Field data by Kris1166 (issue #12). A DPT-5.001 shutter position feedback named only '… RM'
was classified lighting/light and emitted as a bogus HA light entity. Two layers:
- classification (project.py): an ETS Function carrying a shutter role promotes a member to
shutter — but only on POSITIVE evidence (member is a step/stop 1.007/1.010 OR a 5.001 position
STATUS, the Function has an up/down move sibling, and the name has no explicit non-shutter
domain), so a scene-recall/dimmer-feedback/lock sharing a mixed Function is left alone
(LLM-council review; a Function is a free grouping, not proof of domain).
- status recognition (project.py name_is_status): a standalone 'RM'/'FB' token is now a status —
the 'rm '/'rm_' substrings and the stopword tokenizer both missed a name ending in bare 'RM'.
- analyze._is_status_ga delegates to name_is_status; generate_ha then wires it as the cover's
position_state_address.
Passed both gates (self-audit + LLM-council). tests/test_rm_status.py pins it (red/green + negatives:
a light's RM stays light, a stray light / a scene-recall in a shutter Function are not promoted).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Post-fix hardening from an LLM-council review of 0198ae3, with a gate-1 audit pass:
- _is_shutter_control rejects a bare 1.007/1.010 whose tokens are a foreign domain
(light/dim/socket/HVAC) or a central macro (Alle/Zentral) — a 'stop' is an operation
signal, not a domain one, so a lighting 'Licht Stopp' or a house-wide 'Alle Stopp' can
no longer be mis-paired/stolen as a cover's step/stop.
- _rank gains a stable address tiebreak -> deterministic pick across parses on a full tie.
- gate-1 caught 'master' wrongly excluding Master-zone covers; removed. Vent words kept
admissible (roof-window/flap openers are covers).
- tests groups E/F/G pin the foreign-stop, central-steal, and Master-zone cases.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Second pass on issue #11, grounded in Kris1166's real ETS-6 field dump (groups A/B/C),
after the first fix (dd8ecac) proved partial. Three levers:
- project.py: ETS Function role (MoveUpDown/StopStepUpDown) authoritatively promotes a
GA to category=shutter — rescues bare DPT 1.007 step/stops the name classifier left
'unknown' (group A: 26/33 covers were dropped). Guarded to DPT 1.x/5.x; role strings
from Kris's real dump, not invented.
- generate_ha.py: cover builder now gathers candidates and picks the BEST per role
(same-function > name-token overlap > nearest sub) instead of first-match, so a
function-less roller takes its own step/stop, not an awning's sharing the zone token
(group C). Bare 1.007/1.010 admitted only with a shutter name signal, so a foreign
lighting step can't be stolen (audit finding).
- tests: test_cover_pairing.py rebuilt from the real A/B/C structures + a group-D guard
for the foreign-1.007 case. Fails without the fix, passes with it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
generate_ha_package cross-wired move_short_address across unrelated ETS
Functions sharing a zone token (a window and an awning both 'Room A'), and
could leave a cover with no step/stop. Pair siblings by ETS Function
membership (authoritative); a sibling owned by a different function is never
borrowed. Falls back to the existing name-token logic only when the move GA
has no function, so projects without ETS Functions are unchanged.
Adds tests/test_cover_pairing.py (reproduces the field case from #11, fails
without the fix) and wires it into CI.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
detect_topology_issues now surfaces in the human-review project_report markdown
(new '2.4 Topology & addressing' section) and its severities count into the
report totals — closes the audit's deferred LOW (per CLAUDE.md the human reviews
project_report before ETS import, so topology issues must be visible there).
Functional check: Poltavskiy report renders the coupler INFO; suite 27 passed
(pre-existing test_protocol collection error untouched).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New design-time check grounded in the KNX standard (KNX Handbook + xknx via
deepwiki). detect_topology_issues() + check_topology() MCP tool (30->31 tools),
folded into analyze_all counters:
- devices per line vs TP1 segment 64 / line 256 (KNX Handbook p.36/40/55) — TP-only
- individual address valid A.L.D (area 0-15, line 0-15, device 0-255; xknx-confirmed)
- duplicate individual address -> error; missing .0 coupler on a multi-line TP -> info
Medium-gated on the real xknxproject string ('Twisted Pair (TP)'), so IP/PL/RF
lines are never falsely flagged (fixed a dead 'medium==IP' guard caught in audit;
regression test added). senior-code-audit: APPROVED. Functional smoke on 2 real
projects: IP lines clean, TP coupler/count notes correct. 64/256 cited to the
Handbook, not attributed to xknx. tests/test_topology.py 7/7; suite 27 passed
(pre-existing unrelated test_protocol collection error untouched).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the 5 tools missing from the RU table (present in EN + server.py):
- explain_ga → Чтение
- check_policy → Валидация
- check_device_parameters → Починка и дизайн
- validate_room_template, compose_rooms → new Room Library group
Descriptions mirror EN meaning, signatures verified against server.py.
Bump counters: '## 5. Инструменты MCP (25)' → (30) and package layout
'25 инструментов' → 30. RU table now 30 rows, 1:1 with EN.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Compose a new KNX project from parametrised room templates. Adds two MCP
tools (validate_room_template, compose_rooms), six built-in room templates
with a public SCHEMA.md contract, a resolved-IR allocator, and a round-trip
.knxproj generator re-read by the standard loader. Generated house passes
all four linters with 0 errors / 0 warnings. 14/14 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Document both new tools in README (tool count 28 -> 30), add a CHANGELOG entry
under Unreleased, and ship the room_templates/*.yaml + SCHEMA.md as package data.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
tests/test_room_library.py covers all R1-required cases: exact golden RU/EN
object maps for one template; idempotency; room permutation invariance;
sub-address exhaustion raising a clear error; round-trip (generated house ->
load_project -> 0 errors AND 0 warnings across validate_naming /
detect_missing_status / detect_dpt_issues / check_policy); a hand-written
negative-oracle manifest vs fact; plus per-slot preset override, template
rejection and BOM proposal checks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two new tools (28 -> 30): validate_room_template checks a built-in or custom
template against the R1 schema; compose_rooms builds a new project from a room
list and returns the allocation manifest, ETS GA XML/CSV (via the existing
generators on the re-read project), and a device BOM proposal. New projects
only, dry-run by default; writes artifacts into the confined workspace when
dry_run=false and output_dir is given. Never touches a bus.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the Room Template Library core: the public YAML template contract
(schema_version, locale-neutral semantic slot_id, ru/en labels, per-slot
basic/comfort presets, area_m2 as a provenance-tagged hint) with 6 built-in
rooms (bedroom, children, living, kitchen, bathroom, corridor) + SCHEMA.md.
room_library.py implements a function-first pipeline: templates+params ->
resolved IR dataclasses (NOT GARecord) -> deterministic, permutation-invariant
allocation (main=domain, middle=role, sub sequential; hard error on sub
exhaustion) -> a real .knxproj ZIP that is re-read via load_project (same path
as third-party projects, so generation never touches the classifier). Includes
validate_room_template, compose(), manifest + device BOM proposal builders.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Dogfooding the voice layer exposed two pairing defects: (1) _FUNC_WORDS had no
Russian entries, so 'вкл/выкл/цвет/яркость' polluted device identity and an RGBW
colour GA could not find its own switch; (2) subset identity matching let
'RGBW подсветка яркость' {zone, подсветка} grab 'камин подсветка' by subset and
its status by tie-break. Sibling pick and status_for_dpt now prefer EXACT
identity equality before subset fallback. 13/13 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- location phrases no longer vote a domain: "Boiler room light" is a light in
the boiler room, not an HVAC function (location stoplist stripped first).
- dropped the bare "температур" hvac term: a temperature GA takes its domain
from context (уставк/кондиц/тёпл words or its main group), so "Гостиная
температура" in a Sensors main is a sensor, not an hvac actuator outlier.
- "online/offline/heartbeat/watchdog/связь" are diagnostics terms (checked
first), so actuator-monitoring GAs named after the load they watch stay diag.
- check_policy: central macros ("Весь свет выкл") are never taxonomy outliers;
policy_no_reserve now also fires under an INFERRED taxonomy when no reserve
main exists at all; "весь "/"мастер" added to central-macro tokens.
Messy-house dogfood now reports exactly the planted misplacements (3/3) and the
reserve gap; demo 0 outliers. 13/13 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- DPT 1.018 occupancy: category sensor (was diagnostics; the name always says
sensor and the table fought it into a fake conflict).
- 9.001 temperature is SOFT: room temperature defaults to hvac, but a weather-
named temperature re-domains to sensor without a fabricated conflict.
- illuminance terms: "освещённость"/"illuminance" are sensor words and must not
be swallowed by the greedy lighting prefix "освещ".
- passive-range rule: a measurement inside a Sensors/Energy/Diagnostics main
takes that main's domain; a COMMAND is never retyped by a passive range.
Deliberately NOT tuned further: remaining real-project findings (Minsk 10/685,
Razdory 75/3646, demo 2/239) are defensible mixed-main observations a reviewer
should see once — zeroing them out would overfit one integrator's idiosyncrasy.
13/13 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Building a deliberately messy synthetic house instantly exposed four real gaps:
- _build_range_name_map: a middle group 0 starts at the same raw address as its
main, so the modulo heuristic let "Вкл/Выкл" clobber "Освещение" and the
main-range rescue silently died. Depth in the range tree decides now.
- a category from a whole-main DPT fallback (e.g. 20.609 -> hvac) was treated as
strong; only an exact (main,sub) table entry is strong now (is_exact_dpt).
- DPT 1.010 start/stop marked soft: ventilation timers use it, not just shutters.
- terms learned: Cyrillic а/с, сплит, вытяжк, яркост (RU only); weather/метео in
sensor, and sensor now outranks hvac so weather-station temperature is a
sensor, not an HVAC actuator.
Real-project noise after the round: Minsk 1 outlier/685 GAs (was 31 pre-P5),
demo 2/239, Razdory 46/3646. Regressions in test_domain_classifier. 13/13 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A 1-bit switch is domain-agnostic, yet DPT 1.001 defaulted to 'lighting', so an
"AC on/off" was classified lighting and tripped downstream checks. The domain is
now a combination: explicit name > strong domain-encoding DPT > main-group name;
a bare 1-bit GA with no signal is honestly 'unknown', never a guessed lighting.
An explicit name contradicting a STRONG DPT yields 'unknown' (a genuine conflict,
shown by explain_ga as contested), not a silent pick. Word-boundary matching so
"ac" no longer fires inside "terrace"; name terms broadened (EN+RU HVAC incl.
ac/a-c/air-conditioner, RGB/colour lighting).
check_policy taxonomy-outlier now flags only misplaced ACTUATORS (lighting/
shutter/hvac); sensors, scene links, thresholds and timeout params are the
cross-cutting GAs they are. HA generator still emits a switch for a now-HVAC
simple on/off, so no device is dropped.
Verified on 3 real projects: every AC on/off is HVAC; outlier noise fell demo
11->2, Minsk 31->6, Razdory 200->57. tests/test_domain_classifier.py (13/13 green).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Matter readiness, the completeness grade and command/status coverage each now
carry a `math` block: numerator, denominator, the exact formula that reproduces
the percentage, and (for Matter) the functions excluded from the denominator
because their category has no Matter cluster — previously a silent skip and the
biggest "why is this number what it is?" gap. Completeness also states its band
thresholds. Report-only, additive; scores unchanged, only now auditable.
Raised by external review. tests/test_explainable_aggregates.py (12/12 green).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A .knxproj is a user-supplied ZIP-of-XML — untrusted input. New safexml.py
centralises hardened parsing for every place we open an archive or parse XML
ourselves (load_project, check_device_parameters, app-program parser):
XML - dependency-free reject of DOCTYPE/ENTITY (billion-laughs / XXE;
legitimate ETS XML never carries one) + defusedxml parser-level
blocking when installed (added as a dependency).
ZIP - pre-flight against absolute caps (archive size, entry count, per-member
and total decompressed size, compression ratio) rejects zip-bombs;
member names checked for path-traversal / absolute paths; each member
read through a streaming cap so a lying header can't exhaust memory.
Violations return a normal error dict, not a traceback. Still read-only.
Raised by external security review. tests/test_safexml.py (11/11 suite green).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Council #2: the empty GA 2/5/2 spawned missing_status + policy_taxonomy_outlier +
a lighting classification on top of its real defect. An empty name means the
classifier can't be trusted, so detect_missing_status and check_policy now skip
blank-name GAs; the single root cause is empty_name (check_naming). Demo 2/5/2:
3 findings -> 1. tests/test_root_cause.py; full suite green (10 tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three independent reviewers (two external councils + a field integrator) asked for
the same thing: the enriched model mixes ETS facts, DPT structure and name
heuristics, and downstream tools treat it almost as fact. explain_ga makes the
reasoning auditable per GA — evidence per decision with a confidence tier
(authoritative ETS Function > structural DPT > heuristic name), the status-pairing
method, and CONFLICTS (name 'AC' vs DPT 'lighting' -> contested), the silent-
misclassification hotspot. Additive, read-only, no core-model change. Full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
'Does the function have *a* status' missed a dimmer with on/off status but no
brightness status (demo planted error #3), silently inflating coverage/Matter.
detect_role_completeness flags a brightness/position command (5.001) whose device
has no matching value status (missing_value_status), with a device-identity match
so it doesn't borrow a sibling's status. Demo recall 4/5 -> 5/5. Raised by the
external expert review. Test + demo ground-truth updated. Full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
External expert review (real run on the demo house) flagged: scenes (18.001)
wrongly marked missing_status + an unsafe DPT-1.011 status repair for them;
'All blinds down' warned while 'All lights off' was correctly INFO; and a Russian
'(статус)' suffix synthesised on an English project. Fixes: scene_no_status INFO
+ no synth status for scenes; broadened central-macro detection (generic 'all
blinds/shutters/...'); language-matched repair suffixes. Demo missing_status
warnings ~7 -> 1 genuine. tests/test_council_fixes.py; full suite green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rank findings by significance: config_value (setpoint/hysteresis/time/threshold —
an odd device is usually a real mistake) vs config_flag (mode/type) vs label
(per-room text, often intentional). New focus list surfaces the config-value clear
outliers first. On a real 275-device project this turns 422 raw outliers into a
9-item focus that includes the thermostats whose init-setpoint differs from their
siblings. Multilingual name hints (EN/DE/RU). Test updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Validate a project against ITS OWN rules, not a universal standard. With a YAML
profile the declared main-group taxonomy / naming / pairing is authoritative;
with no profile the taxonomy is INFERRED from the project itself and GAs that
deviate from their main group's own majority domain are flagged — never against
an alien standard (a well-organised real 356-GA project drops from 329 false
mismatches vs the default to 31 genuine self-deviations). Answers the recurring
integrator critique that 'your best practices aren't universal'. Example profile
+ test_policy.py included.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New module param_check.py + MCP tool check_device_parameters: reads per-device
ParameterInstanceRef values from the .knxproj project part (xknxproject does not
expose them), groups identical devices by application program, and flags the odd
one out — clear_outliers (a strong majority with a small minority, e.g. one
thermostat with a different setpoint/hysteresis) and split_configs (balanced
variants → review). Parameter names resolved from the app-program; encrypted
projects skipped honestly. Read-only, no ETS/bus. Validated on real 42-275-device
projects; synthetic test_param_check.py. Answers a community feature request.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a community-driven roadmap: cross-device parameter consistency check
(feasibility validated by a PoC on real 42-275-device projects — find the odd
thermostat/sensor out; not yet shipped), a Project Policy Profile, and an
honest note that in-ETS GA linking is left to existing add-ins / ETS7 Smart
Linking while we focus on read-only audit + an evidential model.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The nightly HA sync runs via shell_command in the Core container, not the SSH
add-on where install ran. Core has an empty known_hosts, so git push failed
"Host key verification failed" — the commit was made locally but never reached
GitHub. install.sh now seeds a repo-local known_hosts (ssh-keyscan) and pins
git config core.sshCommand so every context (add-on, Core, cron) uses the same
working ssh. Also drops a stray `.git-sync` arg from the pre-commit chmod line.
Records LESSON-06 + a symptom-table row in SKILL.md and SKILL.ru.md; CHANGELOG
Fixed entry. Field-validated on a live Home Assistant.
Circuit 1: real git in /config (deploy key, pre-commit secret scanner,
meaningful commits). Circuit 2: age-encrypted native HA backups in GitHub
Releases with rotation and a mandatory monthly restore drill. Ships
install/sync/scan/offsite/restore scripts, HA automations/snippets,
setup + architecture + incident + runbook references, EN + RU skill docs.
README/README.ru gain an ops-companion bullet in Scenario 3;
CHANGELOG [Unreleased] Added entry (no version bump).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Reframe README.md, README.ru.md and docs/index.html around what you can
do with the server: (1) design a project from a spec into a full
implementation kit, (2) audit/repair/finish an existing .knxproj,
(3) generate the Home Assistant layer — plus the device-library
foundation. Adds anonymised field numbers from real project validation
(~92% structural match on 3,600+ GA; 145 repair proposals on 3,646 GA).
No functional changes, tool count unchanged (25), not a release.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
