v3.4.0-beta.21
Pre-release
Pre-release
Requires pylxpweb 0.9.36b23 (installed automatically; the manifest requirement is bumped).
This release is the outcome of a four-scanner adversarial bug sweep (two independent Claude passes, Codex, and Antigravity) over the integration and the pylxpweb library, followed by a dead-code cleanup of both. Every fix was independently verified before implementation and every PR passed a dual Opus + Codex review gate.
Fixed
- LOCAL: a physically removed battery now disappears within the staleness bound instead of surviving until a Home Assistant restart (#300): the 6-hour eviction added in beta.19 sat in a branch that only ran when a poll returned zero batteries — but the round-robin merge always returns the full accumulated cache, so on any normal poll the eviction was unreachable and a removed pack's entities stayed populated with frozen values indefinitely. The eviction now runs unconditionally on every merge (matching the HYBRID/CLOUD behavior shipped in beta.19), and the same bound now also covers the HYBRID coordinator's local-only branch, which was previously unbounded.
- HYBRID: schedule time, number and select controls now fall back to the cloud when the local write fails or the local link is down (#301): switches have always retried a failed local write through the cloud API, but every other control type chose the local path purely because a transport was attached — and pylxpweb keeps the transport attached during a link outage — so changing a charge schedule, SOC limit, PV input mode or battery control mode failed until the link recovered even with a healthy cloud connection. All control writes now share the switches' local-attempt-then-cloud semantics, prefer the cloud immediately when the library reports the link down (no timeout wait), and the post-write parameter refreshes are link-down-aware so the service call can't stall against a dead transport. After a successful cloud fallback the acknowledged values are seeded into the parameter cache, so the entity converges on what was written instead of reverting to the stale pre-write value until link recovery. Excluded by design: Quick Charge Duration and the Start Charge threshold (no equivalent cloud write exists).
- Switches and selects now report unavailable during a sustained coordinator outage, like every other entity type (#303): sensors, numbers and time entities already gated availability on coordinator health; switches and selects were missed in 2025-12 and stayed clickable against stale data through a full outage. The gate only engages after three consecutive failed update cycles and single-device or transient failures never trip it, so day-to-day behavior is unchanged.
- Library (pylxpweb 0.9.36b23): a truncated-but-well-formed holding-register response from the dongle or Modbus TCP could silently produce a partial parameter read that blanked parameter-backed entities for up to an hour (pylxpweb#203); a battery whose electrical data arrived before its serial number on a cold start could mint a permanent duplicate battery with frozen values (pylxpweb#204); public cloud-mode schedule getters always returned 00:00 (pylxpweb#205, no integration impact); the link-down probe no longer runs the full six-group input read on every degraded poll, cutting dead-link poll pressure roughly 6× (pylxpweb#205).
Changed
- Dead-code cleanup, both repositories (#299, pylxpweb#202): −106 integration lines and −26 library lines of verified-unreferenced code. Zero functional change — entity IDs, unique IDs and translation keys are byte-identical, verified by both review gates.