CHANGELOG.md

Changelog

All notable changes to this project are documented in this file.

The format is inspired by Keep a Changelog, with one practical rule for this repository: the GitHub release body should be rendered from the matching version section here.

[Unreleased]

Added

• Added an offline MPPT-frame inspector for maintainers, with strict runtime decoding and explicit warnings about overlapping wire formats. That tool does not poll devices or create Home Assistant entities; live optional PV is a separate admission path below.

• Added a generic read-only EyeBond Short-ASCII family profile for the captured MP/Q1/MD dialect seen on some Anern and Maxinn devices (#45). It provides grid/output voltage, load percentage, output frequency, temperature and status diagnostics. Battery Reference Voltage is shown separately from full-pack voltage. Identity is family-only: no retail model (including no Anern SPI-12000W claim), serial number or inverter controls are inferred.

• Short-ASCII also publishes labelled Estimated AC Load Power (estimated_ac_load_power = load% × rated VA from F). It is an estimate, not measured active power; wire load_percent stays separate. Output frequency is not aliased as grid frequency (#45).

• Short-ASCII devices can also expose documented BMS and rated readings through optional read-only RB/F/RH queries. These entities are disabled by default. Missing, invalid or expired BMS samples remove their old values without interrupting basic telemetry. Impossible V/SoC/I/P fields are hard-rejected; the known RB link-loss signature (V=0 ∧ SoC=0) may hold last-good BMS values for up to 180 s without refreshing sample age. Diagnostic counters track link-loss transitions and hard rejects; confirmation-hold pending stays unused in this release. Charge/discharge currents and measured battery_power publish at 19B4 segment-7 multiply=0.1 only while optional RH reports BMS current display accuracy 1 (with decimals) and F ratings are present for I/P bounds; RH 0, unread, expired or failed RH omits those keys. RH=1 is the discriminator (no retail-model gate). One live family member has SmartValue correlation for that path; other members (e.g. Maxinn) are not separately live-proven. This is not full device or control support (#45).

• Optional short-ASCII live PV/MPPT uses the shared framed async_send_auxiliary_read facade for the documented 21-byte 0200 runtime query only (settings 0202 is not live telemetry; no tip AABB harvest or EyeBond TID 0xAABB as aux). Values follow the same OptionalSample pattern as RB (30 s solicit / 60 s TTL), at most one optional read per successful Q1 cycle. Nine matching schema sensors (pv_*, mppt_*, dc_load_current) stay disabled by default; mppt_error_code is diagnostic and quiet. Offline decode_short_ascii_mppt remains a maintainer capture tool, not this live path. Site live-qualification / Support Archive RECEIPT for the family cutover is still pending (#45).

Fixed

• When both are due in one Q1 cycle, short-ASCII optional reads prefer FC4 RB/F/RH over the aux MPPT solicit so a virgin MPPT due time cannot starve BMS or rated refreshes (#45).

• Support archives preserve numeric-looking hex sequences in explicitly typed wire-evidence fields instead of replacing their bytes with identifier masks. Embedded ASCII identifiers remain masked. Short-ASCII archive creation and authenticated download are covered for successful, partial, malformed and offline reads through real Home Assistant.

• Retired collector readers can no longer deliver buffered replies, identity observations or read errors to a replacement connection during slow socket shutdown. Auxiliary read results also retain their original session owner.

• Queued collector commands now stay bound to the TCP session on which they started. A reconnect cannot redirect an old read, write or UART bootstrap to the replacement socket. A confirmed reply followed by normal peer closure remains a valid result.

• Closing an old collector socket no longer clears requests that already belong to its replacement connection. This covers both framed and AT sessions when socket shutdown is delayed.

• Confirmed catalog-backed read-only metadata can now survive an entry reload without a controls profile. Schema-only hints with no matching catalog evidence remain invalid; this change does not enable inverter writes.

• Callback identity checks now return their borrowed shared-listener reference even if sending the trigger fails or the operation is cancelled before waiting for a session. Existing session ownership and callback retry timing are unchanged (#45).

• Proxy capture now offers an explicit reconnect-and-start attempt when only the last connection snapshot is offline and the route is known. The existing live endpoint checks still run before any redirect; inverter identification is not required (#43).

• Read-only cloud analysis now retains its results for collector-only entries with no local inverter driver. Recognized cloud readings no longer cause the result-building step to fail while checking local sensor coverage (#23). Failure artifacts now also include the selected source, workflow stage and exception category without exception messages or credentials.

• Proxy-capture guidance now distinguishes the collector's configured server address from Home Assistant's callback address (#43). Capture still requires a connected collector and a known endpoint that can be safely restored.

[0.3.0-beta.5] - 2026-09-09

Added

• Added a separate SmartClient / ShineMonitor source for read-only device analysis. It collects exact-identity cloud readings, setting descriptions, available daily history and raw-packet fingerprints without changing the collector endpoint or sending control commands. Existing SmartESS/DESSMonitor defaults and active learning are unchanged; this source does not yet offer active verification or a built-in map for previously unsupported PV inverters.
• Added protocol-level fallback profiles for unknown SMG models reporting protocol 1, 2, or 11. Protocols 1 and 11 expose 30 compatible basic controls; the July 2024 GM6200 protocol-2 document supplies a separate 54-capability profile, CT power, schedules, clock and PV-energy decoding. All generic controls remain untested and require Full Control; blocked operations stay unavailable. Exact model profiles retain their existing behavior.
• Added version-specific, document-backed write profiles for Anenji Communication Protocol No. 3-10 variants 3, 4, 5, and 6. The exact HHS-11kW Protocol 3 fingerprint can now opt into its documented controls through Full Control; all new controls remain untested, and destructive reset/counter-clear operations stay blocked.
• Added built-in detection and telemetry mapping for the Sandisolar SD 11KP48V WIFI fingerprint (layout=4, model=0x8003) from issue #13 evidence. Its two hardware-confirmed Secondary Priority controls are available normally; the other compatible controls remain explicitly untested and Full-Control-only.
• Added a typed support-acquisition boundary so read-only cloud evidence can be collected for an identified collector before an inverter driver is known.
• Added 118 opt-in, document-backed settings for the exact Kevolt PD0080G-TPM-EU / Deye-compatible 8 kW fingerprint. They remain untested, hidden in Auto mode, and disabled by default; six Time of Use schedule fields use native Home Assistant time entities.

Changed

• Marked the documented Anenji 0x8401 Protocol 4 control surface as tested for this exact fingerprint after device-owner validation included working OP2 controls that are not exposed by the vendor apps. The irreversible generation-data clear and user-parameter reset actions remain blocked by the independent destructive-action safety policy even though the profile is treated as compatible.
• Unknown inverter models that report documented Anenji Communication Protocol 3, 4, 5, or 6 now select the matching protocol-specific telemetry map automatically instead of requiring a new exact-model catalog entry. Their document-backed controls remain untested and monitoring-only in Auto mode; users can opt into non-blocked controls through Full Control.
• Replaced the ambiguous writable modbus_smg/base.json profile with one explicit, document-backed classic SMG RS232 V1 map shared by compatible layout 1/2/11 model profiles. Exact hardware overlays retain their existing tested controls and model-only registers. Unknown fingerprints now use the compatible protocol profiles described above, not a known model's tested write surface. The V1 document version is no longer presented as proof that register 184 must contain protocol number 1.
• Rebuilt the Anenji Protocol No. 3-10 telemetry layer as one documented shared schema with separate Protocol 3 and Protocol 4 output projections. ANJ-11KW, HHS-11kW and Sandisolar fingerprints now select registers by protocol identity, never by a zero/non-zero value fallback, while ordinary SMG Protocol 1 devices retain their existing map.
• Moved device-support analysis and collector cloud-traffic capture under one Expand device support menu with separate readiness rules for read-only metadata, active learning, and proxy capture.
• Active-learning consent now explicitly covers the temporary collector endpoint change, bounded cloud test commands, local interception, and route restoration.
• Large Modbus settings surfaces now refresh one compact block per poll and keep their cache in non-persisted per-session driver state. Confirmed writes update that same cache only after immediate wire read-back.

Fixed

• A failed temporary proxy or cloud-learning connection no longer disables normal collector callback recovery. Auxiliary port conflicts and upstream failures preserve the primary listener's status; cancelled cloud-route startup also releases its resources and ownership before a retry.
• Collector protocol hints no longer label SmartClient/SmartValue devices as SmartESS merely because they expose the same metadata fields. Cloud source selection remains explicit and independent of those hints.
• Device-learning credentials preserve intentional password whitespace before cloud authentication.
• Modbus exception 03 no longer claims that a rejected write was necessarily outside the inverter's allowed setting range. Diagnostics distinguish the profile's UI bounds from the device's unexplained rejection; write validation and confirmation behavior are unchanged.
• Schema-driven Modbus polling now reports a failed cycle when every runtime block fails, instead of publishing an empty successful snapshot. Cached settings cannot hide the failure; valid partial telemetry remains available, and a disconnected transport stops the sweep immediately.
• Fixed long diagnostic states exceeding Home Assistant's 255-character limit. Write-capability sensors now show a count with the full list in the capabilities attribute; other long texts retain their full value in full_value. Support archives keep the original data.
• Fixed false rejection of successful secondary-priority schedule writes: normal SMG polling now uses the same native capability decoder as write confirmation, including HHMM times such as 655 → 06:55.
• Updated device-registry access for Home Assistant 2026.9, preserving entry-scoped ownership, collector links, entity ids and older HA compatibility. The current real-HA CI lane now runs on 2026.9.1.
• Moved SMG battery voltage and Protocol 3-10 lifetime PV generation out of the diagnostic category without changing registers, entity ids or statistics.
• Corrected the exact Anenji 0x8401 display name to ANJ-11KW-48V-WIFI after owner clarification; existing profile keys and the 0x8000 WIFI-P identity remain unchanged. The fingerprint is not a parallel-capability test.
• Fixed a framed-transport desynchronization where one unexpected unwrapped Modbus RTU response could be combined with the next EyeBond header, consume subsequent valid frames, and stall telemetry until the socket was replaced. Malformed, oversized, or incomplete frames now close only the affected session with a typed diagnostic reason, after which a fresh callback can resume normally.
• Preload the SmartESS semantic catalog with the other JSON metadata caches in Home Assistant's executor, avoiding a first-use blocking file read from the event loop when diagnostics or cloud evidence first resolves a field.
• Restored exact Anenji 11kW identification for the captured Protocol 4 fingerprint layout=4, model=0x8401. Unlike the broad legacy heuristic, the new catalog entry affects only this proven fingerprint; it uses the documented Protocol 4 telemetry map and its own control profile. The old displayed name remains a compatibility alias so an existing entry cannot accidentally select the other hardware variant before a live probe.
• Fixed cancellation cleanup for an in-progress inverter driver sweep on Python 3.14. Expected late probe failures are now consumed by an owned no-fail outcome relay instead of being reported as unhandled event-loop exceptions after Home Assistant cancels or replaces the refresh.
• Fixed AT-text collector restart verification by separating the vendor's dedicated soft-reset command from its staged-settings apply command. A command acknowledgement still does not count as recovery until the old session drops and the same collector reconnects on a new physical session.
• Fixed canonical power-flow estimates that could count the same PV watts both as home consumption and battery charging when asynchronously sampled Protocol 3/4 registers disagreed. Instantaneous routes and their accumulated energy now share bounded source/sink budgets and leave unexplained remainder unattributed instead of inventing a second power path.
• Fixed stale callback-session inventory that could outlive its physical TCP socket, suppress a new callback request, and leave Home Assistant waiting for a connection that no longer existed.
• Fixed Protocol 3/4 optional-register polling so one silent collector session fails fast instead of expanding into minutes of per-register retries. Support archives also refresh their runtime snapshot after a capture reconnect.
• Fixed fully silent callback collectors that open a TCP session without an initial payload. Home Assistant now performs one bounded FC=1 identity challenge on the exact observed session, accepts only a correlated full-PN response, and never falls back to peer-IP routing. Unknown wire dialects are tried only on separate callback attempts and separate physical sockets.
• Corrected the canonical grid and inverter power projection for documented Protocol 3/4 layouts. Grid Power now comes from holding register 340, with protocol-specific output/load blocks; legacy total fields remain diagnostic evidence rather than speculative runtime fallbacks.
• Corrected the Kevolt PD0080G-TPM-EU rated-power scale from 80 kW to its hardware-confirmed 8 kW rating without changing the stable internal fingerprint key or its already-correct battery-value sign.
• Treat a catalog-declared previous model name as the same runtime inverter identity after a display-name correction. Unknown and ambiguous aliases still fail closed instead of replacing a durable binding.
• Fixed the admission failure menu so switching a failed selected-route attempt to manual setup starts a fresh callback transaction instead of raising an internal lifecycle error into the Home Assistant UI.
• Preserve the exact collector-management error in the normal Home Assistant log when restart confirmation fails during recovery verification.
• Renamed the four optional PI30 Q1 temperature channels to neutral numbered names. Clone manufacturers can wire the same protocol fields to different physical components, so the integration no longer presents an inferred tracker, inverter, battery, or transformer location as a confirmed fact.
• Fixed SMG-family setting confirmation when firmware applies a write after a short delay. The integration now records the exact-register read-back and subsequent full-poll convergence, never exposes the requested value as if it had already been observed, and includes the bounded trace in Support Archives for field diagnosis.
• Fixed stale generic-SMG projections that could survive a stronger runtime fingerprint and hide the correct model-specific telemetry surface.
• Added typed per-route probe evidence so support diagnostics can distinguish framed and raw-serial attempts without exposing endpoints or payloads.
• Kept support evidence and proxy diagnostics available when the collector is identified but the inverter driver is still unresolved, while active route-owning operations continue to fail closed.
• Fixed AT-primary collectors that accept AT management commands but carry inverter payloads in correlated EyeBond FC4 frames. The exact live session now negotiates raw serial versus FC4 from replies instead of inferring the data plane from model, endpoint, PN, cloud family, or peer address.

Docs

• Documented the unresolved intermittent cloud-telemetry issue in both READMEs, including how to distinguish it from intentional HA-only operation and what evidence to collect during an outage.
• Added safe instructions for manually testing the unreleased main branch without creating a tag or GitHub release.
• Updated device-learning and proxy-capture guides for the current menu names, trust boundaries, and recovery behavior.
• Added a Kevolt / Deye-compatible advanced-controls guide covering explicit opt-in, read-back guarantees, and deliberately excluded service operations.

Known issues

• Intermittent vendor-cloud telemetry gaps remain under investigation on some Cloud + Home Assistant setups, even while local readings continue. The fixes above do not establish that these gaps are resolved; see the README guidance and issue #13.

[0.3.0-beta.4] - 2026-08-25

Added

• Added a built-in profile for the Anenji HHS 11 kW Wi-Fi inverter family from redacted field evidence, including its model-specific entity scope.
• Added exact-session identity bootstrap for silent framed collectors so a collector can be admitted without trusting a stale or foreign connection.

Fixed

• Fixed mixed-protocol collector negotiation so the framed channel remains the primary inverter and management authority while AT is used only for compatible supplemental metadata.
• Fixed collector metadata ownership and refresh merging so an empty or failed supplemental observation cannot erase a current SSID or other live metadata.
• Hardened capability-aware reboot, recovery, driver detection, and device routing for E500-class collectors that expose both framed and AT interfaces.
• Fixed Support Archive build diagnostics so the loaded integration version is read from the actual package manifest instead of an adjacent runtime folder.
• Replaced collector identifiers in newly tracked regression fixtures with synthetic allowlisted values so release privacy validation remains effective.

Docs

• Added collector-first setup and runtime inverter-detection guides covering discovery results, background discovery, Fast versus Full protocol checks, polling, controls, UART management, and entity availability.
• Updated device-learning, proxy-capture, remote-network, and Support Archive instructions to match the current options flows and safety boundaries.
• Added a plain-language legend to the generated inverter model catalog and strengthened documentation checks for the current public workflows.

[0.3.0-beta.3] - 2026-08-24

Added

• Added a strict typed-telemetry boundary between drivers, runtime state, Home Assistant entities, derived energy, diagnostics, and support tooling. Values now retain their source and observation state instead of being passed through one broad untyped mapping.
• Added focused architecture guards and a tiered validation command for fast, affected, full-unit, real-Home-Assistant, and two-lane release checks.
• Added a fixture-backed release-readiness gate that refuses publication when required hardware evidence is missing or no longer replays cleanly.
• Added a provider-neutral Expand support for this device workflow. Users choose read-only analysis or advanced active verification first, then choose a compatible SmartESS, DESSMonitor, or ValueCloud API without silent cross-provider fallback.
• Added bounded SmartESS and DESSMonitor history evidence plus an optional background series of local read-only snapshots. Cloud labels remain hints; no local sensor or register binding is created without local proof.
• Added advanced diagnostic-command scenarios in the options flow and as a Home Assistant action, with explicit write confirmation and a short-lived, entry-scoped download for the redacted result.

Changed

• Unverified manual attempts no longer create Pending device entries. The setup flow now keeps failures inside the active flow and offers retry or background discovery; an identified session can then be added through the normal collector admission path.
• Collector admission, callback continuation, recovery verification, endpoint switching, runtime orchestration, config/options flows, transport, proxy capture, and device learning now use explicit package boundaries. This is an internal reorganization with no configuration-path migration required.
• Cloud-traffic tools are hidden for local-only ESP EyeBond Collector firmware, while its local endpoint, connection, UART, diagnostics, and Wi-Fi controls remain available.
• Device-support analysis now defaults to read-only evidence. Active verification, temporary cloud routing, and applying locally proven controls remain separate, explicitly confirmed steps.
• Proxy capture duration, start/stop, recovery, live status, and saved downloads are managed through one options-flow screen instead of separate device entities.

Fixed

• Restored Anenji point-read identity detection and made late inverter identification reload the entity set without losing the established collector session.
• Fixed callback and inbound admission around silent collectors, shared/NAT peer addresses, routed collector addresses, exact-session ownership, and cancellation cleanup. A stale or foreign connection can no longer substitute for the collector selected by the user.
• Fixed a discovery-flow completion race that could create the entry and then show Invalid flow specified in Home Assistant.
• Hardened connection-profile switching and recovery: callback-listener routes and direct collector endpoints remain distinct, endpoint formats without an explicit port stay valid for collectors that require them, and a failed transition keeps enough durable state for a safe retry.
• Fixed slow cloud-evidence reads blocking the Home Assistant event loop and tightened callback-listener cleanup during replacement, cancellation, and shutdown.
• Fixed proxy-capture startup after the coordinator package split and preserved safe, localized cloud failure reasons when device learning cannot finish.
• Fixed DESSMonitor login/provider validation, partial device reads, history pagination and time-basis handling; fixed SmartESS history collection and monotonic progress so long cloud waits no longer make the progress bar move backwards.
• Fixed cancelled inverter-detection and cloud-learning tasks so their expected cancellation cannot surface later as an unhandled task exception.
• Hardened diagnostic result downloads: only the redacted shareable copy is served by the signed API route; raw local results remain outside /config/www.
• Fixed sensor, capability, collector-endpoint, cloud metadata, and derived energy projections so they all publish from the same validated runtime snapshot.
• Fixed offline and reload startup so persisted inverter metadata creates the model-specific entity set before a disconnected live refresh can replace it; removed a self-waiting entity-migration timeout from the same lifecycle.
• Fixed PI30 serial identity handling: repeated factory placeholder values stay diagnostic-only, supported extended serials are read through QSID, and stale serial metadata is removed instead of being exposed as device identity.

Docs

• Reworked the English and Ukrainian setup guides around collector-first onboarding, one bounded scan, runtime inverter detection, the current connection/profile menus, and provider-neutral device-support analysis.
• Updated the proxy capture, remote/NAT, Support Archive, collector management, and release-validation guides; added a dedicated diagnostic-command safety guide and an automated public-link/anchor/index check.

Migration

• Existing obsolete Pending device config entries are removed automatically during integration startup. They never owned a verified collector session or endpoint, so removing them does not reconfigure the collector.
• Existing normal collector entries require no manual migration. Their saved connection strategy, endpoint, recovery proof, detected inverter, and entity registry identities are retained.
• Obsolete proxy start/stop buttons and the proxy-duration entity are removed from the entity registry. Existing or interrupted captures remain recoverable from Configure → Cloud traffic tools → Capture collector traffic.

[0.3.0-beta.2] - 2026-08-18

Added

• Added a Show discovered devices again action to the EyeBond Local — Discovery entry. It clears only temporary in-memory discovery suppression and immediately rechecks connected, unconfigured collectors without restarting Home Assistant or the collector.
• Added a persistent EyeBond Local — Discovery service entry so passive callback discovery remains available after all collector entries are removed and across Home Assistant restarts.
• Added model-specific support backed by protocol documents and sanitized hardware evidence:
  • Gootu GT-H2436M14P5 — exact G-ASCII fingerprint, telemetry, and only the charging-priority writes confirmed by the captured device.
  • Kevolt PD0080G-TPM-EU 8 kW — exact Deye-compatible high-register fingerprint and read-only three-phase, PV, battery, generator, and energy telemetry.
  • Anenji ANJ-11KW-48V-WIFI-P — documented Secondary Output Priority read-back; writes remain conditional and appear only in Full Control until hardware confirms them. Its separately confirmed Secondary Charging Priority remains available.
  • Yingfa YF6.2K-LEL-1B — a separate commercial model record and replay evidence confirming the existing PI30 MAX runtime path.
• Added collector-first onboarding: setup saves the identified collector, then runtime detection identifies the inverter and creates its entities on the owned session.
• Added one bounded network scan that combines broadcast responses, reachable local addresses, and already connected unconfigured collectors. The old quick/deep split is no longer presented to users.

Changed

• Collector connection is now presented as one product-level operating profile: Cloud + Home Assistant keeps the vendor cloud working and asks the collector to connect when needed; Home Assistant only verifies a permanent collector connection and stops normal cloud reporting.
• Switching profiles is a verified endpoint transaction. Home Assistant records the previous cloud endpoint when it can, checks that the same collector PN returns, and retains a recoverable transition state if activation is interrupted.
• Proxy capture and device learning now share one exclusive cloud-traffic operation path. New operations start from Cloud + Home Assistant, restore the original endpoint on exit, and retain explicit recovery state if restore cannot be confirmed.
• Inverter protocol selection is now a runtime concern. When more than one protocol genuinely matches, the options flow offers the proven candidates instead of persisting a speculative setup-time identity.
• The legacy writable Collector Operation Mode is retired. Its existing entity identity is retained only as a read-only operating-profile view for compatibility.
• Remote / NAT collectors are identified and owned by collector PN and exact session, never by peer IP. Multiple collectors behind one router remain distinct.

Fixed

• PI30 polling now uses its own measured 2s lower bound instead of inheriting the generic ASCII 10s floor. Manual recommendations use the same driver-owned policy and therefore reflect the observed cycle duration.
• Kept one passive-discovery flow alive while a short collector PN is enriched to its full PN, instead of replacing the flow underneath the Home Assistant UI.
• Fixed the failed inbound-verification and discovery-listener options steps so Home Assistant no longer reports Invalid flow specified / UnknownStep.
• Preserve the full runtime profile and controls for a previously detected, high-confidence inverter when a startup catalog probe temporarily times out; the saved model is resolved through one unambiguous full catalog surface and remains provisional until live detection confirms it.
• Fixed 24 V Anenji control validation so model-specific 48 V limits no longer reject valid writes before the inverter can validate them.
• Fixed duplicate placeholder inverter serials collapsing separate collector-PN devices.
• Fixed signed SMG output-power decoding and clamped negative idle noise.
• Added the confirmed PI30 MAX MCHGC / MUCHGC charge-current controls.
• Hardened silent callback identification, same-IP/NAT admission, cancellation, ownership handoff, and recovery so stale or foreign sessions cannot be saved as the selected collector.

Migration

• Existing entries are migrated automatically to the new connection settings the first time they load — no manual reconfiguration is needed.
• Your collector's server endpoint is not touched on upgrade. If it pointed at the vendor cloud, it still does; if it pointed at Home Assistant, it still does.
• Remote / NAT collectors are identified by their PN, not their IP address. Two collectors behind the same public IP stay distinct, and a collector's public IP is never used as its identity.
• If a collector's vendor app stops showing data because it was pointed at Home Assistant only, use Restore previous collector endpoint to bring it back.
• If anything looks wrong after the upgrade, open the integration's Configure menu and create a Support Archive, then attach it to a GitHub issue.

[0.3.0-beta.1] - 2026-07-06

Added

• Added a catalog-driven generic Modbus driver: supporting a new Modbus inverter family is now a declarative device pack (identity anchors + register schema JSON) instead of a Python driver. Detection uses plausibility anchors (device-type codes, rated power, value envelopes) so one pack covers a whole family, and electrical variants (24 V vs 48 V) never fork the map.
• Added four inverter families from vendor protocol documents. These are datasheet-based and not yet confirmed on hardware — telemetry should work out of the box, and we would love a support package from early adopters:
  • Sandi Aohai FSA (read-only telemetry).
  • Growatt SPF off-grid series (telemetry + controls).
  • Deye single-phase LV storage hybrids (SUN-3/3.6/5/6K-SG04LP1 class; telemetry + controls). String inverters and microinverters share the wire protocol but not the register meanings, so detection strictly gates on the storage device type.
  • Solis / Ginlong storage hybrids speaking the ESINV energy-storage protocol (read-only telemetry; the public protocol document ships without control registers).
• Added untested controls as a first-class support state: controls derived only from a vendor document and never corroborated by cloud evidence or hardware are hidden in the default control mode and appear only when the user explicitly selects Full control. Growatt SPF and Deye LV ship their control sets this way (priorities, charge-current limits, grid/generator charge enables, and similar settings).
• Added model-specific control surfaces for three EyeBond/SmartESS-collector inverters, cross-referenced from cloud evidence and vendor register maps:
  • MUST PV/PH18 now exposes a control surface built from the SmartESS cloud device-settings catalog and the vendor 1.4.15 register map — output voltage/frequency, energy-use mode, grid-protection standard, charge-source priority, battery type, PV/grid/ combined charge-current limits, discharge current, and the battery voltage windows. The 20 controls the cloud exposes ship visible in the default control mode; datasheet-only settings the cloud does not expose (grid charging, the battery-equalization block) stay in Full control. Every control reads back its current value. Hardware write confirmation is still pending a tester report, and the inverter validates its own writes (a locked or absent register self-disables).
  • Aninerel ANL-4200T-24L-W-PRO binds a model-specific full-control surface instead of the read-only family fallback; a tester confirmed writes on hardware, so the family-proven SMG settings are visible in the default control mode. Battery-voltage windows are left wide because this is a 24 V unit and the family templates carry 48 V windows, so the inverter's own validation is the authority.
  • SMG family 4200 variant (a new SMG model) ships controls graduated from a full shadow-learning run that correlated the SmartESS cloud's writes to real registers (input/buzzer/LCD/boot mode, output voltage/frequency, and mains/off-grid battery low-voltage protection).
• Added a raw AT wire-evidence probe to support archives for at_text collectors: when detection fails on a raw-serial ASCII collector, the archive now records a bounded PI30 and G-ASCII read sweep with per-command request/response bytes, and the collector diagnostics surface the raw request/response counters from the AT connection.
• Added a transactional inverter-link baud sweep to the deep scan for ESP EyeBond Collector bridges: when the inverter stays silent on the configured UART speed, the scan walks the candidate speeds a protocol family declares, probing only the drivers expected at each speed, and always restores the original speed. This finds e.g. SRNE/MUST-class units on 19200 behind a bridge provisioned at 9600 (ESP32/ESP8266 bridges only; bk72xx UARTs are fixed at runtime).
• Added Modbus input-register (FC 0x04) support to the shared Modbus core, plus new decode features packs needed: per-spec raw offset (Deye-style (raw-1000)*0.1 temperatures), low-word-first 32-bit counters, and divisor-implied precision.

Changed

• Detection internals were consolidated around single owners: one anchor-matching semantic, one resolution engine (the compiled decision tree), one decision-DAG walker shared by all catalog probers, and one register decoder shared by SMG, SmartESS, and the generic packs. Along the way a real mis-detection was fixed: integral-float signature values (220.0 vs 220) could push a known LVYUAN unit into the read-only family fallback.
• The deep-scan silence verdict is now grounded in transport-observed responses instead of driver-reported outcomes, so "no supported device" and "device never answered" are distinguished honestly.
• Switching the control mode now reloads the config entry, so control entities that only materialize at platform setup (the untested/Full-control set) appear and disappear immediately instead of after a manual restart.
• Support packages are now uniformly share-safe: every archive member masks long numeric identifiers (collector PN, serial numbers) with the same rule, including identifiers embedded in hex payload dumps; replay fixtures keep an anonymized twin, and the manifest always references the real archive file.

Fixed

• MUST PV/PH18 battery power vs current: register 25274 is the battery current (A), which a third-party map mislabeled "Battery_Load"; it is now exposed as Battery Current, and the vendor's actual "Batt power" register 25273 is exposed as Battery Power (W).
• MUST PV/PH18 control read-back: the settings blocks were widened to gap-free register ranges so the inverter no longer rejects them, and every control decodes its current value (previously the wide blocks spanned absent registers and left the controls blank).
• Pack control entities are registered enabled by default, so a device that exposes its pack controls no longer needs each entity enabled by hand; the enabled-defaults self-heal re-enables entities that a previous version had left disabled.
• ESP EyeBond Collector bridges: the Home Assistant callback endpoint written into the collector is now always the entry's own listener port (a proxy-template port could previously leak into the collector and strand it), and bridge entries always resolve to the framed transport profile — both at onboarding and through runtime reconciliation — fixing scans that silently probed the wrong protocol after SmartESS-style metadata answers. Endpoint changes pair best with esp-eybond-collector v0.1.8, which defers the endpoint apply until the acknowledgement is flushed.
• Collector transport lifecycle: a replacing collector session can no longer tear down its successor (previously a live collector could "vanish" until it redialed); teardown no longer inherits a dead peer's TCP timeout or a swallowed task cancellation (both could hang Home Assistant shutdown); unidentified callback sockets are parked under a watcher instead of lingering unwatched where a dead socket blocked same-IP routing; and the byte-shape sniff defers to registered per-PN session owners.
• Phantom daily grid import on SMG-class hybrids: the power-flow split now trusts the charger's own PV measurement over headroom derived from an under-reading pv_power register. A PV-only system no longer accrues fake grid-import energy.
• Onboarding fixes from a deep review pass: the scan aggregator no longer cancels an extendable deadline based on a stale snapshot, probe logs survive scan errors, a slow first refresh no longer blocks entry setup, restricted re-scans no longer spend probes on excluded drivers, and explicit decimals: 0 is honored in register decoding.

[0.2.0] - 2026-07-02

Added

• Added first-class support for the community ESP EyeBond Collector virtual bridge (firmware for inverters without a factory collector). The bridge is detected from the collector hardware-version token read through FC=2 parameter 6 (collector_hardware_version = esp-collector/<version>/<platform...>); detection never uses the collector PN, devcode, or factory firmware-version string, and no extra AT probe is sent during detection. A detected bridge is shown as an honest ESP EyeBond Collector device (manufacturer, model, firmware version, and project link) and its cloud-only actions — device learning / shadow learning, proxy capture, and cloud-assist actions — are hidden unless a future firmware explicitly advertises cloud capabilities. Its collector operation mode is fixed to Home Assistant only (the Cloud + HA choice is not shown), and reverse discovery stays on so it can reconnect. Local actions (runtime settings, diagnostics, and Change Collector Wi-Fi) stay fully available. Already-onboarded bridges keep working from their persisted bridge identity; fresh bridge onboarding requires esp-eybond-collector firmware that emits the esp-collector/<version>/<platform...> token.
• Added an adaptive sensor refresh scheduler. New entries default to Automatic refresh, where EyeBond Local chooses the next poll interval from the observed device response time and protocol-specific limits; Manual refresh keeps the existing fixed-interval behavior for upgraded entries and users who want a fixed cadence.
• Added collector diagnostics for poll mode, current/next interval, poll duration, utilization, recommended interval, and scheduler delay.
• Added poll-context diagnostics so collector-only, detection, and runtime polling cycles are visible separately.
• Added an offline device identification catalog: inverters are now identified by a deterministic register fingerprint (protocol layout + model code + rated power) with explicit support tiers, and the catalog — not heuristics — decides which schema and controls apply.
• Added a detection summary step to onboarding: after the scan you see the identified model, its support tier (full / partial / not recognized), and what to do next, before the device is created.
• Added canonical power-flow telemetry for the SRNE and MUST PV PH18 drivers: pv_power (SRNE: PV1+PV2 sum, MUST: solar-charger power), signed battery_power (SRNE register 270, MUST register 25274), and the six *_to_*_power flow-split sensors now populate from registers these drivers already poll, so the power-flow card renders fully for both families. SRNE has no grid-power register and MUST has no battery-SOC register, so those two slots stay empty until the registers are identified.
• Added a guided control-discovery wizard ("Add controls (device learning)") for partially supported and unrecognized inverters: one linear flow (consent → vendor-app sign-in → progress → review → apply) replaces the old technical action menu. Sessions are fail-closed: the collector is always restored even when a run fails.
• Added read-sensor learning: during a learning session the integration also captures which registers the cloud reads, binds cloud sensor labels to registers by value correlation and enum matching, and can apply the learned read sensors even when no controls were selected.
• Added new supported models from community-donated captures: Anenji 6200 (dual output) — full support including second-output telemetry (power, apparent power, load, voltage, cut-off SOC, overload threshold) and an Output 2 on/off switch — and Anenji 6200 (single output, full support with the SMG control set).
• Added more supported models: Anenji ANJ-4000W-24V (full SMG telemetry), and read-only telemetry support for MUST PV18, SRNE-compatible Modbus inverters, and LVYUAN units on the new SmartValue / EyeBond G-ASCII protocol. The full, always-current device matrix is the generated inverter model catalog.
• Added collector callback identity routing: collector callbacks are routed by collector identity (PN) instead of peer IP, so several collectors behind one router/NAT are tracked separately, and a collector callback-identity diagnostic sensor surfaces the routing state.
• Added bit-level write capabilities: controls that own a single bit of a shared register are written read-modify-write, preserving the other bits (this enables the Output 2 switch).
• Added a memory guard to learning: on memory-tight hosts the scan refuses to start (insufficient_memory) instead of risking an out-of-memory crash.
• Added a contribution toolchain for donated captures: contribution record builder, vetting tool, and a GitHub issue template for sharing support packages.
• Added a validation toggle (EYBOND_FORCE_UNSUPPORTED=1 env var, or a force_unsupported.flag sentinel file in the HA config data dir) that treats every model as unsupported so the learning flow can be exercised on a fully supported device.
• Added a developer-directed diagnostic command runner — the eybond_local.run_diagnostic_commands action and an "Run diagnostic commands" options screen (shown only with Home Assistant Advanced Mode) — that runs a small read/write/write_bit/ ascii scenario against the inverter over the existing collector connection, for adding or debugging device support. It never changes config-entry settings, runs one scenario per entry at a time, and requires an explicit confirm_write before any scenario that writes to the device. Results are saved locally; the shareable copy has known identifiers redacted.

Changed

• Deep scan is now a real extended detection path instead of a lightly longer quick scan: it keeps structured evidence for each target, does not stop a target batch after the first matched inverter, uses a distinct extended timeout budget, and surfaces detection timeouts separately from ordinary collector-only results.
• The deep-scan time budget is now derived from the registered drivers' own signature and probe budgets instead of a hand-maintained constant that was smaller than the worst-case driver sweep, and each connected target gets its own sweep deadline so one slow target cannot starve the identification of another. Together these remove the main cause of connected collectors finishing as "detection ran out of time".
• Driver detection now uses the protocol metadata a collector reports (mapped through the protocol catalog, not a hardcoded family list) to probe the matching local driver first. Previously the probe order fell back to the registry order when the signature pre-pass did not match, so a PI30-family inverter could burn most of the deep-scan budget on Modbus drivers before its own driver was ever tried.
• Deep detection now records a per-driver probe log (driver, elapsed time, outcome) into the detection evidence and the created entry (detection_probe_log), so real installations produce the data needed to validate and tune the per-driver probe budgets instead of guessing.
• The driver-choice step was made human-readable: options are now "model — driver (recommended)" with the measured identification time per candidate, the raw probe-route digits are gone (the device address is shown only when two candidates differ by nothing else), and the summary explains that protocols can differ in polling speed. Candidates show the localized driver display names ("SmartESS 0925 / Modbus", "SMG / Modbus", "PI30") instead of internal keys, and the missing English fallback label for the SmartESS-local driver in the driver selector was added.
• After a deep scan the repeat action is labeled "Repeat deep scan" (repeating always re-ran the same scan mode, but the generic label made it look like a quick scan and sent users through the advanced menu to reach deep scan again).
• Scan targets whose probe was deliberately cancelled because another target already matched are no longer reported as detection timeouts: they carry the dedicated cancelled_first_match_found evidence status, keep what was learned about the candidate, and present by their actual state (collector replied / connected) instead of "detection ran out of time".
• The scan-results screen lets you add a found device directly: the device list and the follow-up actions (refresh, advanced setup) live in one selector, replacing the extra "Add detected device" menu hop and its separate selection step.
• Deep scan now starts immediately when the selected interface has a known, normally sized network; the intermediate confirmation step remains only where the user has a real decision to make (a large subnet, or an unknown network).
• The deep-scan identification headroom now sits on top of the discovery budget instead of sharing one ceiling with it, so scanning a /16 network (whose discovery alone takes ~14 minutes) no longer consumes the identification budget.
• Collectors owned by existing entries are now visible in the scan results as "Already added" (previously the unprobed marker was filtered out of the list, so the summary claimed zero configured devices).
• The scan-results list was decluttered: the status chip is not repeated in the details (no more "SmartESS hint — ... SmartESS metadata ... collector connected"), the "Unconfirmed inverter" filler is gone, serial numbers are shown only when known, and confidence is lowercased mid-line.
• The deep-scan time budget now follows the discovered work instead of being fixed upfront: every collector that connects and is admitted for identification extends the shared scan deadline by one full driver-sweep budget (bounded by a 15-minute runaway ceiling). A site with many inverters gets one sweep per collector instead of all of them starving on a budget sized for one or two.
• Scans no longer probe collectors that already belong to a configured entry: probing stole the collector's callback session from the running entry and burned the shared scan time budget on devices that cannot be added again anyway. Such collectors are listed as "Already added" without being touched.
• A collector that dials back in now triggers an immediate refresh when the entry is not bound, instead of idling until the next scheduled poll — after failed detection cycles that could be more than a minute away. Together with the next change this removes most of the "one update, then unavailable for a minute" churn right after adding a device.
• Consecutive failed refresh cycles no longer re-run the full (slow) collector AT metadata sweep every time: the caches are invalidated once per outage, and collector liveness is proven with the cheap framed query when available. This also stops the failed-cycle duration — and therefore the retry backoff derived from it — from being inflated by our own metadata reads.
• Deep scan now keeps every successful local driver/protocol probe for the same device. When one inverter responds through multiple protocols, onboarding shows a driver choice step with the successful candidates instead of silently keeping only the first match. Choosing an alternative keeps the full probe metadata (SmartESS details included) and re-runs the confirm-time detail refresh with the chosen driver, and if the deep-scan time budget runs out mid-probe the candidates found so far are kept instead of being discarded with a bare timeout.
• New entries persist detection evidence (detection_depth, detection_status, detected_probe_route, and the candidate driver list when more than one protocol matched) so the support package can explain how a device was onboarded.
• Onboarding UX overhaul: the welcome form is gone, the happy path is decluttered, cloud assist is an explicit optional choice (never an interstitial), developer tooling is removed from the options flow, and long wall-of-text screens were rewritten into short actionable guidance (en/ru/uk).
• Unknown-but-SMG-family inverters now onboard at the partial tier with base read sensors and a clear pointer to device learning, instead of being rejected with "no supported driver matched".
• Inverter-communication loss during detection now surfaces as "inverter link down / retry" instead of a false "no supported driver matched".
• Startup is lighter: metadata catalogs are warmed off the event loop, removing blocking file reads during driver detection (matters on slow or throttled hosts).
• The runtime settings flow now hides the fixed poll interval while Automatic refresh is selected and shows it only for Manual refresh.
• High-utilization warnings now point Manual-mode users either to a larger interval or to Automatic refresh.
• Automatic refresh no longer learns from collector-only, offline, or inverter detection cycles, so a missing or unsupported inverter does not permanently inflate the normal runtime poll interval.

Fixed

• The poll-debugging diagnostic sensors (Poll Phase Breakdown, Refresh Phase Breakdown, Slowest Driver Requests) are disabled by default — enable them per collector when chasing a slow poll — and all of them live on the collector device (the refresh breakdown and slow-requests sensors previously landed on the inverter device).
• Added "Refresh Phase Breakdown" and "Slowest Driver Requests" diagnostic sensors: the runtime refresh reports where its time went (collector metadata, driver detection, driver read, snapshot build) and the SmartESS-local driver reports its five slowest register requests per cycle with outcomes.
• Added a "Poll Phase Breakdown" diagnostic sensor: each poll cycle reports where its wall-clock time actually went (network reconcile, session profile, runtime refresh, snapshot profile, endpoint reconcile), so a slow cycle is explained by a sensor read instead of a packet capture.
• The Collector Protocol Asset ID and Collector Devcode diagnostics no longer flip between two values every cycle. Parameter 14 on some collectors returns a composite serial-protocol config string ("02FF,0,0,#0#") — the id is now parsed from its first field, and an asset id is claimed from parameter 14 only when the protocol catalog knows it, so it cannot fight the asset id the bound driver reports. The devcode diagnostic now shows the collector's stable heartbeat devcode instead of the devcode of whatever frame happened to arrive last.
• Whether a collector answers the AT metadata channel at all is now a learned per-device fact, like unsupported inverter commands: framed collectors tunnel AT via raw passthrough and only some firmwares support it, so after two evidence-gated failures the channel is skipped entirely (persisted as collector:at_metadata, cleared by the same "Re-check Supported Commands" button). Collectors where AT-over-passthrough works (Wi-Fi scan, signal metadata) keep it.
• Fixed the actual cause of the stable ~60-second poll cycles on EyeBond collectors with a framed callback session: the collector AT-metadata sweep (12 commands: DTUPN, ATVER, WFSS, ...) ran against an AT channel that never answers, burning a full request timeout per command, and the empty result was never cached — so the sweep repeated every cycle. The sweep now aborts after the first timeout (a dead AT link times out for every command), and its retry cadence is keyed on attempts instead of successful results. The refresh-phase diagnostic additionally splits collector metadata into its framed (FC) and AT parts.
• The SmartESS-local (Modbus) driver applies the same unsupported-command memory to its register blocks: a bulk read the inverter rejects or ignores was retried every cycle (block failures never marked the block as read), spending most of a ~60-second poll on consecutive request timeouts while the wire carried only a few seconds of real traffic. Rejected blocks and dead fallback registers are now remembered per device (block:config, register:5005, ...), the capability values keep coming from the cheap single-register fallbacks, and the same "Re-check Supported Commands" button clears the memory.
• The ASCII drivers (PI30, PI18, EyeBond G-ASCII) no longer re-send unsupported commands on every poll cycle. On inverters that only answer the core command set, every optional or energy command burned a full request timeout each cycle, turning a ~2-second poll into a ~60-second poll. A command that fails twice — in cycles where the device answered something else, so a link outage never counts as evidence — is marked unsupported for this device permanently: the learned set is persisted into the config entry (driver_unsupported_commands), survives restarts and reconnects, and is shown as a diagnostic value. A new "Re-check Supported Commands" diagnostic button clears the learned set and probes everything again (for example after an inverter firmware update).
• One collector seen through two scan sources (the configured-IP marker plus a PN-carrying callback-session result) is now collapsed into a single scan line; the PN-carrying duplicate wins so the line keeps the identity.
• Collector callbacks that no config entry owns are now parked (held open passively, bounded and with a TTL) instead of being closed after classification. Closing made the collector firmware redial within seconds, producing a permanent connect/close loop for collectors whose entry was removed; a parked callback also stays instantly claimable by a scan or a newly added entry, including its already-received identity bytes.
• A collector whose management link answers but whose inverter sends no heartbeat is no longer reported as offline. The runtime now separates the two layers: collector-level sensors keep updating, the state becomes driver_unbound with the new inverter_heartbeat_missing code (instead of the misleading collector_heartbeat_timeout), and a bound inverter still goes through the normal reconnect/recovery path. This fixes the esp-collector bridge appearing stuck as "waiting for collector" after a power cycle even though its TCP session was live.
• Stale UDP discovery details (collector_udp_reply, collector_udp_reply_from) are now dropped when the collector goes offline instead of being shown from the previous session.
• The manual-mode high-utilization warning no longer misfires after a device outage. Cycles that bound the driver or recovered the collector connection measure that recovery work, not the normal poll cost, so they are now excluded from the warning streak, from the poll-duration statistics (average/max/recent sensors), and from the adaptive scheduler's learning samples — a device coming back online no longer inflates the reported utilization or the recommended minimum interval. The warning is also dismissed automatically once polling stays within the configured interval again.
• Device learning now actually works on the partial / unrecognized tier it targets. Two coupled defects made it unusable there: those devices silently inherited the full controls profile (so overlay generation deduped against the wrong base), and the readiness check kept reporting missing_effective_metadata_snapshot because that tier never persists a snapshot.
• Changing the collector Wi-Fi no longer fails with "collector_listener_bind_failed … address in use" while the entry is loaded.
• Guided control discovery is more robust: closing the dialog mid-scan now still restores the collector route (fail-closed), a single bad cloud value can no longer abort a successful run, a successful-but-empty run is no longer shown as a failure, and re-running the wizard no longer shows the previous run's results.
• The Output 2 / bitmask switches no longer get permanently blocked when a transient read error happens while toggling them.
• Partially supported devices: the add flow now tells you the next step to unlock controls, and the learning consent dialog is correctly translated (ru/uk).
• Removed a harmless but noisy "Unable to remove unknown job listener" error on unload.
• Catalog entries are validated on load (support tier and its controls-profile invariant), so a malformed entry fails fast instead of silently degrading onboarding.
• Privacy: all real device identifiers (collector PNs, serials, credentials, network details) were replaced with synthetic stand-ins across tests and catalog provenance, and a guard test now fails the suite if a real-looking identifier enters the source tree.

Performance

• Lighter coordinator refresh on resource-constrained hosts: metadata base-name resolution is cached instead of reading overlay files from disk every poll, the shadow-learning session state is no longer re-read from disk every refresh, and redundant per-poll Modbus reads on multi-output variants were removed.

Docs

• Updated the README hardware table, device-learning guide, and setup walkthrough (en/uk) for the catalog-based identification, the guided learning wizard, and the new Anenji 6200 models.

[0.2.0-beta.1] - 2026-05-12

Beta Notice

• This is a prerelease for the 0.2.0 line; install it only when testing the new collector-first onboarding, local callback ownership, and collector operation mode flows.
• Keep a Support Archive before reporting collector matching, callback ownership, proxy capture, or inverter detection issues.

Added

• Added a collector-first onboarding flow with collector network choice, optional Bluetooth Wi-Fi provisioning, scan-interface selection, and clearer quick-scan / deep-scan / manual-setup branching.
• Added separate collector device management with collector operation mode, Wi-Fi change, restart, proxy capture, and collector-scoped diagnostics.

Changed

• EyeBond Local now presents collector and inverter devices as separate parts of one installation, with user-facing runtime settings centered on collector mode and control mode.
• User-facing setup and runtime copy now explain the two everyday collector modes, Cloud + HA and HA only, in plain user terms.
• Support artifacts now split collector, inverter, and integration roles so beta reports are easier to triage.

Fixed

• Switching between Auto and Full Control now reloads the entity set correctly instead of leaving stale expert entities behind.
• Collector callback endpoint override is available again in Full Control, and hidden full-control entities are cleaned up correctly when returning to safer modes.
• Legacy EyeBond cloud endpoints now resolve with the legacy port default when only the host name is known.
• Legacy collector signal entities are cleaned up from existing installs instead of remaining visible after their defaults changed.
• Removed collector entries and Home Assistant Core restarts now close owner-specific callback sockets and avoid accepting orphan collector callbacks.
• Local onboarding no longer offers cloud assist before creating a local entry, and HA-only mode applies its callback binding silently.

Docs

• Updated the English README around the new onboarding flow and collector modes, added a public collector-management guide, refreshed the docs index, and replaced outdated setup screenshots.

[0.1.53] - 2026-04-25

Added

• Collector diagnostics now track connection churn, dropped pending requests, disconnect reasons, and discovery restarts to make local write-path contention easier to diagnose.

Changed

• Declarative SMG runtime gates now surface advisory warnings instead of locally hiding or hard-blocking controls; the inverter response and immediate readback confirmation are now the final authority for non-action writes.

Fixed

• Non-action writes no longer report silent success when the refreshed value stays unchanged; EyeBond Local now raises an explicit write_not_confirmed error after readback.
• Temperature sensors backed by the affected PI18 and Anenji schemas now report °C instead of plain C.

Docs

• Updated the English and Ukrainian READMEs plus the SMG support matrix to describe advisory runtime warnings, explicit readback confirmation, and the new collector contention diagnostics.

[0.1.52] - 2026-04-23

Added

• Added an explicit built-in anenji_4200_protocol_1 SMG runtime path for classic protocol-1 hardware that matches the documented device_type=0x3501, protocol_number=1, and rated_power=4200 anchors.
• Added documented power_flow_status decoding for classic SMG protocol-1 layouts, exposing diagnostic connection, battery, load, and charge-source states instead of leaving the raw register uninterpreted.
• Added documented classic protocol-1 fault/log support-capture coverage for 700..744, so support archives retain a broader evidence window for SMG 6200 and the document-backed Anenji 4200 path.

Changed

• SMG protocol-1 writable metadata is now layered through a real shared base plus model overlays, so the common path carries only the shared protocol-1 controls and presets while 6200-only extras stay model-scoped.
• The support workflow and support-bundle markers now treat explicit but still-unverified model-specific profiles separately from the read-only SMG family fallback, keeping the new Anenji 4200 path at partial support instead of labeling it as a generic fallback.

Fixed

• SMG common protocol-1 schemas and probe metadata no longer leak 6200-only diagnostics like 341..343 and max_discharge_current_protection into other protocol-1 variants.
• The Anenji ANJ-11KW-48V-WIFI-P schema no longer depends on accidentally inherited low-DC measurement metadata after the SMG common/base cleanup.
• The project-wide support overview export now includes the new built-in Anenji 4200 protocol-1 runtime profile instead of silently omitting it from release documentation.
• Runtime control-mode labels in the options flow are now localized instead of falling back to hard-coded English labels.
• Re-adding the verified default SMG 6200 integration now restores the tested write controls that should be enabled by default, including the equalization settings and the previously verified battery threshold controls.
• Sync Inverter Clock no longer leaks into the verified default SMG 6200 runtime path; it now stays scoped to the Anenji ANJ-11KW-48V-WIFI-P model-specific tooling path.
• Pending onboarding no longer stalls when a saved manual or pending device keeps the default broadcast discovery target as collector_ip; the shared listener now aliases that placeholder to the real collector callback IP.

Docs

• Updated the English and Ukrainian READMEs, the SMG support matrix, and the generated support overview to describe the new Anenji 4200 protocol-1 path, the stricter SMG common-vs-model layering, the verified default SMG runtime path wording, and the current protocol-1 read coverage more accurately.
• Expanded the English and Ukrainian READMEs and onboarding copy to explain what the Pending Device / EyeBond Setup Pending state means, what to expect from it, and which retry steps to use before opening a support issue.

[0.1.51] - 2026-04-18

Fixed

• Anenji inverter date/time sensors no longer stay unavailable on hardware that returns valid clock registers only when the optional inverter clock range is read as one contiguous block.

[0.1.50] - 2026-04-18

Added

• Added a separate deep-scan onboarding path that can probe the full selected IPv4 network from both the first setup step and the scan-results screen.
• Added BusyBox-compatible IPv4 interface parsing for Home Assistant OS, so deep-scan network size and broadcast metadata still resolve correctly when ip -j is unavailable.
• Added runtime-schema-aware entity selection for model-specific SMG variants, which restores Anenji PV1/PV2 and other variant-only entities when the detected runtime metadata differs from the generic driver defaults.

Changed

• The Anenji ANJ-11KW-48V-WIFI-P model-specific write surface is now marked as tested on real hardware, so its validated controls can participate in normal high-confidence auto exposure.
• The setup wizard now distinguishes quick scan from deep scan explicitly, with scan-mode-aware hints, timing estimates, and follow-up actions.

Fixed

• Quick scan now stays effectively broadcast-first by removing duplicate broadcast targets and shortening reverse-connection waits when no UDP reply was received.
• Deep scan no longer reports zero-address networks on BusyBox-based Home Assistant OS hosts and remains available from the results screen even when candidates were already found.
• The scan progress bar now publishes its first determinate update immediately instead of briefly jumping from an indeterminate-looking state.

Docs

• Updated the English and Ukrainian READMEs, SMG support docs, and generated support overview to describe deep scan, the validated Anenji control surface, and the current onboarding fallback flow more accurately.

[0.1.49] - 2026-04-17

Added

• Added a dedicated SMG family_fallback runtime path with explicit read-only/unverified markers in the support workflow, support bundle, and exported support archive.
• Added broader built-in Anenji ANJ-11KW-48V-WIFI-P monitoring, including PV1/PV2 telemetry, inverter date/time readback, native PV day/total counters, and a Sync Inverter Clock tooling button.
• Added broader SMG read-side diagnostics for the verified default path, including program_version, protocol_number, device_type, battery_type, warning_mask_i, dry_contact_mode, and automatic_mains_output_enabled, with cautious hidden-by-default exposure for lower-value raw fields.

Changed

• SMG writable metadata is now layered through shared family/base/default/model profiles with capability templates instead of one duplicated monolithic profile file.
• The Anenji profile is now shipped as a real 47-capability model-specific control surface, but those writes remain intentionally untested and stay out of normal auto exposure.
• Runtime metadata and support reporting now label internal runtime paths separately from commercial hardware names, so docs and exported reports are less misleading.
• The daily grid-export helper now stays available when signed or direct export power keys are present, even if solar_feed_to_grid_enabled is missing.

Fixed

• The default SMG binding now stays limited to verified 6200-class hardware; other SMG-like power classes fall back to the read-only family path instead of inheriting the default write surface.
• Optional SMG probe diagnostics now backfill missing details during normal refresh, so the surviving probe-only sensors remain available after Home Assistant restarts.
• Placeholder all-zero SMG device_name values are now suppressed instead of surfacing as misleading identifiers on the verified SMG 6200 path.
• Local draft and SmartESS bridge generation now copy fully resolved profile JSON, so profile shims and layered metadata do not leak into generated local files.

Docs

• Updated the README, Ukrainian README, SMG support docs, and generated runtime-profile reports to describe the verified default SMG path, the Anenji-specific path, and the read-only SMG family fallback more explicitly.
• Release docs and CLI examples now use the changelog-first flow with version placeholders instead of stale hard-coded historical tags.
• Removed the extra README card badge while keeping the companion card link in place.

[0.1.48] - 2026-04-17

Added

• Added optional cloud assist for onboarding and diagnostics, including reusable cloud-evidence export for one collector identity.
• Added JSON-first SmartESS protocol and model-binding catalogs plus imported SmartESS assets 0912, 0921, and 0925 for metadata ownership, diagnostics, and local draft tooling.
• Added SmartESS local collector helpers for collector query/set commands, protocol-id parsing from query 14, and known-family metadata planning.
• Added read-only SMG model coverage and wider support-archive capture windows for Anenji ANJ-11KW-48V-WIFI-P / Protocol 3-10 devices.

Changed

• Create support archive is now the main diagnostics flow and can include saved cloud evidence automatically or refresh it inline before the ZIP is built.
• Runtime diagnostics, support export, and local draft tooling now resolve effective profile/register-schema ownership from saved or live SmartESS metadata hints, so imported SmartESS assets can be used before a native SmartESS runtime driver exists.
• PI30 default metadata now uses the canonical SmartESS 0925 compatibility paths, while user-facing naming presents raw VMII-NXPW5KW devices as PowMr 4.2kW.
• Advanced metadata tools now focus on raw JSON export plus SmartESS draft and bridge generation instead of duplicating routine archive and reload actions.

Fixed

• Metadata cache priming now also warms catalog-driven metadata, avoiding blocking file reads when Home Assistant starts or reloads local overrides.
• Support archives now store matching cloud evidence only once inside the ZIP under evidence/cloud_evidence.json.
• Support-archive raw register capture now follows the effective schema name, so model-specific SMG evidence windows are not dropped for variant overlays.
• External relative metadata overrides can now fall back to built-in parent profile and schema files when the local parent file is missing.

Docs

• Public docs now explain cloud evidence, inline archive refresh, and the retention behavior of saved cloud-evidence files.
• Public docs now call out PowMr 4.2kW and Sandisolar SD-HYM-4862HWP as the currently verified commercial examples for the PI30 and SMG families.

[0.1.43] - 2026-04-15

Added

• First public GitHub release of EyeBond Local.
• Built-in local support for SMG / Modbus and PI30-family collectors, plus experimental PI18 replay coverage.
• Support Archive export workflow for unsupported or partially supported inverters.
• SMG writable and readback coverage for registers 341, 342, and 343.

Changed

• Config-flow runtime copy now loads from private flow_translations/ bundles while Home Assistant-validated translation files remain Hassfest-compatible.
• Public release validation now passes HACS Validation, Hassfest, and the repository quality gate in GitHub Actions.

Fixed

• Publication blockers around config schema exposure, manifest ordering, and unsupported translation key placement.

Docs

• Public README and release metadata were aligned for the first published release.