Repository navigation
A permission model and update gating for third-party plugins #10295
Xenobones
started this conversation in
Suggestions
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The problem
omarchy plugin addinstalls arbitrary git repositories, and third-party plugin code runs unsandboxed with full user privileges inside the Quickshell process. A plugin manifest today declares UI metadata only (kinds, entry points, bar settings) — there is no concept of capabilities, and nothing sits between plugin code and the user account.Concretely, any third-party plugin can, by design:
Process,Quickshell.execDetached)XMLHttpRequest)That's fine for first-party plugins, but the third-party ecosystem is real and growing (Better Displays, Network Scanner, Search Bar, Workspace Lens, App Launcher, Google Drive, and more on omarchyplugins.com). Every one of those repos is a supply-chain dependency that runs with full user privileges — and the only thing standing between a malicious or compromised plugin update and full user-account code execution is the user remembering to manually diff the repo before pulling. And the stakes aren't just per-user: one bad plugin incident would burn trust in the entire plugin ecosystem overnight, and that kind of reputation damage spreads much faster than any fix.
GNOME Shell extensions have this exact model, and it's been a long-standing criticism there. Browsers and VS Code solved it years ago with declared permissions plus process isolation.
Proposal, phased
Phase 1 — disclosure and update gating (pure Omarchy, no Quickshell changes needed)
Manifest permissions field. Extend the plugin manifest schema with an optional
permissionsarray using a small vocabulary:A bar widget that only renders text has no business requesting
process-exec, so the declaration itself surfaces intent.Consent at install/enable. When a plugin is added or enabled, show the user which permissions it requests — the Android-install-screen moment. Plugins that declare no permissions would be the default-safe case.
First-class update gating. Record the approved commit hash when a plugin is installed or enabled. Make
omarchy plugin update:--yesfor scripted use)This closes the "update on autopilot" hole for every user by default, without changing the runtime model at all.
Phase 2 — soft enforcement (needs Quickshell cooperation)
Load each plugin in a restricted QML context where capability-backed APIs (
Process,execDetached,XMLHttpRequest) don't resolve unless the corresponding permission was granted. Undeclared use fails at load time instead of silently succeeding. Quickshell controls the import surface, so a mediated API layer looks technically feasible.Phase 3 — isolation
Per-plugin processes communicating with the shell over a defined IPC bridge — the VS Code extension-host model. This is the only true security boundary (a compromised plugin process can't touch the user's files), and it's obviously the largest lift. Worth having on the roadmap even if it lands last.
One data point
After reviewing a third-party plugin recently, I threw together a local "update watchdog" for my own machine: an interactive gate that replaces
omarchy plugin update(fetch, show commits/diff, approve-before-pull) plus an inotify sentinel that disables any plugin whose files change without approval. It took an afternoon and already surfaced one edge case worth designing around: deliberate local edits to installed plugins are a legitimate workflow, so approval should hash the working-tree state rather than just HEAD, or users who customize plugins will fight the tool.Ask
Just thinking out loud as a user who'd love to see the plugin ecosystem grow safely:
All reactions