.
• The usage text was a wall of prose in a single paragraph. Three short groups now.

4.0.1

A bug-fix release, and most of what it fixes was found by a tool rather than by reading code.

packaging/test/differential.py runs 4.x and a live 3.2.0 against the same config.json and diffs every endpoint — status code, JSON type, shape, value. It started at 142 differences and ends at none that are not documented with a reason. That check exists because a unit test agrees with whatever the code does, so it cannot catch a misunderstanding of the contract; the reference is still installable, so the honest test is to ask both and compare.

Fixed

• A bad enum answered 422; the reference answers 400. The split is inherited rather than designed — pydantic rejected numeric bounds while parsing, so FastAPI produced its own 422, while an unknown mode name survived parsing and was refused with an explicit 400. Answering one status for everything is invisible until a client branches on it.
• The bad-enum message wording, back to the reference's Unknown mode: X.
• /api/system reported a null init system on a hardened deployment. It read /proc/1/comm and nothing else, which ProtectProc=invisible makes unreadable — and that directive is in the project's own hardening drop-in. It also meant a null init on macOS, Windows and all three BSDs, which have no /proc at all. Now checks /run/systemd/system first, as the reference does, plus OpenRC, procd, runit and s6 markers, with pid 1 kept as the last resort for the container case where the server itself is pid 1.
• Six settings.* fields were missing from /api/system — the ones 4.0.0 turned into constants. "No longer configurable" is not "no longer a fact": the panel displays them by name and showed blanks. Reported with their fixed values.
• os.* used our own field names where the reference has established ones, so distro_id, distro_version, pretty_name, system, hostname, kernel_version and libc all read blank in the panel.
• Also added to /api/system: network.bind_host / bind_port, cpu.cores / arch / endianness, paths.package, server.installed_at, and per-unit online and cached capabilities.
• capabilities sent 30.0 where the reference sends 30. JSON has one number type and consumers do not: invisible to JavaScript, fatal to a client that asks for an integer.
• b64_decode accepted url-safe base64 only; the reference accepts both. Python's urlsafe_b64decode leaves an already-present + or / valid, so standard base64 has always worked against 3.x. Refusing it broke such clients intermittently — a signature encodes without a + or / roughly 7% of the time, so about one request in fourteen succeeded, which is far worse than failing outright. Neither the web panel nor the Android app was affected; both already emit base64url.

Added

• An access log, on by default. One line per request: peer, method, path, status, duration. 4.0.0 logged a single line per process lifetime, so a misbehaving deployment left nothing to look at. BREEZE_LOG=0 silences it.
• BREEZE_DEBUG=1 traces the control path — the JSON a client sent, what the unit reported before the merge, the setpoint sent to it, what the unit echoed back — and dumps the exact canonical string behind any rejected signature.
• A reference byte vector for the control frame. get_state was pinned to msmart-ng's real bytes from the start; set_state, the frame that actually changes a unit, had none.
• Container images — see below.

Not in the 4.0.1 release

No FreeBSD, NetBSD, OpenBSD or OPNsense packages: those are built on real machines and the build hosts were unreachable. They will be added to the 4.0.1 release rather than held for the next one.

Containers, new in 4.0.1

Two images replacing the Python line's five:

Tag	Base	On disk
4.0.1	scratch	4.3 MB
4.0.1-debug	busybox:musl	5.9 MB

Against 137 MB for the Alpine image and 291 MB for UBI. All five of the old ones existed because an interpreter needs a distribution around it, and that reason is gone. amd64 and arm64. Details: Installing with containers.

4.0.0

The Rust reimplementation. Same REST API, same store files, same AC_* variables, same service name, same panel, same Android app — meant to be installed over a 3.x deployment rather than migrated to.

	3.2.0	4.0.0
what installs	interpreter + ~40 wheels, or a ~25 MB frozen bundle	one ~2.5 MB executable
package dependencies	python312 / python311 and friends	none
resident memory	~62 MB	~2.4 MB
disk installed	~59 MB	~3 MB
architectures with packages	amd64, arm64, armhf	those plus riscv64, ppc64le, s390x
OpenBSD	source install into a virtualenv	a signed pkg_add package
build	emulate the target, compile on it	cross-compile, no emulation

Those memory and disk figures are measured on one x86-64 server with the same three units paired, before and after the in-place upgrade.

What changed underneath

• The Midea LAN protocol is implemented directly. No msmart-ng, no Python: V2/V3 framing, the handshake, the encryption and discovery all live in breeze-proto, pinned against the reference implementation's own byte vectors.
• The panel is compiled in. A build script generates an include_bytes! table from static/, so there is no directory to deploy and no WorkingDirectory to get wrong.
• breeze-store reads and writes 3.x's files byte-for-byte, including fields it does not use and null for an absent optional rather than omitting the key. That is what makes the upgrade an upgrade.
• No async runtime, no web framework, no TLS server. Fourteen external crates. See Architecture.

New

• breeze-core control — set a unit from the shell, positionally, without curl and a JSON body.
• breeze-core login — enrol a machine's CLI as its own device.
• /metrics in Prometheus format.
• A signed FreeBSD pkg repository and the project's first OpenBSD package; the Python line could only manage a source install there.
• riscv64, ppc64le and s390x as first-class packages rather than proof-of-concept builds. See Ports and architectures.

What went away

• /docs. No framework, so no generated OpenAPI schema. REST API is the reference and it is complete.
• The zsh tools, replaced by subcommands. The flags they took are still accepted.
• Brotli. gzip only, deliberately.
• The ability to loosen some settings — LAN-only admin approval, the code TTL, the token TTL and the clock-skew window are constants now.

Version numbers

4.0.0 follows 3.2.0 because the implementation changed, not because the contract did. Nothing about the wire format, the store files or the service name moved, so a client cannot tell which is answering except by GET /api/version.

The major bump is a warning about what you are installing, not about what you will have to change.