Fix the issues found in the review of 1.0.3

CHANGELOG.md

@@ -3,6 +3,59 @@
 All notable changes to this project are recorded here. The format follows
 Keep a Changelog; versions follow Semantic Versioning.
 
+## 1.0.4 - 2026-09-06
+
+Fixes from a fourth review, of 1.0.3, which drove the real socket path the
+test suite fakes and found two behaviours the original library had and
+this one had lost. One device fix carried from upstream.
+
+### Fixed
+
+- A connected socket that went bad (interface bounce, host address change,
+  container network restart) was never replaced: the request waited out
+  its timeout and every later request did the same until `aclose()`. The
+  original opened a socket per call, so it healed on the next one. Now a
+  send failure the socket reports (no route, address gone) fails the
+  waiting request at once with that `OSError`, and a request that fails
+  for a network reason (that, or a timeout) drops the socket so the next
+  call opens a fresh one. A transport asyncio closes from its side also
+  wakes the waiting request instead of leaving it to time out. An ICMP
+  "port unreachable" (a host that is up with nothing listening, or a
+  device mid-reboot) is logged and treated as silence, since the
+  original's unconnected socket never saw those, so the timeout decides
+  as before.
+- `discover()`, `hello()`, `ping()` and `setup()` passed hostnames straight
+  to `sendto`, which resolved them with a blocking call on the event loop
+  and swallowed the failure: a name that did not resolve made `hello()`
+  wait out its timeout and `ping()` return without sending. The
+  destination is now resolved once, off the loop, and `socket.gaierror`
+  propagates as it did from the original's socket. A send failure in
+  `ping()` and `setup()` is raised too.
+- The A2 air quality sensor's request frame was two bytes short and
+  declared the wrong length, and real units answered every read with
+  error -5. The frame now follows the SP4/LB1 layout, which is byte for
+  byte the packet upstream pull request #826 tested on an A2. That is the
+  one oracle case re-recorded on purpose; the fix is carried on the
+  strength of that report, not of hardware we have.
+- `xdiscover()` closes the `scan()` generator it wraps, so the discovery
+  socket is closed when the caller stops iterating rather than by the
+  finalizer a few turns later (1.0.2 claimed this and only `Device.hello()`
+  did it).
+- Two identical captures compare equal: `CapturedSignal.captured_at` no
+  longer takes part in equality or hashing.
+- Async generator functions are annotated `AsyncGenerator`, which has the
+  `aclose()` the library and the README call; `AsyncIterator` does not.
+- The `TICK` docstring tells the same story as the README: 8192/269 from
+  protocol.md's measured conversion, not a 32768 Hz clock.
+
+### Added
+
+- A loopback test module that drives the real datagram endpoint, including
+  the socket-error path, since every other transport test fakes it.
+- README: which errors `discover()` and `hello()` raise, that a failed
+  request drops its socket, and that `CaptureInProgressError` can come
+  from the `capture()` call or from the first iteration.
+
 ## 1.0.3 - 2026-09-06
 
 Fixes from a third review, this one of 1.0.2. No change to the wire

README.md

@@ -162,6 +162,11 @@ If the device is locked, it may not be discoverable with broadcast. In such case
 device = await broadlink.hello("192.168.0.16")
 ```
 
+`discover()` and `hello()` raise `NetworkTimeoutError` when nothing answers
+within the timeout, `socket.gaierror` when a hostname does not resolve, and
+`OSError` when the socket cannot be opened or the send fails (no route, for
+example), the same errors the original library raised from its socket.
+
 If you are a perfomance freak, use `broadlink.xdiscover()` to create devices instantly:
 ```python3
 async for device in broadlink.xdiscover():
@@ -202,9 +207,14 @@ await device.aclose()
 
 The socket reopens by itself on the next call, so closing is cheap and
 safe to do at any time. A request that is in flight when `aclose()` runs
-fails with `EndpointClosedError`. An integration that creates devices
-should close them when it unloads; a device that is never closed holds
-its socket until it is garbage collected.
+fails with `EndpointClosedError`. A request that fails for a network
+reason (a timeout, or an `OSError` from the socket such as "network is
+unreachable" after an interface change, raised at once) also drops the socket, so the
+next call starts fresh rather than reusing one that has gone bad, which
+is how the original library behaved by opening a socket per call. An
+integration that creates devices should close them when it unloads; a
+device that is never closed holds its socket until it is garbage
+collected.
 
 The next steps depend on the type of device you want to control.
 
@@ -280,7 +290,9 @@ By default the window closes after the first signal. Pass
 because the device holds only one code per learning session. A universal
 remote has a single receiver, so only one capture window can be open on a
 device at a time: opening a second one raises `CaptureInProgressError`
-while the first is still held. Always close a window you leave early
+while the first is still held, either from the `capture()` call itself or
+from the new window's first iteration, depending on what the first window
+was doing at that moment. Always close a window you leave early
 (`aclosing` above does it), otherwise it stays open until Python collects
 the generator.
 

broadlink/__init__.py

@@ -1,7 +1,7 @@
 """The python-broadlink library."""
 
 import contextlib
-from collections.abc import AsyncIterator
+from collections.abc import AsyncGenerator
 
 from . import exceptions as e
 from .alarm import S1C
@@ -294,15 +294,15 @@ async def xdiscover(
     local_ip_address: str | None = None,
     discover_ip_address: str = DEFAULT_BCAST_ADDR,
     discover_ip_port: int = DEFAULT_PORT,
-) -> AsyncIterator[Device]:
+) -> AsyncGenerator[Device]:
     """Discover devices connected to the local network.
 
     Yields each device as soon as it answers.
     """
-    async for resp in scan(
-        timeout, local_ip_address, discover_ip_address, discover_ip_port
-    ):
-        yield gendevice(*resp)
+    responses = scan(timeout, local_ip_address, discover_ip_address, discover_ip_port)
+    async with contextlib.aclosing(responses):
+        async for resp in responses:
+            yield gendevice(*resp)
 
 
 # Setup a new Broadlink device via AP Mode. Review the README to see how to enter AP Mode.

broadlink/device.py

@@ -15,7 +15,7 @@
 import logging
 import random
 import socket
-from collections.abc import AsyncIterator
+from collections.abc import AsyncGenerator
 
 from cryptography.hazmat.backends import default_backend
 from cryptography.hazmat.primitives.ciphers import Cipher, algorithms, modes
@@ -48,12 +48,20 @@
 _CLOSED = (None, None)
 """Sentinel put on the receive queue when the endpoint is closed."""
 
+_QueueItem = tuple[bytes | Exception | None, tuple[str, int] | None]
+"""What the receive queue carries: a datagram with its source address, an
+error the socket reported (address ``None``), or ``_CLOSED``."""
+
 
 class _Protocol(asyncio.DatagramProtocol):
-    """Datagram protocol that hands every received packet to a queue."""
+    """Datagram protocol that hands every received packet to a queue.
+
+    Errors the socket reports go on the same queue, so the request that is
+    waiting fails at once instead of waiting out its timeout.
+    """
 
     def __init__(self) -> None:
-        self.queue: asyncio.Queue[tuple[bytes, tuple[str, int]]] = asyncio.Queue()
+        self.queue: asyncio.Queue[_QueueItem] = asyncio.Queue()
         self.transport: asyncio.DatagramTransport | None = None
 
     def connection_made(self, transport: asyncio.BaseTransport) -> None:
@@ -65,18 +73,30 @@ def datagram_received(self, data: bytes, addr: tuple[str, int]) -> None:
         self.queue.put_nowait((data, addr))
 
     def error_received(self, exc: Exception) -> None:
-        # ICMP unreachable and the like. Surface it as a receive of nothing;
-        # the retry loop will time out and raise NetworkTimeoutError.
-        pass
+        """Queue a send failure or an ICMP error for the waiting request."""
+        self.queue.put_nowait((exc, None))
 
     def connection_lost(self, exc: Exception | None) -> None:
-        """Nothing to do; a waiting request is told through the queue."""
+        """Wake the waiting request if asyncio closed the transport on us."""
+        self.queue.put_nowait((exc, None) if exc is not None else _CLOSED)
 
     def drain(self) -> None:
         """Drop anything that arrived before the current request."""
         while not self.queue.empty():
             self.queue.get_nowait()
 
+    def raise_if_error(self) -> None:
+        """Raise the error a send just reported, if it reported one.
+
+        asyncio delivers a failed ``sendto`` to ``error_received`` before
+        ``sendto`` returns, so a fire-and-forget sender can check right
+        after sending and raise the ``OSError`` the way a plain socket did.
+        """
+        while not self.queue.empty():
+            item, _ = self.queue.get_nowait()
+            if isinstance(item, Exception):
+                raise item
+
 
 async def _open_endpoint(
     local_addr: tuple[str, int] | None = None,
@@ -95,6 +115,21 @@ async def _open_endpoint(
     return transport, protocol  # type: ignore[return-value]
 
 
+async def _resolve(host: str, port: int) -> tuple[str, int]:
+    """Resolve a destination once, off the event loop.
+
+    Sending to a hostname through an unconnected datagram socket would
+    resolve it with a blocking call on the loop and hide the failure. A
+    name that does not resolve raises ``socket.gaierror`` here, as the
+    original library's ``sendto`` did.
+    """
+    loop = asyncio.get_running_loop()
+    info = await loop.getaddrinfo(
+        host, port, family=socket.AF_INET, type=socket.SOCK_DGRAM
+    )
+    return info[0][4][:2]  # type: ignore[return-value]
+
+
 def _hello_packet(local_ip_address: str, port: int) -> bytearray:
     packet = bytearray(0x30)
     packet[0x08:0x14] = Datetime.pack(Datetime.now())
@@ -119,13 +154,14 @@ async def scan(
     local_ip_address: str | None = None,
     discover_ip_address: str = DEFAULT_BCAST_ADDR,
     discover_ip_port: int = DEFAULT_PORT,
-) -> AsyncIterator[HelloResponse]:
+) -> AsyncGenerator[HelloResponse]:
     """Broadcast a hello message and yield responses as they arrive.
 
     The hello is repeated every ``DEFAULT_RETRY_INTVL`` seconds until
     ``timeout`` elapses. Each device is yielded once.
     """
     local_addr = (local_ip_address, 0) if local_ip_address else None
+    target = await _resolve(discover_ip_address, discover_ip_port)
     transport, protocol = await _open_endpoint(local_addr=local_addr, broadcast=True)
     try:
         if local_ip_address:
@@ -140,7 +176,7 @@ async def scan(
         discovered: set[tuple[tuple[str, int], bytes, int]] = set()
 
         while (loop.time() - start) < timeout:
-            transport.sendto(packet, (discover_ip_address, discover_ip_port))
+            transport.sendto(packet, target)
             deadline = min(DEFAULT_RETRY_INTVL, timeout - (loop.time() - start))
             slot_end = loop.time() + deadline
             while True:
@@ -151,7 +187,11 @@ async def scan(
                     resp, host = await asyncio.wait_for(protocol.queue.get(), remaining)
                 except TimeoutError:
                     break
-                if len(resp) < 0x80:
+                if resp is None:
+                    return  # The transport was closed under us.
+                if isinstance(resp, Exception):
+                    raise resp
+                if host is None or len(resp) < 0x80:
                     continue
                 entry = _parse_hello(resp, host)
                 key = (entry[1], entry[2], entry[0])
@@ -167,9 +207,11 @@ async def send_setup_packet(
     payload: bytes, ip_address: str, port: int = DEFAULT_PORT
 ) -> None:
     """Broadcast one Wi-Fi provisioning packet to a device in AP mode."""
-    transport, _ = await _open_endpoint(broadcast=True)
+    target = await _resolve(ip_address, port)
+    transport, protocol = await _open_endpoint(broadcast=True)
     try:
-        transport.sendto(payload, (ip_address, port))
+        transport.sendto(payload, target)
+        protocol.raise_if_error()
     finally:
         transport.close()
 
@@ -181,11 +223,13 @@ async def ping(ip_address: str, port: int = DEFAULT_PORT) -> None:
     Useful to prevent reboots when the cloud cannot be reached.
     It must be sent every 2 minutes in such cases.
     """
-    transport, _ = await _open_endpoint(broadcast=True)
+    target = await _resolve(ip_address, port)
+    transport, protocol = await _open_endpoint(broadcast=True)
     try:
         packet = bytearray(0x30)
         packet[0x26] = 1
-        transport.sendto(packet, (ip_address, port))
+        transport.sendto(packet, target)
+        protocol.raise_if_error()
     finally:
         transport.close()
 
@@ -407,6 +451,22 @@ async def aclose(self) -> None:
         if protocol is not None:
             protocol.queue.put_nowait(_CLOSED)  # type: ignore[arg-type]
 
+    def _drop_endpoint(self) -> None:
+        """Throw the endpoint away after a failure; the next call reopens it.
+
+        A connected datagram socket can go bad for good (the interface
+        bounced, the host's address changed), and the original library
+        never noticed because it opened a socket per call. Dropping the
+        endpoint whenever a request fails restores that self-healing.
+        """
+        transport = self._transport
+        self._transport = None
+        self._protocol = None
+        self._endpoint_addr = None
+        if transport is not None:
+            transport.close()
+            _LOGGER.debug("%s: endpoint dropped after a failure", self.host[0])
+
     async def _endpoint(self) -> tuple[asyncio.DatagramTransport, _Protocol]:
         if self._transport is not None and self._endpoint_addr != self.host:
             # The caller changed host; the connected socket points at the
@@ -510,6 +570,19 @@ async def _exchange(self, packet: bytes) -> bytes:
                     raise e.EndpointClosedError(
                         -4013, "Endpoint closed", "The device endpoint was closed"
                     )
+                if isinstance(resp, ConnectionRefusedError):
+                    # ICMP port unreachable: the host is up and nothing is
+                    # listening, or the device is rebooting. The original
+                    # library's unconnected socket never saw these, so keep
+                    # waiting and let the timeout decide, as it did.
+                    _LOGGER.debug("%s: port unreachable, still waiting", self.host[0])
+                    continue
+                if isinstance(resp, Exception):
+                    # A send failure (no route, address gone) or a fatal
+                    # transport error: fail now and throw the socket away.
+                    _LOGGER.debug("%s: socket error: %s", self.host[0], resp)
+                    self._drop_endpoint()
+                    raise resp
                 resp = self._validate(resp)
                 reply_count = int.from_bytes(resp[0x28:0x2A], "little")
                 if reply_count == count or reply_count not in self._recent:
@@ -521,6 +594,7 @@ async def _exchange(self, packet: bytes) -> bytes:
                 )
             if loop.time() - start >= timeout:
                 _LOGGER.debug("%s: no reply within %ss", self.host[0], timeout)
+                self._drop_endpoint()
                 raise e.NetworkTimeoutError(
                     -4000,
                     "Network timeout",