[evnchn]

The GUI read the ODrive on the event loop: NiceGUI polls every widget bound to a fibre object every 0.1 s, and each read is a blocking USB round-trip. On a real board, main logs binding propagation … took 0.016–0.020 s on every tick. With this PR it logged one such warning in about 4 minutes.

Two things on main this also fixes, both found on the bench:

• An Idle click could be lost. The state toggle drops a click on the option it already shows, and the state shown can be a refresh old. So: click Loop, a refresh still reads Idle, the firmware then enters Loop, and the click on Idle is never sent. The state is now three buttons that look the same, and every click is sent.
• Save and reboot logged a traceback. import odrive loads libfibre twice (odrive.pyfibre.fibre and top-level fibre). Device objects come from the second copy, so the GUI's ObjectLostError and EmptyInterface checks never matched on hardware. It now checks both.

Would you rather have the libfibre fix as its own PR? It is the last two commits and stands alone.

Bench run: ODrive 3.6 clone, real USB, motor never energized

Board: ODrive 3.6 clone, FW 0.0.0 (dev), 19.96 V DC, both axes Idle throughout. Step 4 was skipped on purpose: this clone has no hardware overcurrent protection.

step	result
read latency	0.328 ms per read (200 × Iq_measured)
page load, live values	pass. One binding propagation warning (0.013 s) in about 4 min, where main logs 231 in about 37 s, one per tick at 16–20 ms each
write path	pass. pos_gain 5.0 → 5.5 survived refreshes and Save (this firmware reboots on save); the panel came back reading 5.5 from the device; restored to 5.0
Loop/Idle on hardware	skipped (see above); covered by the tests and a Chromium run against the mock
device drop	partial: every save/reboot drops the device with the page open; app and page stayed up, panel back within ~1.5 s. USB was not physically unplugged
tracebacks on save/reboot	main: logged, e.g. fibre.libfibre.ObjectLostError: the object disappeared from reboot(), then 'EmptyInterface' object has no attribute 'Iq_measured' from the power timer. Fixed by the libfibre commits

How the device is read now

• Widgets bind to a small local mirror per fibre object instead of the fibre object.
• One worker thread (_usb) refreshes the mirrors at 10 Hz. The power/voltage/error timers and the plot sampler read there too.
• Every write and command (setting edits, state requests, motion inputs, save, reboot, clear errors) is queued on that worker in click order. So "lower current_lim, then Loop" reaches the device in that order.
• A refresh that read a value before the user edited it does not overwrite the edit.
• Unchanged: the reads made while a panel is built still happen on the loop, once per panel.

Tests: each new one fails without its fix

test	without the fix
test_live_values_are_not_read_on_the_event_loop	main: 604 device reads on the loop thread in 1.2 s
test_a_refresh_that_read_before_an_edit_does_not_undo_it	fails with the write-order check removed
test_a_quick_loop_then_idle_is_not_dropped	fails when clicks are compared with the cached state
test_a_command_does_not_overtake_a_setting_changed_before_it	fails when commands bypass the worker queue
test_idle_is_requested_even_while_idle_is_the_state_shown	the toggle version on main sends [8] instead of [8, 1]
test_a_device_dropping_off_mid_save_or_reboot_is_not_an_error	ODrive write failed logged
test_lost_device_drops_only_its_panel (existing)	fails once the mock's lost device uses the top-level EmptyInterface, as on hardware

The mock now has a separate top-level fibre with its own classes, like the real package. It also counts reads per thread and can give each read a latency (MOCK_ODRIVE_READ_MS).

The browser half of the lost Idle click (Quasar ignoring a click on the selected option) is outside NiceGUI's User simulation, so it was checked in Chromium against the mock, with a firmware that never switches: click Loop, wait 2.5 s, click Idle. main sent only 8; this branch sends 8, then 1.

Gates: ruff, ruff format, mypy clean; pytest 18 passed (repeated, also with 1 ms mock reads). src/main.py boots against the real locked odrive and serves the page.

Known limits

• One worker for all devices: a hung device stalls every panel's updates, but the page stays responsive (before, it froze the app).
• pyfibre is called from the worker while panel builds read on the loop. The bench showed no problem; that is not proof.
• A failed write is logged; later commands are not held back, as before.
• Each open browser tab polls on its own, as before.

Review

Five Codex passes on the worker change; each blocker was reproduced, then fixed with a test. They are the ordering, lost-click and stale-state cases in the test table. Last verdict: "suitable for supervised bench testing … 8/10". The bench above followed.

The libfibre fix was reviewed by GLM-5.3 (Codex was out of quota). It flagged an unguarded sys.modules['fibre'] lookup that could fail on older odrive releases; that is now importlib.import_module('fibre'), the same import odrive's discovery thread does. Its other high finding, that the new test asserts nothing, does not hold: NiceGUI's user fixture fails a test on any ERROR log, and the test fails with the fix reverted (ODrive write failed). .trigger('click') fires every visible button.

🤖 Generated with Claude Code

[falkoschindler]

@evnchn Thanks, this is a clean way to get the USB traffic off the event loop. The bench run and a test for every change that fails without it make it easy to trust. Keeping the libfibre fix in this PR is fine, since it sits in its own commits.

One thing for later, not for this PR: sync_mirrors still re-reads every bound config value at 10 Hz, although only current_state needs to feel live. Syncing the config mirrors at about 1 Hz would remove most of the worker's load. That load is not new, because NiceGUI polled the same bindings before, so it can wait for a follow-up.

Two small fixes pushed to the branch

• No repeated Loop request (d8583b7): The state buttons sent every click, so a click on Loop while the axis was already in closed loop wrote CLOSED_LOOP again. As your own comment notes, 0.5.x firmware then leaves and re-enters closed loop. Idle is still sent on every click, but any other state is now only requested when the mirrored state differs. At worst a stale mirror swallows a Loop click, which starts nothing.
• Lost sub-objects (e4b79d0): _send and _on_usb only checked whether the object passed in was lost, but the power label reads the current controller and the plots read the encoder. libfibre tears a lost device down children first, so such a read could fail while the axis still looked alive and log a traceback. is_lost now lives in controls.py and is used there and in main.py, so there is one definition of "the device left the bus" again.

Both come with a test that fails without the fix.

※