feat(settings): Ctrl/Cmd + and - step the window zoom

[Adam-Dalloul]

Settings → Window zoom is a discrete menu. This binds the usual keys to those same rungs:

• Ctrl/Cmd + or = zooms in one step
• Ctrl/Cmd - zooms out one step
• Ctrl/Cmd 0 resets to 100%

Rungs are 80 / 90 / 100 / 110 / 125 / 150 / 175 / 200 / 250 / 300. Capture-phase so the webview does not eat the keys as page zoom. Same persist path as the Settings menu.

Also listed under Settings → Shortcuts.

[xintaofei]

Thanks for this — ⌘/Ctrl + - 0 is a genuinely missing affordance, and I like that you reused the existing discrete rungs instead of inventing a second, continuous zoom model. Putting the listener in AppearanceProvider is also the right call: it's mounted from the root layout, so it covers the settings and pet windows and web mode, whereas WorkspaceChromeController would have missed them. Extracting stepZoom with a unit test and remembering to keep VALID_ZOOMS in the pre-paint inline script in sync are both nice touches — that second one is easy to miss.

I ran the branch: tsc --noEmit, ESLint and Prettier are clean, and the full suite is green (313 files / 4198 tests). Hygiene is good.

There's one thing I'd like fixed before merging, plus a few smaller ones. Details below.

---

Blocking

1. "Zoom in" / "Zoom out" show up as rebindable in Settings, but nothing reads the binding

zoom_in / zoom_out / zoom_reset are added to SHORTCUT_DEFINITIONS and DEFAULT_SHORTCUTS, so ShortcutSettings renders three rows with the full record / conflict-check / reset-to-default UI. But at runtime only zoom_reset is actually consulted (appearance-provider.tsx:754, :782) — zoom in/out go through the hardcoded isZoomInShortcutEvent / isZoomOutShortcutEvent, which take no binding argument (appearance-provider.tsx:772, :777).

So if a user rebinds "Zoom in" to, say, ⌘⇧Z: the value is normalized, persisted, the button label updates and a success toast fires — and then nothing happens when they press it, while ⌘= keeps zooming. Worse, rebinding frees mod+= in the settings conflict checker, so they can now legitimately assign ⌘= to another action and get both behaviors (the zoom handler only calls preventDefault(), not stopPropagation()).

Two ways out, either is fine:

• Wire them up: match against shortcuts.zoom_in / shortcuts.zoom_out like every other action. If you want to keep the shift/numpad tolerance, put it in the matcher for the bound key (treat =↔+ and -↔_ as the same physical key) rather than in a separate predicate that bypasses the binding.
• Or make them honest: drop zoom_in / zoom_out from SHORTCUT_DEFINITIONS, keep them fixed, and just document them. zoom_reset is already wired correctly and can stay.

2. if (event.repeat) return runs before the match, so held keys aren't prevented

appearance-provider.tsx:771. Two consequences:

• Holding ⌘- doesn't repeat, so it's 7 taps to walk 100% → 300%. Every other app repeats here.
• In server/browser mode (codeg-server / Docker / remote access) the un-prevented repeat events reach the browser's own zoom, which then compounds on top of the app's rem zoom.

Moving the repeat guard after the match — and still calling preventDefault() — fixes both. Or just allow repeats.

3. The terminal loses ⌃_ (readline/emacs undo)

The listener is a bare window capture handler with no target check, and mod in this codebase means Ctrl or Cmd (hasMod = event.metaKey || event.ctrlKey). isZoomOutShortcutEvent also deliberately ignores shiftKey, so ⌃⇧- (= ⌃_) matches too.

Net effect with the integrated terminal focused — on macOS as well as Linux/Windows — pressing ⌃_ sends undo to readline and steps the app zoom. Both fire.

terminal-context.tsx already has the pattern for this (isInTerminalRegion, terminal-context.tsx:304); reusing it here would keep terminal muscle memory intact. I know ⌃K-vs-toggle_search has the same shape today, so this isn't a regression you introduced — but zoom keys collide with a lot more shell muscle memory than ⌘K does, so I'd rather not widen it.

4. Recording any shortcut in Settings also zooms the UI

The recorder in shortcut-settings.tsx:59 calls event.stopPropagation(), but both it and the zoom handler are capture listeners on the same window node — stopPropagation() doesn't stop sibling listeners on the same target. So while the "Record" button is armed and the user presses ⌘= or ⌘-, the whole settings window zooms underneath them.

(And per #5 below, ⌘⇧= is then rejected as invalid anyway, so they get a zoom plus an error toast.)

5. + can't actually be recorded

shortcutFromKeyboardEvent joins parts with "+", so ⌘⇧= serializes to "mod+shift++"; normalizeShortcut splits on "+" and returns null, so updateShortcut fails and the user gets "invalid shortcut". normalizeShortcut("mod++") === null — I checked.

This one predates the PR, but the PR is what puts a +-shaped row in front of users, so it becomes reachable here.

6. i18n: the 9 non-English locales got English strings

zoom_in / zoom_out / zoom_reset titles and descriptions are English in ar, de, es, fr, ja, ko, pt, zh-CN and zh-TW. The convention here is that new UI strings ship translated in all ten (cf. toggle_custom_style in 4e54708).

Separately, Appearance.zoomLevel.sectionDescription was only extended in en.json, so the shortcut hint is invisible in every other language. Two copy notes on the English while you're in there: it says "Ctrl" only, and on macOS it's ⌘; and "one Settings rung" is a bit opaque as user-facing text.

---

Non-blocking

• Split the 300% commit. feat(settings): raise window zoom to 300% widens ZOOM_LEVELS from 6 rungs to 10 and is an independent product decision from the keybinding. I checked and it's less risky than I first assumed — Monaco (file-workspace-panel.tsx:2291) and xterm (terminal-view.tsx:456) both scale their font size off zoomLevel, and the Rust side already clamps 50–300 — but I'd still rather see it land on its own with a screenshot at 300%.
• The event.code === "Minus" / "Equal" fallback misfires on non-US layouts. { key: ")", code: "Minus", ctrlKey: true } returns zoom-out (that's the AZERTY shape). The event.key branch already covers the real - / + / = on those layouts, so I'd narrow the code fallback to the numpad keys or drop it.
• matchShortcutEvent(event, "mod++") in isZoomInShortcutEvent is dead — normalizeShortcut("mod++") is null, so it's always false. Harmless (the fallback catches +), but it means the ⌃⇧+ test is passing through a different branch than it looks like.
• The three new SPECIAL_KEY_ALIASES (keyboard-shortcuts.ts:135-137) are inert on the webviews we ship on: normalizeKeyToken returns early for single characters, and Chromium/WebKit report "+" / "-" / "=", never "Add" / "Subtract" / "Equal" ("Equal" is a code, never a key). And normalizeShortcut("mod+add") → "mod++", which the parser then rejects — so the round-trip is broken. I'd drop them, or fix them along with feat: Markdown 预览模式（文件面板） #5.
• zoomLevelRef.current = zoomLevel in the render body (appearance-provider.tsx:481). No demonstrated bug, but terminal-view.tsx:449 does the same sync inside an effect — worth matching.
• No isComposing guard on the global listener, so an IME's own Ctrl/Cmd +/- commands can be intercepted. Probably falls out of whatever target policy 如何让ai写的代码符合规范 #3 gets.
• Tests. keyboard-shortcuts.test.ts:134 and :138 are the same assertion twice. More useful coverage would be: the configurable path (rebind → new binding fires, old one doesn't), repeat events, and one non-US-layout event.
• Docs. The canonical shortcut table lives in the docs site (guide/workspace.md and its zh/ twin), so the three new rows want a companion PR there.

---

None of this is structural — the shape of the change is right, it's mostly about making the Settings rows tell the truth and picking a target policy for the keys. Happy to look again once #1 is sorted; the rest can come along with it or as follow-ups if you'd prefer. Thanks again for taking this on! 🙏

[Adam-Dalloul]

Thanks for catching that. The Settings rows for zoom in/out weren't actually wired, which is on me.

Follow-up is up. Those bindings drive the listener now, and = / + and - / _ still count as the same key. Held keys repeat and still preventDefault so the browser doesn't also zoom. Terminal keeps Ctrl+_, recording a shortcut no longer zooms the settings window, and + can actually be recorded. Strings are translated, and the zoom hint no longer says Ctrl-only.

Also dropped the dead Add/Subtract/Equal aliases and limited the code fallback to the numpad so AZERTY Ctrl+) doesn't zoom out.

300% is still its own commit here. Happy to split it if you'd rather land that separately.