fix: restore dropped IPC handlers and fix several main-process bugs

[devoliverluccas]

Six main-process bugs, three of which have open issues that turn out to have the same root cause: an IPC handler that is not registered at runtime.

1. Autostart and badge IPC handlers were never registered

native/autoLaunch.ts and native/badges.ts register their handlers as an import side effect, and nothing imports either module.

For autostart this is a regression in 1.5.0:

• e00f3a8 (fix: Do not enable autostart on first launch #237) removed the last direct use of autoLaunch, leaving the import unused
• 6894231 (refactor: add mise, update electron-forge, add ci/cd workflow for lint #252) then removed the now-unused import during the lint pass

ipcMain.handle("getAutostart"/"setAutostart") went with it, so desktopConfig.setAutostart() rejects with "No handler registered" and the settings toggle silently does nothing — matching #287 exactly ("I had removed the option for Stoat to auto start in the options menu. However it still auto started. I verified that the windows startup option was still ticked").

Badges were never wired up at all: git log -S 'native/badges' shows the module has not been imported since the initial commit, even though the preload bridge has exposed native.setBadgeCount() since #25.

Verifiable on the current main:

$ pnpm package && grep -c 'getAutostart\|setBadgeCount' .vite/build/main.js
0

Both modules now expose an explicit init*() called from main.ts, so registration is tied to a call site the linter cannot quietly drop.

Closes #287. Refs #213 — this only restores the plumbing; the UX described in that issue still needs client-side work.

2. Discord RPC (refs #203)

• Turning it on did nothing. The discordRpc setter ran initDiscordRpc() before writing the new value to the store, and initDiscordRpc reads that setting back out of the store — so it observed the stale value and took the if (!config.discordRpc) return early exit. The setters now persist first and act second.
• Turning it off left the old client running. initDiscordRpc only called removeAllListeners() on the previous client and never destroyed it, so every toggle stacked another client that stayed logged in to Discord and kept broadcasting the activity.
• login() is asynchronous, so its rejection escaped the try/catch as an unhandled rejection and never triggered a retry.
• destroy() rejects when the transport never connected, which is now swallowed rather than surfacing as an unhandled rejection.
• disconnected only schedules a reconnect for the client that is still current, and reconnects no longer stack.

3. Spellchecker not applied on launch (refs #69)

webPreferences.spellcheck was hardcoded to true and setSpellCheckerEnabled only ever ran from the setter, so the stored preference was never applied at startup — disabling it silently reverted on the next launch.

4. Screen sharing breaks after a cancelled picker (refs #267, #263)

screenPickerCallback is registered with ipcMain.once, so a picker the user dismissed without choosing a source leaves it attached forever. The next screen share request registers a second listener behind the stale one, which then consumes the reply and hands it to a request that is already dead — leaving screen sharing broken until the app restarts. Stale listeners are now cleared before registering, mirroring what the renderer side already does in world/window.ts.

Alongside it:

• the bounds check rejected idx < 0 || idx > sources.length, so idx === sources.length passed and sources[idx] handed undefined to the callback as if it were a real source
• getSources() had no rejection path, so a failure left the renderer waiting on a callback that was never invoked
• NativeImage.resize() returns a new image instead of mutating in place, so discarding the result sent full-size icons over IPC

5. Window restored onto a disconnected display (refs #176)

The saved position was restored without checking that it still lands on a connected display, so unplugging or rearranging a monitor left the window opening off screen with no way to reach it. It now falls back to centring the window.

The guard also changes from x > 0 || y > 0 to x !== 0 || y !== 0, since negative coordinates are legitimate for displays positioned left of or above the primary one.

---

Testing

mise lint, mise format and mise build all pass.

I have not exercised these interactively — no manual autostart, screen share or RPC run — so runtime confirmation from someone on each platform would be welcome, particularly the screen picker path. The autostart and badge handlers are confirmed by diffing the packaged bundle before and after (0 → 1 occurrences of each handler, plus the 12 badge assets now resolved).

🤖 Generated with Claude Code

[Dadadah]

for-desktop follows the same contribution guidelines as for-web. We do not accept contributions made by LLMs.