tuya2ildevice

The one place that interprets Tuya. Sans-IO Python: takes a Tuya device as rustuya / rustuya-bridge know it and produces an ildevice: descriptor, live values and events from dps packets, and dps for ildevice commands. It opens no sockets and reads no clock. Other projects (rustuya-local, ildevice hosts, rustuya-manager) import this instead of carrying their own Tuya knowledge.

pip install tuya2ildevice              # no dependencies;  tuya2ildevice[host] adds paho-mqtt for `tuya2ildevice.host`,  [test] for the tests

from tuya2ildevice import TuyaDriver, Connected, Message, Command, descriptor_of

descriptor_of(device)                 # just the ildevice descriptor of a tuyadevices.json entry
d = TuyaDriver(device)                # the state machine (il.md section 8)
d.handle(now, Connected())            # -> [Descriptor]
d.handle(now, Message("active", {"1": True}))   # -> [Value(...), Event(...)]
d.handle(now, Command("switch_1", "off"))    # -> [SendMessage("set", {"dps": {"1": False}})]  or  [Reject(...)]

Packets and commands

• Numeric properties use the native unit of their values. For example, a plug reporting 500 mA publishes 500 with unit mA; a suggested display unit such as A does not change the wire unit or magnitude.
• Message(channel, json): active (device push: fires events, accumulates add_ele-style deltas) or passive / state (readback, snapshot: values only; with delta.accept_passive, a passive report adds its delta too, see Delta settings). json is a flat {dp: value} map (the host has already decoded whatever the bridge's own wire payload looked like — see Hub.on_bridge_message below).
• Command(prop, value) is checked as il.md section 5 says before anything is sent; failures come back as Reject. Values convert as in Home Assistant core (brightness goes through 0..255), so a written value can differ slightly from what is read back. A cover's position is the device's own number, never mirrored (core mirrors it); a device that counts the other way is set so by its cover settings (below).
• A passive report (a live one: the host drops retained ones) that changes a value is the device's own push, like active: converters see it as live; one that changes nothing is a readback. Events fire only from active. state, the bridge's merged view of what active / passive carried, is a snapshot: the first one after a connect sets everything, a later one only what it alone brings, and never as a push. An increment (report_type: sum) is added only from active.
• Covers of class garage/gate are read only unless TuyaDriver(..., allow_hazardous=True) (il.md S-1).
• A dp no platform table claims gets no property, as in Home Assistant core; TuyaDriver(..., expose_unused=True) gives each one a property chosen by its Tuya type. Integer, Enum and Boolean are writable only if the dp is in function; String, Raw, Json and Bitmap are read-only. config if writable, else diagnostic.
• Assembled: switch, button, select, number, sensor, binary_sensor, event, light, cover, fan, siren, valve, humidifier, climate, alarm (kind alarm), vacuum. Not assembled: camera (a stream is outside the IL; driver.unsupported lists it). A doorbell alarm_message event loses its message text (an ildevice event has only a kind).

User overrides

Tuya devices are fragmented: a cloud schema can give a dp a non-standard code, leave a dp out, speak other words (on/off/pause instead of open/close/stop) or run a position the other way. Overrides fix that in Tuya's own terms, before classification, so the fixed device goes through the same tables as any other and every IL consumer (il-ha, a discovery publisher, ...) sees the fix.

TuyaDriver(device, overrides=mapping) / Hub(devices, overrides=mapping); mapping is {product_id or device id: block}, already loaded (merge_all([...]) merges several, later wins; tuya2ildevice.host.load_overrides reads a directory). Details and the block format are in overrides.py:

{ "<product_id>": {
    "dp":     {"104": {"code": "percent_control", "type": "Integer", "values": {"min": 0, "max": 100, "scale": 0, "step": 1}},
               "101": {"code": "control"}},
    "remove": ["percent_state"],
    "category": "cl",
    "remap":  {"control": {"alias": {"on": "open", "off": "close", "pause": "stop"}}},
    "cover":  {"invert_position": true, "settle": 5},
    "delta":  {"accept_passive": true},
    "props":  {"countdown_1": {"label": "Timer", "category": "config", "rw": false}},
    "device": {"model": "Sliding Window Opener"},
    "converters": {"glow": {"on": "bright"}},
    "expose_unused": true } }

A property defined from a dp builds what the tables cannot, such as a cover from a position dp alone (a window opener whose category has no cover table):

{ "<product_id>": {
    "dp":     {"104": {"code": "percent_control", "type": "Integer", "mode": "RW",
                       "values": {"unit": "%", "min": 0, "max": 100, "scale": 0, "step": 1}}},
    "device": {"kind": "cover", "class": "window"},
    "props":  {"position": {"src": "percent_control", "role": "position"},
               "open":     {"src": "percent_control", "type": "trigger", "role": "open", "send": 100},
               "close":    {"src": "percent_control", "type": "trigger", "role": "close", "send": 0},
               "battery":  {"src": "residual_electricity", "role": "battery", "category": "diagnostic"}} } }

Its type follows the dp's (Boolean binary, Integer number, Enum select, others text) unless given; min, max, step, unit and options come from the dp unless given; it is writable when the dp is in function (rw: false makes it read only); a trigger writes its send value. Values go through remap like any other.

Unknown keys raise OverrideError. Hub.reload(mapping) applies new overrides live: republishes changed descriptors, clears removed properties, and works a changed device's values out again from the dps it last had.

Cover settings

Which way a cover's position and commands run, and whether its motion is inferred, are the cover's settings, not remap: a curtain whose direction is wrong is fixed from its own configuration switches. Hub(devices, device_settings=True) (a host with somewhere to keep them) offers each switch only where it changes something:

A cover with a target opens and closes by writing it (100 / 0), and its command dp only stops it: it has no Invert control. The switches are binary, the state source is select; all are rw, category: config, and show the settings in effect; a hazardous cover's (garage, gate, ...) are read only unless allow_hazardous. A later cover of the device (control_2, ...) has its own, named by its group (control_2_cover_invert_position).

Writing a switch sends nothing to the device: the driver answers with the value and SettingsChanged(block), and the Hub with SaveSettings(device_id, block), the device's whole settings block (e.g. {"cover": {"invert_position": true}}) for the host to keep. tuya2ildevice.host keeps it in zz_settings.json (host.SETTINGS_FILE) in the overrides directory (Runner(..., on_settings=watcher.save_settings)), which then reloads at once.

The same settings are a cover block in any override file, where a product's block is the starting point for its devices and a device's (the switches write the device's) wins:

A later cover's keys go in an object under its group's name: {"cover": {"invert_position": true, "control_2": {"invert_position": true}}}. An inversion is laid over the dp's remap (a remap.invert still there cancels it), and the swap is composed with the dp's remap.alias, so a device that says on / off / pause keeps its alias. remap.<code>.invert on a cover's position or command dp and converters.cover_motion still work, and log that they are deprecated: move them to cover.

control maps live open/close/stop reports directly to opening/closing/stopped, even without position feedback. Boolean true/false and FZ/ZZ/STOP use the same control vocabulary as inference. The order on input is remap.alias, existing invert_control (where applicable), then invert_reported_motion. The last setting swaps only opening/closing from control reports; it never changes sent commands, positions, inferred direction, or endpoint states. Live stop is always stopped. Position updates alone do not settle a directly reported movement.

Delta settings

A delta dp (report_type: sum, a plug's add_ele) reports the energy used since its last report, not a meter reading: its property is the running total tuya2ildevice adds up (series: counter), so a repeated increment (5, then 5 again) adds twice. The total is published in the DP’s native unit with its scale applied: for scale: 3 and unit kWh, a raw increment of 5 adds 0.005 kWh. By default only an active push adds: a passive report is a readback of what the device already pushed. A device that never pushes it, and reports it passively only, needs passive reports counted too:

It is the delta block's accept_passive (default false) in an override file, and saves as the cover switches do ({"delta": {"accept_passive": true}}). One report seen twice with the same Tuya t adds once; a readback the host asks for (a get, on a reconnect) carries no new increment but adds once more, so turn it on only for a device that does not push. A reload (an override or a setting changed) keeps the totals.

Retained snapshots/readbacks (state, and non-live passive reports) never start motion: their control may be the last command. A relevant snapshot reports only the position's resting state (closed at 0, open at 100, stopped between), or no state without position. Disconnect clears motion; reconnect/reload seeds only resting state and waits for a live report. Duplicate live reports do not republish unchanged state. Direct report mode has no settle timer.

inferred preserves CoverMotion: a control word can start motion; subsequent position feedback corrects its direction, while an independent target can be compared with current position. A shared current/target DP is never compared against itself. Without independent feedback, command direction and arrival cannot be verified; a target-only update does not prove movement. none generates only resting state from position and never movement. Legacy infer_motion:false keeps its previous behavior of omitting the separate cover_state property (position remains available); explicitly selecting none also provides the resting state property.

The old Infer motion switch is replaced by the state source select, not exposed alongside it. Existing JSON and saved zz_settings.json Booleans remain readable; the select shows their mapped mode and subsequent selections save state_source. Explicit state_source wins even if an older infer_motion entry remains. A user converter owning a cover's cover_state retains ownership: that cover gets neither the built-in state producer nor its source/direction controls. Other covers remain independent, including when legacy converters.cover_motion owns the first one.

Settings remain saved while their controls are hidden: target inversion is dormant while current and target share a DP, and reported motion inversion is dormant outside control mode. Switching back restores them. Current-position inversion follows the selected display source; when that source is also the target, it inverts both reads and writes. Property IDs and cover group IDs stay stable through source changes. Host reload cancels old timers and clears the old cover state before publishing a fresh snapshot. A bare sans-IO caller must save SettingsChanged and reload, just as for the existing inversion switches.

For f6jujmx0is5td50x, remove its old remove, remap and converters.cover_motion JSON entries, then configure the device (no product ID is hardcoded):

1. Use target as current position: ON.
2. Invert current position: ON.
3. Motion state source: control.
4. Invert reported motion: ON.

This selects percent_control (DP 2 in the referenced schema), inverts its reads and writes, and reverses only the reported motion. The DP name does not establish that it reports physical position: if firmware reports only a target, the displayed position is only that target; intermediate travel, arrival and physical resting position cannot be determined from it.

The previous rustuya-homeassistant JSON selected DP 2, inverted position reads and target writes, and mapped control directly; it was not enrolled in the Python converter because it had no state_stream: "derived". Its MQTT open/close commands used control open/close. This package continues to send target 100/0 when set_position exists (after position inversion), so its command path is different from that earlier implementation. Report inversion does not change that path.

The package carries no block for any product or device. A fix for a model goes in a converters file: the user's own, or the override pack, which hosts copy into that directory.

Code converters

A per-device Converter object sees the driver's dp state (after remap) on every packet and returns property values (and timer requests, since it cannot read a clock). Its properties may replace the tables' ones of the same name, and one with rw: true (or a trigger) is written through its write(prop, value, codes), which returns the dps to send as {code: value}. See converters.py.

class MyConverter(Converter):
    def __init__(self, config): ...
    def props(self):  return {"filter_low": {"type": "binary", "class": "problem"}}
    def update(self, now, codes, changed, active):  return {"filter_low": codes.get("filter_life", 100) < 10}  # or Result(...)
    def write(self, prop, value, codes):  return {"<code>": value}        # only for rw / trigger properties

TuyaDriver(device, converters={"<product_id or device id>": [lambda device: MyConverter({})]})
TuyaDriver(device, converter_types={"my": MyConverter}, overrides={"<product_id>": {"converters": {"my": {}}}})

The built-in CoverMotion gives a cover its cover_state role (open / closed / opening / closing / stopped); the cover settings turn it on (state_source: "inferred", or legacy infer_motion) on the dps the cover actually has, after its inversions. A command word only says a move started, the position reports after it decide which way; a snapshot never starts motion. Naming it in a block ({"converters": {"cover_motion": {...}}}) still works, deprecated. Timers come out as SetTimer (driver) or Schedule (Hub); the host calls back with Timer / hub.on_timer(now, id, name).

Overrides as files

tuya2ildevice.host.load_overrides(path) reads a custom_converters/ directory (or one .json file) and OverrideWatcher(path, runner, base=...) follows it, reloading the Hub when a file changes:

• *.json: override mappings, deep-merged in filename order (99_local.json refines 10_base.json).
• *.py: define CONVERTERS = {"name": factory}; an override block turns one on by name. The code runs in-process.
• A bad file is reported and left out; the rest still loads. Overrides the Hub refuses leave the ones in effect.
• zz_settings.json (SETTINGS_FILE, loaded last) holds the settings written through IL, a block per device: await watcher.save_settings(device_id, block) replaces the device's (an empty block removes it), writes the file atomically and reloads at once. Keep your own overrides in other files; with one .json file instead of a directory nothing is saved.

The override pack

Fixes for non-standard devices can reach users before the next release: pack/ on master holds override files and tuya2ildevice.host.pack.sync(directory) copies them into a host's custom_converters/ directory, where the watcher loads them like the user's own (rustuya-local does this at start and daily, unless turned off). Every file is checked against the manifest's SHA-256, and sync only writes or removes the files it put there, as recorded in .tuya2ildevice_pack.json. A user's file with the same name, or a pack file the user has edited, is left alone. A manifest entry can be limited to a range of tuya2ildevice versions (pack.LEVEL). .py pack files run in the host's process, like the user's own files. The pack never writes zz_settings.json.

tuya2ildevice <-> an IL host over MQTT

Hub (mqtt.py) owns one driver per device and maps il-mqtt.md on the IL side. On the bridge side it takes already-decoded input, not a raw topic/payload: on_bridge_message(now, device_id, inp, retained=False) where inp is Connected() / Disconnected() / Message(channel, dps), and its writes/reads to the bridge come out as abstract BridgeCommand(device_id, action, dps) — Hub has no idea rustuya-bridge's own MQTT topics exist, let alone that they're configurable. Rendering BridgeCommand into a real topic+payload, and turning a real bridge MQTT message into Connected/Disconnected/Message, is the host's job — correctly, that means using pyrustuyabridge's bindings (match_topic, render_template, tpl_to_wildcard, parse_seed_dps), which mirror the real bridge's own template/payload parsing, not a hand-rolled one. rustuya-local is that host for a real rustuya-bridge; its bridge_client module is the reference implementation.

tuya2ildevice.host is the IL-side host: Runner drives a Hub on one IL transport (MqttTransport over paho, or the in-process InProcessTransport), keeps the Last Will presence (M-12), reconnects, adds/removes devices while running (set_device, remove_device, sync_devices), follows rustuya-manager's tuyadevices.json (DeviceWatcher), and routes every BridgeCommand Hub produces through an injected on_bridge_command callback — supplied by whatever owns the real bridge connection — plus a matching runner.on_bridge_message(device_id, inp, retained=False) entry point for feeding decoded bridge input back in. A host that already heard a device before it drives it (a restart: the bridge's retained link state and state) hands that over as seed (set_device(device, seed), sync_devices(records, seed=...)), so the device goes out as it was — available: true with its values, never false in between — and stop(offline=False) keeps the presence online for a host that starts again at once: IL consumers see no gap.

One producer per IL prefix and source: await producer_running(il, hub.il.presence) before starting says whether another one is serving it. It returns True when a running Runner answers a probe on <presence>/probe (the answer comes on <presence>/alive; neither is retained, and IL consumers, subscribed to _producer/+, do not see them). It returns False when presence is not online. It returns None when presence is online but nothing answers: a stale presence whose Last Will never reached the broker, or a producer on tuya2ildevice before 0.3.5.

Register the devices on the bridge yourself (add); the hub never touches keys.

Layout

src/tuya2ildevice/
  driver.py    TuyaDriver: packets/commands <-> outputs          io.py      inputs and outputs as data
  assemble.py  engine entity plans -> ildevice props/roles       checks.py  il.md section 5 command checks
  mqtt.py      Hub, IlTopics                                     tuya/      the DP engine (see below)
tuya/          classify + platforms, adapter (raw dps <-> values), quirks, ops, codecs, units
tuya/tables/   per-platform description tables, generated from HA core     tuya/quirks/   from tuya-device-handlers
tuya/data/     HA's allowed units per device class
scripts/       generators for those data files, and golden/ (builds the golden data; needs HA core + oracle venvs)
docs/          engine-spec.md (the engine's behaviour), analysis/ (how it was derived from HA core)
tests/         golden/ = 324 HA core fixtures + core's own snapshots; chain/ = the same through an IL host's planner

The engine reproduces Home Assistant core's tuya integration exactly, including its quirks, and the golden tests pin it: tests/golden/golden.json is core's own entity snapshots for every fixture. To follow a new HA core / tuya-device-handlers release, run the generators in scripts/ against it, then the tests; a difference is either a real change to adopt or a regression.

One deliberate exception: where Tuya's own category list (tuya/tables/_tuya_categories.json, generated by scripts/gen_tuya_categories.py from Tuya's standard instruction set page) disagrees with core, Tuya's list wins. The decisions are data too (_tuya_standard_rules.json, applied by tuya/standard.py): a kg ("Switch") category's switches are switch, not core's outlet plug icon; cz ("Socket") and pc ("Power strip") stay outlet. The golden tests apply the same file to core's expectations, and each such difference is a tagged deviate row in docs/engine-spec.md (P-28).

A second one: a cover's position is not mirrored. The bridge shows the device's number and so does the IL; an installation inverts a device that counts the other way itself (its cover settings). The golden tests mirror core's expected position and position writes where core mirrored them (core_reverses in tests/test_golden.py).

Tests

pip install -e .[test] && python -m pytest

tests/test_adapter.py compares the raw-dps adapter with the Tuya SDK and is skipped unless tuya-device-sharing-sdk==0.2.15 is installed. The spec repository (../ildevice, or $ILDEVICE) supplies the schema and the language-neutral vectors (commands, wire values, topics) read directly from its checkout; those tests are skipped if it is not there. tests/chain/ compares every Home Assistant core tuya fixture with core's entity snapshots through an IL host's HA-free planner (needs that host's ildevice.core on sys.path; not collected without it).