TODO: a check-off on the phone waits behind the reminder scan (218-220)

[Timtam]

Records Toni's report of 2026-10-06 and the three decisions it led to. Docs only; nothing is implemented yet.

The report

Completing a Vikunja task with the rotor action took more than a second to confirm on the iPhone.

The log shows:

• The write itself is quick. It takes about 0.2 s for a one-off task and about 0.4 s for a recurring one. The confirmation waits for it on purpose.
• The rest was a queue wait. upcomingRemindersJson runs on Expo's shared serial default queue. A scan mostly takes 3.4 to 4 s. updateTaskJson and the settings reads queued behind it.
• The first check-off had started that scan itself (09:33:57 to 09:34:01 UTC, 4.0 s). Its list reload flushed through cacheObserver.ts, which restarted the 2.5 s reminder debounce. The second check-off's PATCH came 32 ms after the scan ended.

TODO B13

• 218: the scan gets its own native queue on iOS and Android. rescheduleReminders runs again after a scan instead of dropping a request that arrives during one. A log line records when the check-off fired.
• 219: decided against a 5-minute cache on the phone; the scan stays live.
• 220: TasksFeature::update_task takes the previous row, so adapters send only what changed. Vikunja then skips the assignee PUT and the GET /tasks/{id} behind reconcile_parent. This is a deliberate behaviour change: a newer change from another device is no longer overwritten. The row has to cross the plugin interface, which today carries a bare Task, so this includes the ABI question.
• Open: a later scan stood still for 17 s and ended with every socket dead, as after each return from the background. The app was most likely suspended. Whether such a break also happens in the foreground is open.

First check (ee3e5e8)

Two checkers (code, log) and one skeptic per finding. Seven findings were confirmed and fixed:

• The trigger chain. The scan starts 2.5 s after the cacheObserver flush, not 2.5 s after the check-off. That is about 3.8 s after the write.
• Scan times. They are now stated as measured: mostly 3.4 to 4 s, once 5.2 s, once 21 s.
• The plugin interface in 220. Every external adapter is reached only through it, and ffi_update_task decodes a bare Task.
• 219. The phone has no push, so a change from another device arrives at the next scan, not at once.
• Wording. The reminder scan is now called "Erinnerungs-Durchlauf" throughout, so it no longer shares the word "Abgleich" with sync. The GET is named as the task's own read.

Three were refuted.

Second check (b543594)

Five findings were confirmed and fixed:

• Which scan. The scan the second check-off waited on is now named with its time and length, so it cannot be read as the 21 s one.
• The 17 s silence. The Vikunja server answered in 51 ms; then the whole log was silent. Every socket was dead afterwards, including ones to servers the scan was not using, as at each return from the background. The flag now says the app was most likely suspended.
• The plugin interface in 220. The device adapter is reached directly, not through the plugin interface. Only six plugins have ffi_update_task.

Two were refuted.

Third check (8bdb03d)

Three findings were confirmed and fixed:

• Built-in adapters. The local store is the second task adapter that both hosts call directly, besides the device adapter.
• The 17 s silence. The Vikunja server sent its headers in 51 ms, but the body never arrived.
• Grammar. One genitive was wrong.

220 now also says that the host must read the cached row before every write.

Fourth check (528fec9)

Three findings were confirmed and fixed:

• Where the cache is read today. Both hosts already read the cached list after every external task write (write_through_task), and again after completing a recurring task. Neither read comes before the write.
• The 90 ms. They were measured on the iPhone. The two saved requests apply to both hosts.
• Who calls whom. The sentence on the built-in adapters now says plainly that the hosts call them.

One was refuted.

Fifth check (528fec9)

No findings. Three wording suggestions were refuted: each sentence is correct as written.

Docs

TODO only. DESIGN describes no behaviour that changes here.

🤖 Generated with Claude Code