К основному содержанию
скилл CX Codex

fmess-live-check — скилл для Claude Code: проверки fMess

Скилл fmess-live-check для Claude Code проверяет поведение fMess при сбоях сети, сна, перезагрузке и работе уведомлений. Запускается перед изменениями этих функций и при жалобах на зависшую отправку или потерю фокуса.

MIT ★ 122 обновлён 2026-11-15

Установить в Claude Code · Через Skills CLI (Vercel) Все способы и платформы ↓
npx skills add toanbku/fmess --skill fmess-live-check -a claude-code
Открыть на GitHub

Что делает fmess-live-check

Скилл запускает проверки в установленном приложении fMess: имитирует сон компьютера, обрыв и возврат сети, сбой воркера зашифрованных чатов, уведомления и звуки. Отдельная проверка после перезагрузки и сетевого сбоя тестирует поиск, сочетание ⌘K и выделение текста сообщения. Окно приложения остаётся на заднем плане; тест ввода набирает «zq» и удаляет символы, не отправляя сообщение.

Для проверки нужен debug-сборочный вариант fMess. Скрипты tests/app/live-check.sh и tests/app/focus-check.sh рассчитаны на macOS; сетевой тест требует сборки с --features test-proxy. Первый скрипт использует локальный прокси tests/app/net-proxy.py, выводит результат для каждого случая и сохраняет логи. Полный запуск четырёх сценариев занимает около 25 минут, проверка фокуса — около 8 минут.

Какие задачи решает

  • проверить восстановление fMess после сна и обрыва сети
  • найти причину зависшей отправки или пустого экрана
  • проверить, возвращаются ли звуки и уведомления
  • убедиться, что после перезагрузки поиск и выделение текста работают

Примеры запросов агенту

  • Проверь fMess после сна компьютера и отключения сети.
  • Разбери, почему уведомления fMess приходят без звука.
  • Проверь фокус поиска и выделение текста после перезагрузки приложения.

Как установить fmess-live-check

1Через Skills CLI (Vercel)

npx skills add toanbku/fmess --skill fmess-live-check -a claude-code

2Вручную: положить файлы

~/.claude/skills/fmess-live-check/SKILL.md

Папка скилла целиком; для одного проекта — .claude/skills/fmess-live-check/

По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов

Вопросы о fmess-live-check

Как установить fmess-live-check в Claude Code?

Выполни npx skills add toanbku/fmess --skill fmess-live-check -a claude-code. Файл скилла будет доступен по пути ~/.claude/skills/fmess-live-check/SKILL.md.

Что делает fmess-live-check?

Скилл проверяет fMess в реальном приложении, имитируя сон, потерю сети и сбой воркера зашифрованных чатов. Он также тестирует уведомления, звуки и фокус элементов интерфейса.

Что нужно для работы fmess-live-check?

Нужны macOS и debug-сборка fMess. Сетевые сценарии требуют сборки с --features test-proxy; проверки запускаются скриптами tests/app/live-check.sh и tests/app/focus-check.sh.

Исходный текст

SKILL.md · MIT
---
name: fmess-live-check
description: Check in the real fMess app what happens when the computer sleeps, the network dies or comes back, Messenger's encrypted chats' worker dies ("Couldn't load chats"), when a notification rings, and whether the keyboard and mouse still go where they're put (search field, ⌘K, selecting a message's text) after a reload or a dead network. Run it before shipping any change to connections, reloading, notifications, sounds, timers or focus (19-recover, 14-notifications, 05-messenger-prefs, 06-routes, 01-start, notify.rs), and when a user reports stuck "Sending", missing notifications or sounds, a blank or error screen, a field that can't be typed in, typing that lands in the message box, or text that can't be selected. Fix by simulating the case, not by releasing and waiting for reports.
---

# Live check

`tests/app/live-check.sh [wake] [network] [chats] [sound]` (all four by default;
macOS). It needs a debug build that can go through a proxy:

```
npx tauri build --debug --bundles app --features test-proxy
```

It quits the running fMess, runs the debug app through `tests/app/net-proxy.py`
(a local CONNECT proxy that can cut the network), and opens the installed fMess
again at the end. The window isn't brought forward (`FMESS_PROBE_NO_FOCUS`).
It prints `PASS`/`FAIL` per case, then `PASSED` or `FAILED (n)`; logs go to
`$LIVE_OUT` or a temp folder (each case's `app.log`, `proxy.jsonl`, `stats.json`).
About 25 minutes for all four.

## How things are made up

| What | How |
|---|---|
| Network dies silently (sleep, Wi-Fi drop, NAT) | proxy `blackhole`: open tunnels pass nothing and never close; new ones get no answer |
| Offline | proxy `refuse`: open tunnels close, new ones refused at once |
| Every connection dropped | proxy `reset` |
| Sleep | a clock jump between two ticks of 19-recover's timer (`window.__fmessCheckAwake(now)`, debug) |
| A message being typed | a hidden contenteditable with text, focused (not Messenger's box, so nobody sees "typing…") |
| "Couldn't load chats" (#24) | Messenger's own `MAWSetupWorker.setPreventWorkerRestart(true)` + `terminateWorker`, then work for the worker (reading ahead encrypted chats, which opens none) |
| Messenger's sound | its real `MWLSChatPopperCommon.playChatTabSound(message)` |
| Do Not Disturb, Messenger's sounds | `__fmessSetMessengerPref('muteFor' / 'soundsOn')`, reported at once (`__fmessReportMessengerPrefs`), put back as they were |
| The Mac muted | the probe logs `sound: mute`; the script runs `set volume output muted true` and puts it back |
| Windows' ring decision on a Mac | `window.__fmessRingAnywhere = true` (debug) |

Debug builds log what the page does: `[page] main recover: … @<ms>` (19-recover),
`window.__fmessSockets` (every WebSocket: opened, closed, last frame each way),
and `[fmess] notify #n: <title> (…, sound true|false)`.

## Cases

**wake**
- a timer a minute late (a hidden page's timer) doesn't reload; a 2-minute sleep doesn't
- after a 6-minute sleep: no reload while a message is being typed, reload once it's sent
- after the reload, messages (gateway.facebook.com) and encrypted chats (web-chat-e2ee) talk to Facebook within 20 s
- waking with the network off, then dead: no reload until it's back, reload within 10 s after

**network** (no sleep)
- a minute of dead network: messages back within 10 s by Messenger itself; fMess reloads at most once, within 60 s, for the connections Messenger leaves down
- every connection dropped: back within 10 s, no reload
- 30 s offline: back within 60 s
- 2.5 minutes offline: fMess reloads within 60 s of the network coming back (Messenger had given up on its calls connection), and everything talks to Facebook within 75 s

**chats** (#24)
- the worker dies for good: Messenger shows its error screen, fMess reloads within 30 s, encrypted chats talk to Facebook again
- the same within 10 minutes of fMess's last reload for it: the screen stays (Messenger's own Reload can also clear and refetch the encrypted chats stored on the computer)

**sound** (decided as on Windows, plus macOS's own rule)
- after Messenger's sound played, the notification is quiet (one sound per message)
- no sound from Messenger (its 4th message in a chat, or its sound failed): it rings
- a notification asking for silence: quiet
- Do Not Disturb on: quiet, and Messenger quiet too; off again: rings
- Messenger's sounds off: quiet; on again: rings
- the Mac muted: Messenger's sound still counts as played, so quiet; without it, it rings (the system keeps it quiet)
- macOS: notifications stay silent (Messenger plays its own)
- each case checks the page's decision and what fMess's native side was told

## What Messenger and WebKit do on their own (measured 2026-11-05)

Messenger keeps four WebSockets to gateway.facebook.com (`/ws/lightspeed`
messages, `/ws/realtime`, `/ws/rpsignaling` calls, `/ws/streamcontroller`) and
the encrypted chats' worker keeps one to web-chat-e2ee.facebook.com.
- A silent death: it notices in 11–13 s and keeps trying; lightspeed and
  rpsignaling are back within 2 s of the network, but realtime and
  streamcontroller were left down (over 2.5 minutes, until a reload).
- Offline: it retries with growing gaps, lightspeed every ~75 s after a few
  minutes, so it came back 77 s after the network did; rpsignaling tries every
  second and gives up after ~110 s, for good.
- The dead socket of a silent death is never closed: the newest socket of each
  kind is the one in use.
- A dead worker is restarted by itself, unless it can't (the #24 screen).
- WebKit keeps sending HTTP requests to a host down its pooled connection after
  a silent death until it gives up on it, 44 s later; a request to a host the
  page doesn't use (m.facebook.com) goes out on a new one. 19-recover asks both.

## Focus check (#51)

`tests/app/focus-check.sh [fresh] [reload] [chats] [network]` (all four by
default; macOS, debug build; `network` needs `--features test-proxy`). About 8
minutes. With a chat open, real clicks and keys (see below) check that:

- a click into the chat list's search field, then typing, types there
- ⌘K from the message box, then typing, types in the search field
- dragging across a message's text selects it
- a reply started (double-click beside a message) and dropped (the bar's ×)
  leaves the message box its width (the Zalo look of 0.1.35 shrank it to its
  buttons)

| Case | State |
|---|---|
| fresh | a chat just opened |
| reload | the page reloaded at that chat (fMess's own reload, Ctrl+R) |
| chats | after opening four chats in a row |
| network | a minute of dead network, then back (fMess reloads by itself) |

It prints `PASS`/`FAIL` per case and, for a failure, every move of the
keyboard: each `focus()` call with where it came from (`user-script` is
fMess's own) and each change of address. The window stays behind what you work
in. It types "zq" and deletes it again; nothing is sent.

What it found: Facebook swaps `window.setInterval`/`clearInterval` for its
scheduler's once its scripts load, and that `clearInterval` can't stop a
timer started before. After a reload at a chat, 06-routes' "cursor in the
message box" timer ran for ever (every 50 ms), so no field could be typed in and
no text selected until fMess was quit. fMess's script now keeps the browser's
own timers (01-start). A timer started as the page loads is the case to test.

## Fuzz check

`tests/app/fuzz-check.sh [seed] [steps] [runs] [looks]` (macOS, debug build):
random sequences of what a person does, with real clicks and keys, checking
after every step what must always hold. Bombadil's idea (property-based UI
testing), on WebKit with this probe. Runs behind what you work in.

- Steps: open a chat, the search field (or a theme's own search box), ⌘K, the
  photo viewer (zoom, Escape), a reply started and dropped, right-click then
  Escape, selecting text, scrolling, a reload; and, only in whitelisted chats,
  typing, sending, replying, a draft left while doing something else.
- Rules: the keyboard goes where it is put; typing lands there; the message
  box keeps its width; Escape closes the viewer (box empty) and menus; text
  can be selected; a sent message shows and the box empties; no error from
  fMess's script; no "Couldn't load chats"; a chat opens within 20 s; no chat
  off the whitelist gets typed in or marked read.
- Looks: one per run in turn (messenger, zalo, whatsapp, teams, …), put in
  the page as Settings would, not saved.
- `PASS`/`FAIL seed … step n: rule: detail (picture)`, every step in app.log
  as it goes (`[fuzz]`). The same seed and look do the same steps again.
- It runs on anh's account: it may type and send only in chats listed in
  `~/.config/fmess/fuzz.json` (`{"sendTo": {"<thread id>": "who"},
  "maxSendsPerRun": 6}`), kept on that Mac, never in the repo. Without the
  file it sends nothing. Elsewhere: read chats only, no typing in a message
  box, no menu items, reactions or calls; a draft of anh's own is left as is.
- Most first failures were the probe's (a click under a panel, a caret in the
  middle of old text, text in a link): look at the picture before blaming
  fMess. What it found that was real: encrypted chats hung after an early
  reload (below).

## Real clicks and keys in a probe

Synthetic events (`dispatchEvent`, `el.focus()`) passed where people failed:
#51 never showed with them. A probe asks the app for real ones (probe.rs,
macOS): set `window.__fmessInputRequest` to steps in CSS pixels, then wait for
`window.__fmessInputDone` to count up.

```js
const input = async (steps) => {
  const done = window.__fmessInputDone || 0;
  window.__fmessInputRequest = steps;
  while ((window.__fmessInputDone || 0) === done) await new Promise((r) => setTimeout(r, 100));
};
await input([{ click: [x, y] }, { type: 'zq' }, { key: 'Escape' }]); // also Enter, Tab, Backspace
await input([{ drag: [x1, y1, x2, y2] }]);                       // selects text
await input([{ click: [x, y], count: 2 }]);                      // double click
await input([{ rightClick: [x, y] }]);                           // contextmenu
```

They go through the window when it is the key one, otherwise straight to the
web view (macOS won't let fMess come in front of the app you're in), so no
Accessibility permission and the pointer doesn't move. A probe that returns
`"again"` runs again `FMESS_PROBE_DELAY` seconds later, after a reload too:
keep where it is in `sessionStorage` (focus-check.js does).

## Encrypted chats hung after a reload

A reload in the page's first seconds (fuzz: 2 of 8 at ~3 s) can leave
Messenger's encrypted chats' worker hung: open encrypted chats spin with no
message, and Messenger logs `BridgeTimeoutError … WorkerHealthStatus:
shared_exists_and_connected`. A later reload brought them back in the run
measured; anh's morning one needed fMess quit. 19-recover: no reload in a
page's first 30 s; a reload when Messenger reports that hang twice in two
minutes, when an open encrypted chat spins 60 s with no message, or when
every encrypted chat in the list stays blank 90 s; stuck again within 10
minutes of that reload, fMess starts again (recover.rs, at most every 30
minutes, kept in `last-chats-restart` in the app's data folder; delete it
after testing a restart). Turning SharedWorker off on macOS (as #67 does on
Windows) broke the encrypted chats after a reload: don't.

## Dark wakes (0.1.36)

A Mac with its lid shut wakes in the dark for a few seconds every 15–20
minutes (`pmset -g log | grep DarkWake`) and sleeps again. 0.1.36 reloaded
there, the reload was cut short by the next sleep, and the encrypted chats
stayed on grey bars and a spinner, with no error screen, until fMess was
quit. 19-recover now waits until the computer has been awake 20 s, and
reloads when every encrypted chat in the list stays blank for 90 s.
`tests/recover.mjs` has both; a real dark wake can't be made up here, so
when someone reports a stuck morning, read `pmset -g log` around the time
the page's process started (`ps -o lstart -p <WebContent pid>`).

## Still by hand

- Windows: the real toast sound (a `test build` labelled PR; install it in the
  Windows VM). Settings → Send Test Notification rings; one message rings once;
  Do Not Disturb or Messenger's sounds off: silent.
- A real night's sleep (the morning after: send a message, it goes).
- A muted group: Messenger doesn't notify for it, and fMess's own notification
  leaves muted chats out (14-notifications, unread chats); needs someone to write
  in a muted group.
- Messages waiting at "Sending" go out after a reload: needs a chat you may send
  test messages to.

## Adding a case

Make it up at the proxy or with Messenger's own calls, never with a real chat
someone else sees. A probe that gets the page reloaded loses its result: log
from the page with `window.__fmessErrors?.push('<what>: … @' + Date.now())` and
judge in `live-check.sh` with `judge`/`at`/`reloads`/`back`, or return
`"again"` and pick up after the reload (see focus-check.js). Click and type
with real events, not the page's own (above).

© toanbku · MIT · toanbku/fmess

Похожие скиллы

Каталог скиллов SKILL.md →
live-check NahumLitvin Скилл live-check для Claude Code проверяет, как интерфейс отображает ответы после визуальных изменений. Он запускает рабочую копию Prismantis и показывает её вывод в рендере. CX Codex +1 ★ 482звёзд репо developer-live-demo-design Samuel Berthe Скилл developer-live-demo-design для Claude Code проектирует надёжные живые технические демоверсии для выступлений. Превращает сценарий показа в подробный ранбук с точками отката и фразами на случай сбоя. CX Codex +1 13 400установок gemini-live-api-dev официальный google-gemini Скилл gemini-live-api-dev обучает агента собирать потоковые голосовые и видеоприложения через Gemini Live API поверх WebSockets. Помогает настроить сессии на Gemini 3.8 и перевести проект со старых версий API. CX Codex +1 11 300установок build-live-game Unity-Technologies Скилл build-live-game ведёт агента по сборке серверной части живой игры на Unity Gaming Services: вход игроков, прогресс, конфиги, лидерборды. CX Codex +1 7 300установок gemini-live-api официальный google Скилл gemini-live-api для Claude Code собирает WebSocket-клиенты для Gemini Live API и помогает разбирать сбои двусторонних сессий. CX Codex +1 5 200установок chrome-cdp-live-browser reason-machines Скилл chrome-cdp-live-browser подключает агента к уже открытому Chrome по Chrome DevTools Protocol. Работает с вашими вкладками, логинами и текущим состоянием страницы. CX Codex +1 1 500установок

Из того же набора3

Спонсоры проекта

Компании, которые помогают держать аналитику открытой