[miroslav-ondra]

The findings of the review of #276 (Evok 3.1.0), fixed in dev-3.1 before the release.

Changes

A failed end of a pulse left the output on (blocks/sequencer.py)

• The last step of a sequence ended it before its write. When the write failed (a noise or a timeout on RS485, an error of the unit), the sequence was gone, nothing wrote the end again and the relay stayed on; finish_all() at the shutdown did not write it either. A failed middle step lost the final of a sequence the same way.
• A step failed by an error of the communication (UnitCommunicationError, TModbusError, TimeoutError, OSError) is written again until it succeeds or the sequence is cancelled or replaced, the sequence stays running meanwhile. The delay starts at STEP_RETRY_DELAY (1 s) and doubles up to STEP_RETRY_MAX_DELAY (30 s), a missing unit does not load its bus every second. The failure is logged, again when the kind of the error (its root cause) changes and every 5 minutes, a later success too.
• A step refused by the unit (an exception response other than Server Device Busy, Acknowledge or of a gateway, e.g. Illegal Data Value or Server Device Failure; an invalid request) or another error ends the sequence and is logged; written again it would be refused forever and the output refused the changes.
• The final state of such a sequence (if it has one) is written as its last step while it still runs: a change of the output meanwhile is refused (409), finish_all() waits for it, an error of the communication writes it again; it is not written by a task the sequence no longer owns, nor when it is the refused step itself (the end of a pulse).
• The step is done (step of the state) after its successful write; the last step ends the sequence after it (under the lock of the unit, the state read after it is not running).
• finish_all() waits for the steps being written, then writes the final state of the sequences whose last step was not written (Running.finished): a failed end of a pulse while Evok stops is written by the final state.

Watchdog (modbus/special.py)

• A changed timeout is an event, the clients kept the old one.
• timeout, value and reset are checked before any change, timeout must be a whole number of ms: {"value": 1, "timeout": "1.5"} enabled the watchdog and set the alias before int() failed (400 after a partial change), 2.9 was cut to 2, an invalid value or reset set the alias or wrote the registers before; reset of a unit without the reset coil is ignored as before.

An invalid alias after the write (modbus/base.py, devices.py)

• set_devices() checks the aliases of all assignments (Devices.check_alias()) before the first write, in the order of the changes (by the units), with the aliases released and taken by the earlier assignments. A used, invalid, reserved alias or one equal to a circuit switched the output and then the request failed with 400; the same alias twice in one bulk is refused, an alias moved to another device in one bulk works; an alias which is not a string (0, false, a list) is refused as invalid; alias: null is not given, the alias is kept. All APIs set the devices by set_devices(); set_alias() uses the same check.

AccessorBit writes with its pin (#279), a bit accessor with pin=False was pinned.

Docs: the retried step in sequencer.md and the API changes, the watchdog timeout, the alias.

Testing

• pytest tests: 1302 passed; new tests: alias: null keeping the alias, the refused end of a pulse sent once, reset of a watchdog without the coil, the final state written while the sequence runs (also again after a noise), finish_all() waiting for it, Server Device Failure refused and Server Device Busy retried, an error with a changing text logged once, a step refused by the unit ending the sequence with its final state, the doubling delay, the changed error logged, the invalid watchdog value and reset, an alias which is not a string, the failed step written again and logged once, a refused step ending the sequence, the step done after its write, the last step ending the sequence after its write, the failed end of a DO pulse written again, the final state after a failed last step at the shutdown, the same alias twice in a request, an alias moved in a request, an alias released by a later unit, the watchdog timeout as an event, the invalid timeouts (1.5, 2.9, -1, x) without any change, a whole timeout as a string, the invalid aliases before the write, the bit without a pin. Each fails without its fix. ruff check and ruff format --check src clean.
• Patron M527 (192.168.221.159): a pulse of 0.5 s of led ext2_01; {"value": 1, "alias": "tmpx"} with tmpx used by ext2_02 is 400 and the LED stays off; wd ext3 {"value": 1, "timeout": "1.5"} is 400 and the watchdog stays disabled with its timeout; a bulk moving the alias pump from led ext2_02 to led ext2_03 succeeds; wd ext3 {"alias": "wdx", "value": "maybe"} is 400 without the alias.

🤖 Generated with Claude Code