Best-practice guide for authoring a Muxy extension — how it should look and behave so it reads as a native part of the app. Covers theming (follow the theme, never hardcode colors), the sizing scale, and which surface to use. Mechanics (manifest fields, permissions, the window.muxy API) live in the linked docs.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Best-practice guide for authoring a Muxy extension — how it should look and behave so it reads as a native part of the app. Covers theming (follow the theme, never hardcode colors), the sizing scale, and which surface to use. Mechanics (manifest fields, permissions, the window.muxy API) live in the linked docs.
Muxy Extension Guide
A Muxy extension is an npm + Vite project: source under src/, an entry HTML that vite build emits into dist/, and Muxy reads dist/ when present (otherwise the project folder). The manifest is the "muxy" object in package.json. There is no fixed folder layout — every entry/background/icon path is an arbitrary relative path inside the build output (the vanilla starter kit emits its panel to panel/index.html); package.json and dist/ are the only names Muxy fixes. During development you don't copy into the config folder — Load Unpacked in the Extensions modal points Muxy at any folder (your git checkout is the install).
The build script must copy package.json into dist/. The publish pipeline ships onlydist/, and the app reads the manifest from the install root — so the manifest has to be inside the build output. vite build alone emits your entry/asset paths but not the manifest, so use "build": "vite build && node scripts/copy-manifest.mjs" where copy-manifest.mjs copies package.json into dist/. Easy to miss because Load Unpacked falls back to the root package.json in dev, so it loads locally but fails validation/install when published. The vanilla starter kit already wires this up.
This skill is the guidance layer — how an extension should look and behave. For the API and manifest mechanics (every field, the permission strings, the full window.muxy surface, events, scripts), read the reference docs. Start from the LLM-friendly index, which lists every page and links to its raw Markdown source:
The goal of everything below: an extension should be indistinguishable from a native Muxy surface. Match the theme and match the scale, and it will be.
Pick the right surface
A global full-window destination outside project tab restoration → a homeView. Use it for an overview or launch surface that spans projects and worktrees; see Home Views for current presentation availability.
Showing something inside a project workspace or transient app chrome → a UI page (tab, panel, or popover). Page scripts get the full window.muxy API.
A persistent, full-height navigation or control surface that replaces the built-in left sidebar → a sidebar (one per extension; the user selects it in Settings → Sidebar). It fills the entire region — the project list and the footer — so own your own navigation. Same window.muxy API and theme variables as a panel.
Reacting durably to events, coordinating multiple webviews, or running shell commands headlessly → a background.js script. It can also call muxy.tabs.open to show a result in the active workspace. Most extensions don't need one.
One-shot logic from the palette → a runScript command, not a hidden tab. Its muxy.* calls are synchronous except for the Promise-based muxy.execAsync. It has tabs/panes/projects/worktrees/browser/agents/files/git/exec/execAsync/dialog/modal/topbar/statusbar/notifications, but nothttp, events, remote, panels, or popover. It can open a modal and act on the choice inline — no page or background listener needed.
Your own HTML in a modal on a keypress (a form, info, a list, or mixed — not just forms) → a webview modal (muxy.modal.openWebview). A top-centered omnibox-style overlay rendering your HTML. Reach for it over the native muxy.modal.open picker when a list won't express what you need. — it's always running, so a shortcut works with nothing else on screen, and you can pass dynamic and the result. Use the declarative command action only for a static, self-contained modal that needs no result and no . The modal returns a value via only if the opener wants one; an informational modal just calls .
Don't open a hidden tab to run logic, and don't put durable event-driven work in tab JS where closing the tab loses it. Use muxy.events.emit('extension.<name>', payload) plus a background listener when a webview needs to ask background.js for shared or long-lived work.
Theme — follow it, never hardcode
Muxy ships paired light/dark themes and a user-chosen accent. Every extension webview inherits CSS custom properties on document.documentElement that track the live theme and update automatically when the user switches it.
Rules:
No hex literals for chrome. Use var(--muxy-…) for every color. The only exception is decorative art meant to be theme-independent.
The variables already invert for light/dark — never sniff the color scheme to pick a color. Only branch on muxy.theme.colorScheme for things a variable can't express (e.g. swapping a logo image).
--muxy-accent is the only saturated color. Use it sparingly — primary action, focus ring, one key number — so it stays distinctive. Text on an accent fill must use --muxy-accent-foreground, which is resolved for contrast against the active theme's accent.
Use --muxy-surface-solid for component backgrounds. It is the native surface overlay precomposited over the active theme background, so cards, inputs, code blocks, and buttons match Muxy without letting content show through. --muxy-surface, --muxy-border, --muxy-hover, and --muxy-accent-soft are translucent overlays for material-relative effects. Do not apply additional opacity to these tokens or to an entire control.
Re-read the theme for JS-drawn color. Canvas/SVG that doesn't pick up CSS variables must redraw in muxy.onThemeChange(theme => …).
Popovers leave the body transparent (body { background: transparent; }) — they sit over native macOS popover material that is already light/dark-aware. Tabs and panels do paint --muxy-background on the body.
The variables (the complete injected set):
Variable
Use for
--muxy-background
Page background
--muxy-foreground
Primary text
--muxy-foreground-muted
Secondary text, labels, captions
--muxy-surface-solid
Opaque cards, inputs, code blocks, buttons
--muxy-surface
Translucent surface overlay for popovers and intentional layering
(muxy.theme.colorScheme gives "light"/"dark" in JS; there is no --muxy-color-scheme CSS var.)
Sizing — match the app's scale
Muxy's native views are built from one scale of values, and all of them scale with the user's interface-scale setting (Settings → Interface). Pick from this scale rather than inventing numbers, so your surface tracks scale changes the way native views do. These are the base (100%) values in px:
Spacing (padding, gap, margin) — 2 · 4 · 6 · 8 · 10 · 12 · 16 · 20 · 24 · 32. No in-between values. Panel rows and content pad 10px left/right; an icon-and-label gap is 8px; adjacent icon buttons sit 4px apart.
Font sizes — 10 caption · 11 footnote/section labels (often uppercased) · 12 body (paths, row text) · 13 controls · 14 titles (weight 600) · 16+ headings. Body is 12, not 13. Use the system font for UI; "SF Mono", Menlo, monospace for code, counts, and hashes.
Icons — 12–14px glyphs at weight 600 (a thinner default weight is the most common reason an extension's icons look foreign). Custom SVG strokes are 1.5px, round caps/joins.
Controls — an icon button is a 24×24 hit target wrapping a 13–14px glyph; text buttons are 28px tall with 10px horizontal padding.
Radii — 4 chips/badges · 6 buttons/inputs · 8 cards/panels · 10 large containers. Buttons are 4–6, not 5.
Topbar height is the exception — never hardcode it. It scales with interface scale and is injected pre-scaled as --muxy-topbar-height. A tab fills its whole pane, so render your own topbar to match native surfaces (so sibling panes line up): use that variable for the height and keep box-sizing: content-box so the 1px border-bottom lands on the same line as native tabs. Split-child tabs appear as bordered panes inside their owner and do not have a local tab strip. Omit the topbar for edge-to-edge content.
Declare the scale once at the top of your stylesheet and reference it everywhere, so there are no stray magic numbers:
Least privilege. Declare a permission only when you add the call that needs it.
Panel float/dock preferences are per panel.mode sets a panel's initial mode, then Muxy persists the user's choice independently for that extension and panel ID while the native pin control is available. Hiding pin with hiddenControls, or all native chrome with hideTopbar, keeps the panel at its declared mode; do that only when the extension intentionally owns the behavior.
Workspaces can be remote. When the active workspace is a remote (SSH) workspace, muxy.exec, muxy.execAsync, muxy.git.*, and worktree work run on the remote server with the selected SSH device's environment, and paths are remote paths. Write extensions against the active workspace, not a hardcoded local machine — the same code works for local and remote because Muxy brokers the SSH connection. See Scripts.
Use muxy.execAsync for cancellable long-running runScript commands — it returns { id, result, cancel() }; result resolves to the same shape as muxy.exec, and cancellation rejects with error.code === "cancelled" / error.cancelled === true. Local cancellation terminates the command process group; remote cancellation closes the SSH command channel, so detached remote processes may outlive the job. Keep muxy.exec for short synchronous calls. An extension may run at most 32 commands concurrently across both APIs — beyond that, starting a command fails. See Scripts.
Use muxy.git for repository work (status, diff incl. { raw: true }, repoInfo, log, branches, PRs incl. pr.number/pr.diff, tags, init, checkout/cherryPick/revert, branch delete/deleteRemote, worktrees incl. worktree.switchTo and pr.checkoutWorktree) instead of shelling out via muxy.exec — it's the app's own git core, returns structured data, and caches reads. Reads need git:read; writes need git:write and prompt for consent. Reads are cached per project/worktree (HEAD/index aware); pass to bypass. Available to tabs, panels, popovers, commands, and background scripts. See .
Checklist
Every color is var(--muxy-…); muxy.onThemeChange wired for any JS-drawn color.
Component backgrounds use --muxy-surface-solid; translucent overlay tokens are only used intentionally.
Spacing, font, icon, control, and radius values come from the scale above — no off-ramp numbers (rows pad 10px, body is 12px, icons 12–14px at weight 600).
Tab topbar uses --muxy-topbar-height with box-sizing: content-box.
Hover/active states are visible in both themes.
tooltip is the accessible name of an icon-only topbar control (title bar and rail).
permissions declares only what is used.
Durable event-driven work is in background.js, not tab JS. Webview coordination uses extension.* events. No background script unless events, shared state, or background exec are needed.
build copies package.json into dist/ (e.g. vite build && node scripts/copy-manifest.mjs) — only dist/ ships, so the manifest must be inside it.
Every declared localization bundle is copied into dist/, contains only resources, includes the catalog for its declared language, and preserves each key's format placeholders.
Built with npm run build, then Reload in the Extensions modal (a Reload alone won't pick up unbuilt source).
Get the signed-in GitHub user with muxy.gh.user() instead of shelling out to gh api user — it returns { login, name, avatarUrl } from the local gh CLI login, needs gh:read, and caches for five minutes so a header/badge can call it on every render. The account is global to the gh login, so no { project } argument; it always reflects the local login, even on a remote (SSH) workspace. See GitHub.
Manage the project list with muxy.projects write verbs instead of editing projects.json — add(path) registers an existing folder as a project, makes it the active project, and returns its project id, rename(identifier, name), setColor/setIcon(identifier, value), setLogo(identifier, storedLogoFilename) (pass null to clear), and reorder(identifiers) (all local non-home project ids in the new order, each exactly once). All need projects:write and mutate Muxy's live project store, so the native sidebar updates immediately. add only accepts an existing directory; create the folder first via muxy.files/muxy.exec if needed. Subscribe to the projects.changed event (declare events: ["projects.changed"], grant projects:read) to notify a webview/sidebar that can refetch via muxy.projects.list() after any change, whether made by your extension or Muxy itself. The home project cannot be renamed, recolored, re-iconed, or reordered. See Permissions.
muxy.projects.create(path, { createIfMissing, name, workspace }) creates or opens a project in one call — unlike add, it can create the directory (createIfMissing: true, which prompts the user for runtime consent before writing to disk, like muxy.files.mkdir), optionally rename it, and optionally place it directly into a workspace by name or id. It rejects if the given workspace does not exist or is a remote SSH workspace, and never resets worktreesEnabled on a project that already existed. muxy.projects.attach(identifier, workspace) moves an existing project into a workspace (same workspace rules as create); muxy.projects.detach(identifier) removes a project from all workspaces. All three need projects:write. attach and detach accept local and device-backed remote projects, but reject the home project and projects owned by an SSH workspace. See Permissions.
muxy.projects.delete(identifier) deletes a project and is irreversible — it cleans up the project's worktrees, branches, and directories on disk. It needs the projects:delete permission (separate from projects:write) and prompts the user for confirmation on every call. The home project cannot be deleted. Reserve it for explicit user-driven actions, never silent cleanup. See Permissions.
Manage workspaces with muxy.workspaces — list() returns each workspace's id, name, project count, and whether it's active; create(name) adds one and makes it active; switchTo(identifier), rename(identifier, name), and delete(identifier) accept a name or id. delete fails while the workspace still has projects in it — detach or move them first with muxy.projects.detach/attach. list needs projects:read; create/switchTo/rename/delete need projects:write. See Permissions.
Use muxy.files for workspace filesystem work (list, read, stat, write, mkdir, rename, move, delete) instead of muxy.exec — paths are sandboxed to the active worktree root and returned relative to it. The root is valid for list/stat and as a move destination, but cannot itself be mutated. Reads and UTF-8 writes are capped at 5 MiB. Rename rejects collisions; move uniquifies them. Reads need files:read; writes need files:write and prompt for consent. Pair with the file.changed event to stay reactive (e.g. a file tree). See Files.
Persist your own state with muxy.storage instead of shelling out to a config file — set(key, value) / get(key) (any JSON value; get returns null when absent) / delete(key) / keys(). Storage is isolated per extension and shared across that extension's surfaces (a panel and its background.js see the same keys), and survives restarts. Needs storage:read (get/keys) / storage:write (set/delete); a key is ≤256 chars, a value ≤1 MB. Good for layout/collapse/preferences. See Storage.
Ask for a value or a folder with muxy.dialog.prompt / muxy.dialog.pickFolder instead of osascript — prompt({ title, message, default?, placeholder?, confirm?, cancel? }) resolves the entered string (or null), pickFolder({ title?, message?, default? }) resolves an absolute path (or null). Same surfaces and no-permission rule as confirm/alert. See Dialogs.
Surface copyable diagnostics with muxy.dialog.alert. Alerts with a non-empty message include a Copy button and Command-C shortcut that copy the complete message, including content beyond the 2000 characters shown in the sheet. See Dialogs.
React to branch changes with the worktree.headChanged event instead of polling git.worktrees() — it fires when a worktree's checked-out branch changes (e.g. a git checkout in a terminal), with the new branch and worktree path. Declare events: ["worktree.headChanged"] and grant worktrees:read. See Events.
Give large worktree removals an explicit budget.muxy.git.worktree.remove({ path, force?, timeoutMs? }) defaults to 30 seconds, allows up to one hour, and returns { path, dirRemoved }, where path is the canonical physical path resolved by Muxy rather than necessarily the caller-supplied representation. The budget covers path resolution, teardown, Git removal, stale-directory cleanup, and final verification. Extension-triggered removal runs only pre-authorized per-machine hooks from $XDG_CONFIG_HOME/muxy/worktree.json when XDG_CONFIG_HOME is non-empty or ~/.config/muxy/worktree.json otherwise. Project hooks from <project>/.muxy/worktree.json require the native confirmation that displays every command, so this API skips them. A failed per-machine hook or invalid per-machine configuration blocks removal. Remote and external worktrees do not run local hooks. With force: true, Muxy can remove a tracked Muxy-created checkout from its own default app-managed directory when a stale .git reference proves that it belonged to the selected repository. If ownership cannot be verified, Muxy forgets the stale entry but preserves the directory and reports dirRemoved: false; custom, external, cross-project, foreign-repository, and symlinked paths remain protected. Treat dirRemoved: false as a residual-directory warning and dirRemoved: null as an unknown result because verification exhausted the deadline.
React to AI agent activity with the agent.status event instead of polling — it reports an agent's lifecycle per worktree (working > waiting > idle, with providerID and the owning paneID), driven by the provider's hooks (Claude Code, Cursor, Codex, Droid, Grok, OpenCode, Pi). idle covers a finished, cancelled, or ended turn; Muxy's native finished indicator is separate UI state. Coverage depends on what each CLI's hooks expose — Pi and Cursor do not report waiting; the other providers report all three states. It fires only when a worktree's status changes and turns idle when the last agent pane closes. Declare events: ["agent.status"] and grant agents:read (both the event subscription and muxy.agents.list() need it); pair the event with muxy.agents.list() to hydrate current statuses on load. Good for a live per-worktree indicator. See Events.
React to an entirely sleeping worktree with the worktree.offline event when a resource belongs to the whole worktree (a dev server, a container, a watcher) — it reports offline: "true" only when every terminal pane in that worktree is offline, and offline: "false" again when any pane wakes, a new terminal pane is created there, or the last terminal pane closes, so every suspend is paired with a resume. It carries projectID, worktreeID, worktreePath, and offline, so a background.js listener can act on the directory directly (background scripts have no muxy.worktrees.list()). Declare events: ["worktree.offline"]; no permission is required. See Events.
Use muxy.http.fetch to call external APIs from a tab/panel/popover instead of the webview's fetch() — the request goes out via native code, so it is not CORS-blocked, and a panel needs no background.js (no subprocess) just to reach the network. Pass (url, { method?, headers?, body?, timeoutMs? }) and await{ status, headers, body, truncated }. No manifest permission; the first call to a host prompts for consent, "Allow & remember" whitelists that host. Private/loopback hosts (localhost, 127.*, 192.168.*, 169.254.*, .local, …) are blocked. muxy.http is a webview-only surface — neither background scripts nor runScript commands have fetch; they shell out via muxy.exec(['curl', …]). See HTTP.
Use muxy.modal.open for a list picker (the native searchable picker overlay) instead of building your own — pass { items: [{ id, title, subtitle? }], placeholder?, onSelect(choice) }; the choice (or null if dismissed) arrives in onSelect. Muxy owns the search, navigation, and open/close. No permission needed. Available on every surface: on runScript/background modal.open returns immediately and you read the result in onSelect; on webview pages you can also await it. It has no shortcut of its own: bind a palette command with a defaultShortcut (its action can be the runScript that opens the modal, or an event a background.js listener reacts to). Pass searchToolbar: true only when the picker should show the footer search option toggles (Aa, W, .*). For large lists (a file picker over a big repo), pass items as a function items(emit) instead of an array — the picker opens instantly and you stream rows with emit(batch) while Muxy filters them natively, so typing never calls back into your code and the UI can't hang. For results that depend on the query (server-side/async search), pass onQuery(query, emit) — Muxy debounces the field and calls it per query so you supply a fresh list, dropping responses for superseded queries; native filtering still runs on top. See Modal.
Bind keyboard shortcuts to your extension — for a static binding declare a palette command with a defaultShortcut ("cmd+shift+e"); for a runtime one (e.g. configurable in your settings) call muxy.shortcuts.register({ id, combo }) from background.js and subscribe to the same command.<id> event. register returns { ok, conflict? }; unregister(id) and list() round it out. Runtime shortcuts need shortcuts:register, are not persisted (re-register on launch), and reject an id that collides with a manifest command. See Palette Commands.
Build your own session restore from background.js — Muxy has no built-in session restore, so an extension owns it. Record sessions by subscribing to the enriched tab.* events (which now carry kind/projectID/worktreeID/areaID/cwd/data), then recreate each terminal with muxy.tabs.open({ kind: 'terminal', directory, command }), which resolves the new tab's id (for extensionWebView tabs, the instance id usable with setTitle/setIcon). directory stays inside the worktree root; command adds a one-time runtime consent on top of tabs:write. See Events and Tabs.
Use extension.* events for webview ↔ background communication — pages and background scripts can muxy.events.subscribe('extension.<name>', handler) and muxy.events.emit('extension.<name>', payload). These events are same-extension only, need no permission, and are not listed in the manifest events array. A webview emit is relayed through the extension's background.js, so it rejects when no background script is running — webviews can't reach each other directly. Workspace events (pane.*, file.changed, etc.) still require manifest events.
Update bar items live with muxy.topbar.set / muxy.statusbar.set — pass { id, icon?, visible? } (topbar) or { id, icon?, text?, visible? } (statusbar) from background.js or any page to swap the icon/text or show/hide without reloading; text: null clears back to the manifest value. Only togglePanel items may appear on the rail; the rail sits under the title bar to the window bottom. Keep muxy.topbar.*, and treat tooltip as the accessible name of an icon-only topbar control (title bar and rail). Decide visibility at runtime: declare the item with "visible": false and call muxy.topbar.show(id) / .hide(id) (or muxy.statusbar.show(id) / .hide(id)) when it applies. The item must be declared in topbarItems / statusBarItems; needs panels:write. Good for live indicators (e.g. a PR badge that only appears inside a repo). See Topbar / Status bar.
Retitle a tab live with muxy.tabs.setTitle(title) / muxy.tabs.setIcon(icon) from the tab's own page to reflect changing state (e.g. an editor showing the open file). icon is "<sf-symbol>", { symbol }, or { svg }; setTitle("") / setIcon(null) reset to the manifest defaults. Needs tabs:write; runtime-only (resets on restart, so set it again on load). See Tabs.
Autofocus your input on muxy.onFocus from a tab/panel/popover page so the surface behaves like a native one — when its tab is opened or switched back to, move keyboard focus into your editor or search field. The callback receives true on focus gained, false on lost; muxy.focused reads the current state. No permission, no manifest field. See Tabs.
Register project file handlers with fileOpeners. Users choose extension-provided openers only in Settings; they never appear in the topbar project-target control. If an opener becomes unavailable, Muxy preserves the selection and resumes using it when the extension is available again. The referenced tab receives source: "terminal", the project-relative filePath, and optional line/column through muxy.data or muxy.onDataChange when its singleton tab is reused. See Tabs.
Provide app translations with localizations. Keep only the provider metadata in package.json; put the full translation catalog in a resource-only .bundle copied into dist/. The bundle must not declare CFBundleExecutable and needs <language>.lproj/Localizable.strings or Localizable.stringsdict. Every translated value must keep the format placeholders of its key at the same argument positions — use %1$@/%2$lld to reorder — or Muxy rejects the bundle. Categorize published providers as localization so users can discover them from Settings → Interface → Language → Browse Language Extensions…. After installation and enablement, their providers appear automatically in the language picker; a missing provider temporarily falls back to built-in English while preserving the selection. See Localizations.
Guard a close with muxy.lifecycle.onBeforeClose(handler) from a tab/panel/popover page when closing could lose work (a dirty editor). Return/resolve true (or { prevent: true }) to prevent the close, anything else to allow it; the handler may be async, so await muxy.dialog.confirm(...) and decide. Closing a top-level tab asks its own and every split-child surface in parallel; any veto cancels the entire hierarchy close. Call muxy.lifecycle.close() to finish the close yourself without re-asking. No permission, no manifest field — registering the handler is the opt-in, and it fails open (no handler / timeout / throw ⇒ closes). It does not fire on app quit, project switch (panels are torn down and restored per project), or an outside-click popover dismiss; for those, persist reactively instead. Open panels are session-scoped per project — switching projects recreates webviews, so use muxy.storage / project.switched for durable panel state. To merely react after a close, subscribe to tab.closed / panel.closed / popover.closed. See Lifecycle and Panels.
Drive and automate the built-in browser with muxy.browser.*. Tabs: open(url, { split }) returns the tab ID; navigate(tabId, url), reload/back/forward(tabId), list(), read(tabId) ({ title, url, text }, ~1 MB cap), close(tabId). Automation: eval(tabId, script) (returns the parsed JS result), click, type(…, { submit }), fill, press(tabId, key, selector?), select, hover, scrollIntoView, setChecked. Waiting: wait(tabId, { selector|text|urlContains|function, timeoutMs }), waitFor, waitForNavigation. Inspection: getText/getHTML/getValue/getAttribute/getCount, is(tabId, property, selector), find(tabId, kind, value), snapshot(tabId, selector?) (visible interactive elements — let an agent "see" the page), screenshot(tabId) (base64 PNG). State: storage.get/set/clear(tabId, key, value?, kind) for local/session storage; cookies.get/set/delete/clear(tabId, ...) per profile. Reads and JS-running calls (eval, click, type, waitFor, get*, screenshot, storage.*) need browser:read/browser:write; navigate, cookies, list do not. All of them work headlessly on any open tab in the active project — the tab need not be visible or focused (screenshot renders off-screen). Every call fails when the user disables the built-in browser. Capture the tab ID from open/list and reuse it. See Browser.
Make hover and active states visible in both light and dark — background: var(--muxy-hover); border-color: var(--muxy-accent); is the standard pattern. Change the relevant color instead of lowering the whole control's opacity, which also fades its content.
Respect prefers-reduced-motion — Muxy users opt into Reduce Motion at the OS level; avoid long transitions, large translations, autoplay.
No hardcoded ~/.config/muxy paths from inside the extension — rely on the working directory Muxy sets, or pass cwd to exec.
